The errata register — every published claim this project got wrong, with the fix attached. Part I of the ledger; the changelog and diary are beside it, and the index explains what each is for.

Part I — The errata register

Errata Register — Public Explainers vs. Statute Text

Opened 17 August 2026. Rule of this register: a research draft may honestly contain unresolved issues; it may not contain silent ones. Every entry below quotes the public claim, quotes the statutory text it diverges from, states which one is wrong, and states the fix and its status. Five entries are resolved by changing the statute (the explainer stated the design intent; the v3.3 text fell short of it — the text moves to match the promise). One entry (E6) was a plain copy error and is corrected in the explainer today. Nothing is deleted; corrections travel with the claims they correct.

Explainer line numbers are as of the 17 August 2026 revision. Statutory citations are to model_act_v3_3.txt.

E1 — Engineer exemption: claimed as written; in fact implied, not yet express

The claim (docs/03-whats-in-the-act.md, line 16): “NOT liable: you, the engineer. written into the definitions: rank-and-file employees are exempt.”

The text (SEC. 4(a)): a controlling person is “any natural person who, regardless of title, possesses or exercises material practical authority over a covered system through any of: (1) deployment, expansion, or access decisions; …” No subsection expressly excludes status, credentials, ministerial execution of another’s decision, or technical ability standing alone.

The gap. The exemption is real in design — “material practical authority” is the operative screen — but it is inferred from the definition, not “written into” it. An aggressive reader could argue an engineer with production access makes “access decisions.”

The fix (statutory). The controlling-person definition is narrowed to final material independent decision authority, with express textual exclusions for status, credentials, ministerial execution, and technical ability standing alone — making the explainer’s sentence true on the face of the text. Queued for the public working branch (v3.4 cure list). Landed: the cure entered the statute verbatim on 19 August 2026; tagged v3.4. The explainer claim stands as the binding design intent.

E2 — Certification cadence: “every quarter” is not in the statute

The claim (docs/03-whats-in-the-act.md, lines 10 and 42): “the CEO personally signs a safety certification every quarter” / “CEO signs a safety cert every quarter, personally.”

The text (SEC. 8): “Before material deployment and following material change to a covered model or configuration, the chief executive officer … shall personally certify …” The trigger is event-based. No quarterly or other periodic cadence appears anywhere in SEC. 8.

The gap. The explainer promised a rhythm the statute does not contain; event triggers alone are also independently attackable as vague (“per-configuration cadence,” audit finding).

The fix (statutory). Certification triggers are being defined precisely, with a periodic batch cadence below the material-change line — giving the certification both the defined events and a regular clock. Until that lands, the accurate public sentence is: the CEO signs before material deployment and after material change, personally and non-delegably.

E3 — “No signature, no shipping”: the signature is not a gate, and a signed confession currently counts

The claim (docs/03-whats-in-the-act.md, line 10): “no signature, no shipping.”

The text (SEC. 8): “a certification disclosing identified noncompliance satisfies this section.” And SEC. 3(c)(2)(B): provisional validation may “document… the conformity … or disclose… identified nonconformity and the compensating measures taken.”

The gap. Two divergences. First, certification is a duty with criminal consequences for lying (SEC. 6(b)(1)), not a shipping gate. Second — the serious one — a certification candidly disclosing unremediated noncompliance both satisfies SEC. 8 and, with “compensating measures” of unspecified adequacy, can support provisional validation. Truthfully disclosed unsafe deployment is punished nowhere in that configuration: it punishes lying, not shipping.

The fix (statutory). Validation is being rebuilt to require a reasonable documented conclusion of material conformity; compensating measures must be equivalent, judged against stated criteria; disclosure of nonconformity becomes a report, never a validation; and a certification disclosing unremediated material nonconformity constitutes neither compliance, validation, cure, nor a defense. Queued for the public working branch. Landed: the cure entered the statute verbatim on 19 August 2026; tagged v3.4.

E4 — “No waiting for an agency” vs. the Agency-approval validation mode

The claim (docs/03-whats-in-the-act.md, line 33, and the project’s standing no-permit-regime design): the Act does not gate anyone’s deployment on an agency’s say-so.

The text (SEC. 3(b)): “The Agency shall specify for each standard the mode of validation (internal attestation, independent audit, accredited certification, or Agency approval).”

The gap. If the Agency selects the fourth mode for any standard, deployment lawfully waits on an affirmative agency act — a permit regime through the back door, contradicting the Act’s own design and its commencement architecture, which was rebuilt specifically so that agency inaction can never stall the statute.

The fix (statutory). The Agency-approval validation mode is struck. Queued for the public working branch. Landed: the cure entered the statute verbatim on 19 August 2026; tagged v3.4.

E5 — “It was never going to be you”: true for the weekend model, not yet true for the startup

The claim (docs/03-whats-in-the-act.md, line 22): “your fine-tune, your weekend model, your use, study & modification of weights: untouched. sec. 1 says so, out loud.” (README.md, line 100: “Your startup is not in these chairs.”)

The text (SEC. 1(b)(3)): “‘deployer’: the entity that … operates a covered model or system”; the express carve-outs are personal, noncommercial operation (SEC. 1(b)(3)) and use, study, or modification of weights “except as part of the deployment of a covered system” (SEC. 1(b)(9)).

The gap. The hobbyist claims are accurate. The startup claim is not yet: a company commercially operating a configuration of a covered frontier model is a deployer with SEC. 2 duties, and v3.3 contains no de-minimis rule and no reliance rule for non-modifying deployers. Every thin wrapper is, on the current text, a criminal-statute deployer.

The fix (statutory). A conduct-based de-minimis and reliance rule: non-modifying deployers discharge the duty by documented adoption of an upstream validation plus their own tool, credential, permission, and monitoring manifest. Never revenue-based. First in the cure queue.

Addendum, 17 August 2026 (evening): the same claim also appears in the dossier Q&A (dossier/05_questions_and_answers.md, startup answer). That page now carries the gap and the queued cure inline, same-day.

E6 — Commencement: the copy error, corrected today

The claim (docs/03-whats-in-the-act.md, line 33, as published until 17 August 2026): “it starts working day one — no waiting for an agency that doesn’t exist yet.”

The text (SEC. 3(c)(1)–(2)): the truth-telling, incident-reporting, records, whistleblower, and liability provisions “arise and operate from the effective date, and do not depend upon … any … act of the Agency”; but “No duty arises under SEC. 2, and no offense lies under SEC. 5(a), before the provisional commencement of paragraph (2)” — which begins [180] days after the effective date, on interim standards adopted verbatim from enacted California, New York, and Illinois law.

The status. “No waiting for an agency” was and is correct — that is the point of the interim-standards design. “Starts working day one,” unqualified, was wrong as to the core deployment offense. Corrected in copy today on the same card, with a dated note; the superseded wording is preserved here, per the no-deletion rule.

E7 — The withdrawn PDF still called itself “the introducible text”

The claim (CHANGELOG.md, integrity patch, 17 August 2026): “‘Introducible’ retired everywhere until a gated sponsor release earns it back,” and “model_act_v3_3.pdf withdrawn … The file stays in the tree for link integrity; the README no longer offers it.”

The artefact (model_act_v3_3.pdf at the repository root, cover note, as shipped 16 August 2026): “[This file is the introducible text: SEC. 0 (uncodified findings) and SEC. 1 through SEC. 13. …]”

The gap. “Everywhere” missed the inside of the withdrawn file. De-listing removed the README’s link, but the file still opened normally at the root path, carried no withdrawal notice, and self-described with the retired word — the last live instance of it in the tree. The .txt got its signpost at the old filename; the PDF’s cover line got nothing. v2 and v3.2 both received one-page signpost PDFs; v3.3, the current version, was the only one without.

The fix (packaging, applied 18 August 2026). The v2/v3.2 practice, applied late: the typeset edition moves unchanged to archive/model_act_v3_3_withdrawn.pdf with its correction attached in the archive README, and a one-page signpost PDF holds the old root path so old links land on the honest label. Outside catch (a reader running a link-checker over the tree — 160 internal links, zero broken, one PDF that missed the memo). This is a packaging entry, the register’s seventh; the five/one split described above covers E1–E6.

E8 — “In one paragraph”: true of the duty, silent on the entity, in the paragraph built to be quoted

The claim (README.md, “In one paragraph,” as published to 20 August 2026): “The base duty is due care, not strict liability; the custodial tiers are maximums…”

The text (SEC. 10(a)): “Entity: civil penalty of up to $[1,000,000] per violation for each day the violation continues… strict liability.” And SEC. 6(c): “Entity liability under SEC. 10(a) is strict.” And SEC. 1(a): offenses under the Act “are public welfare offenses within the meaning of Morissette v. United States, 342 U.S. 246 (1952), except where a greater mental state is expressly required.”

The gap. The claim is exactly true of what it names — SEC. 2 is headed “PUBLIC WELFARE DUTY” and its base duty is the exercise of due care — and it is incomplete as a characterisation of the Act, which contains an express strict-liability limb and classifies its offenses into the Morissette family. The README states the qualified version correctly three times further down (SEC. 6 summary: “strict liability buys entity fines only”; the scienter-drift passage: “strict liability survives only where the modern Court tolerates it, in the entity’s civil penalty under SEC. 10(a)”; the section walk: “strict liability is for entity fines only”). Only the summary drops the qualifier.

Why it is graded higher than its size. “In one paragraph” was written for machines to lift verbatim, after a search engine’s AI summarised the project with the doctrine unnamed and the scope inflated (diary, 19 August 2026). It is the single most quotable passage in the repository and therefore the one place an omission propagates without the surrounding correction travelling with it. A reader who takes “not strict liability” from a summariser, opens SEC. 10(a), and finds the words “strict liability” has been handed a contradiction the project put in his way — in a project whose whole premise is that its claims survive being checked.

The fix (explainer, applied 20 August 2026). One clause, no statutory change: “The base duty is due care, not strict liability — strict liability reaches an entity’s civil penalty alone, never a custodial sentence; the custodial tiers are maximums…” This conforms the summary to the three correct statements below it and to SEC. 6(c). Outside catch. The statute needed no amendment, which is the finding worth keeping: the drafting was right and the shop window was thin.

Precision notes (audit record)

N1 — New York penalty phrasing. audit/chunk3_penalty_architecture.md §A.3 states the New York figures flatly (“$1,000,000 first / $3,000,000 per subsequent”). The enacted text (GBL § 1427, consolidated through 2026-04-03, pinned 17 August 2026 against nysenate.gov) phrases both as caps — “not to exceed” — with the amount “determined based on the severity of the violation.” Chunk 3 §D.1 already characterizes the family as severity-scaled; §A.3’s flat phrasing stands corrected to caps, severity-scaled. Public copy should say “up to.”

N2 — Explainer section numbering. The “whole act, plainly” list on docs/03-whats-in-the-act.md uses its own compressed numbering (certification at “SEC.5,” reporting at “SEC.6,” penalties at “SEC.7”), which does not match the statute (certification SEC. 8; reporting SEC. 9; penalties SEC. 10; whistleblowers SEC. 11). The card’s header already directs readers to the statute as authoritative. Logged so the divergence is a recorded choice, not an oversight; the list will be renumbered in the next docs pass.

This register is append-only. When a statutory cure lands on the working branch, its entry gains a dated “landed” line; entries are never removed.

E12 — 20 August 2026 (internal catch, same day). F1 — that nobody on the predecessor FDA docket names an upstream person — was published in four places at a strength the finding’s own file forbids.

The claim. The front page’s Recent entry and its contents table both described the reading notes as “the predecessor comment file read end to end” reporting “the element none of them names,” of “the 51 comments.” The standing watch made it the punchline of its headline finding: “fifty-one commenters named no upstream person, and four frontier statutes name none either. Two independent evidence bases, the same vacancy.” The diary carried the same sentence.

The text it contradicts. F1’s own strength note, in bold, in the file all four were describing: “the wider claim is true of tiers 1 and 2 on reading, and is not certified across all 51.” The register said the same of F3. The repository asserted on its front page precisely what it stated in bold, two clicks away, was not certified.

What the roster then established. The substance of 29 of the 51 has never been read (§ 1.3). The claim was running over twenty-nine unopened comments. “Read end to end” was false on its face about a file whose three-tier structure exists because it was not read end to end.

Why it is graded at E8’s level rather than below it. E8 was a summary of our own statute, where the drafting was right and the shop window was thin. This is an empirical claim about other people’s documents, and its entire rhetorical force came from the number 51 — the pairing “two independent evidence bases” does no work at 22. A hostile reader following the front page to the notes would have been handed the contradiction by us, in a project whose sole authority is that its claims survive being opened.

The fix (copy, applied 20 August 2026). No finding is withdrawn; each is restated at the strength its evidence carries. The front page now reads “rostered in full and read in part — every filer named, the substance of 22 of them read, and the element none of those 22 names.” The standing watch and the diary carry dated corrections in place, superseded wording preserved. The statute needed no amendment. Internal catch, prompted by an outside reader asking how the roster was counted.

E11 — 20 August 2026 (outside catch; the roster). The reading notes on docket FDA-2024-D-4488 were compiled from 22 comments read in full or as posted text, 13 more known by title, and 16 never enumerated at all. The complete 51-filer roster was then read from the docket’s three result pages, and it falsified four published claims in that file. Nothing here touches the statute.

(a) The substantive one. The file called the National Multiple Sclerosis Society “the file’s only patient organisation,” in the § 1.1 census and again in § 2. It is not: the National Health Council (0034) — the American patient-advocacy umbrella — and Pathway for Patient Health (0047) are both patient-side bodies, and both sat in the sixteen this file had never enumerated. NMSS is the only single-disease patient organisation, which is the claim that survives. The error is instructive about its own cause: a finding about who is absent from a file was published while a third of the file was unread, and the two missing filers were identifiable from their names alone. Corrected in both places, with the count stated: four filings of fifty-one come from the patient side, twenty-one from industry.

(b) “Anonymous” is not “unattributed.” Comment 0012 was recorded as unattributed with the filer left —, on the reasoning that the attachment carried no signature block. The docket names the filer Anonymous. It is a filer who took the option, not one who forgot the letterhead — and it is one of three anonymous filings (0012, 0038, 0050), where this project’s campaign copy had said two.

(c) An inference from a gap that was not there. The file reasoned from the ID range that “a small number were received and not posted.” Comment IDs run contiguously 0003–0053 with no gaps — exactly 51, the two docket documents taking 0001–0002. Nothing was withheld on this docket, and the received-vs-posted lag observed live on FDA-2026-N-7874 gets no support from here. A gap was asserted from arithmetic that was never done.

(d) Two late comments were four. § 5’s guidance-dockets-never-close finding named Ikeda (0051) and Yang (0053). 0050 (posted 12 Aug 2025) and 0052 (posted 17 Apr 2026) were also filed after the April 2025 close. The finding strengthens; the count was wrong.

Two smaller catches in the same pass, below erratum grade and recorded for completeness: Jitendra Pund’s comment ID, previously —, is 0049; and 0028 is posted under the individual name Cythika Bopearachchi, one of the cohort members, not under the university or Prof. Sampat.

What the roster unlocked. F3 — no frontier model developer filed — is certified against the complete list, with the qualifier that must travel with it: none filed in its own name, and two trade associations whose membership includes frontier developers (CTA 0035, Connected Health Initiative 0039) did file. The project’s running list had recorded F3 as blocked on capturing five comments’ substance. It never was: F3 is an absence-of-filer claim and needed only a page-through of the roster. F1 and F8 are the findings that need the substance. Two different blockers had been filed under one line, and the cheap one went unrun. Status: cured in the reading notes; the substance of 29 comments remains uncaptured and is stated as such.

E10 — 20 August 2026 (internal catch). The statute’s header bracket at line 5 refers to “the open items for v4”; the next revision is v3.5, as the companion and the README now say. The text is non-operative — the bracket is apparatus, not statute — and it is left uncorrected on purpose. model_act_v3_4.txt is tagged, checksummed in the changelog, published as the source hash of the reviewer’s copy (399c725adcd117aa7736a63b716328226eb24f33a48695115d941b68caace1bf), and archived at CERN under DOI 10.5281/zenodo.22029795. Editing a byte would falsify the reproducibility chain rather than improve the text: the PDF would no longer rebuild to its published hash, and the claim a reviewer is invited to check would become false. Corrected at v3.5, when the file re-hashes anyway. Status: open by design, closes at the next revision.

E9 — 20 August 2026 (internal catch). The register’s own rule — “when a statutory cure lands on the working branch, its entry gains a dated ‘landed’ line” — was applied to E1, E3, and E4 but not to E2 or E5, which landed in the same tagged revision. Both are recorded here rather than edited into the sealed block above. E2 (certification cadence) landed at v3.4 on 19 August 2026 as cure 12: SEC. 8 now defines its triggers and requires changes below the material line to be certified in a periodic filing at least once in each [calendar quarter] in which any occurred (companion n.39). The explainer’s “every quarter” is true on the face of the text as to sub-material changes; material deployment and material change remain event-triggered. E5 (deployer and startup reach) landed at v3.4 on 19 August 2026 as cure 1: SEC. 2(b) gives non-modifying deployers a conduct-based reliance rule — documented adoption of an upstream validation, a manifest of tools, credentials, permissions, and external access, monitoring within the deployer’s control, and reporting within its knowledge — never conditioned on revenue, size, or resources (companion n.28). The dossier Q&A’s startup answer, which carries the gap inline per E5’s own addendum, is accordingly overtaken and should be conformed at the next docs pass. Status: both cured; the omission was in the register, not the statute.

E30 — 19 August 2026 (internal catch, same day). (Renumbered from “E8” on 22 August 2026: this entry and the “In one paragraph” entry above were both filed as E8, unrelated to each other. The strict-liability entry keeps E8, because E12 cites it by that number in published text; this one takes the next free number. The collision itself is recorded at E31. Original number retained here so a reader following an old citation lands correctly.) The consolidated front page claimed that the original Sacramento scorecard table was “preserved verbatim in the diary”; it is not — the diary never carried the table. The claim is corrected to point to the repository’s pinned history (docs/06-track-record.md at commit 6f48eff), where the original card is preserved unchanged. Two smaller corrections landed in the same patch: the DeCoster chronology is reconciled (sentenced 2015; affirmed on appeal 2016), and an opening sentence overstating the general law is tightened to the statutory-gap claim the project actually makes. Caught by our own hostile read-through within the hour of publication; the fix is live on the front page. Status: cured.

E13 — 20 August 2026 (internal catch). SEC. 8 of the tagged v3.4 text contains a punctuation error: “risks, or merits of any model or system, A certification disclosing identified noncompliance…” The comma before “A certification” should be a period. The error does not alter the provision’s meaning, and the archived v3.4 file remains unchanged to preserve its published hash and reproducibility chain. The one-character correction is queued for v3.5. Status: open by design; closes at the next revision.

E15 — 21 August 2026 (internal catch, same day; the sweep’s own grading). The research sweep of 21 August marked several details ✅ that were read in a secondary source quoting the primary, not in the primary itself. Under the standing rule adopted in E14 the same day, that is the wrong grade.

The claim (the sweep’s working notes, unpublished, and — for two items — carried into dossier/README.md by the corrections patch of the same date): the sweep’s Part 1 and Part 2 entries were graded ✅ on the strength of first-party authorship of the source quoted, without regard to whether the sweep had opened that source.

What was actually read. One primary was fetched and read in full: the arXiv abstract page for 2608.10218. Every other ✅ in the sweep rests on a secondary outlet quoting or paraphrasing a first-party document. That distinction is invisible in the sweep as written.

The affected details, regraded.

✅ stands. The Mind Viruses corrections (viral persona; near-total immunity; the context-wipe experiment; “a real but currently limited risk”; the Fellows Program and EPFL affiliation). All four are stated in the abstract, and the abstract was read at arxiv.org/abs/2608.10218, submitted 10 August 2026, 20:37:57 UTC. E14’s locator rule is satisfied.
⚠ pending fetch, and the patch overstates them. The four-accounts role breakdown (staging and outbound relay, data storage, two read-only) is reported by SC Media as OpenAI’s own 28 July update. It was not read in OpenAI’s post. The 3 million GPU hours and the “collaborative knowledge sharing between models” quotation are attributed to JFrog CTO Yoav Landman’s blog and appear to be quoted from it at length, but the blog was not opened.
⚠ pending fetch, in the unpublished sweep notes only, not in the dossier. The nine Artifactory CVE numbers and the release-note chaining caveat; the Hugging Face forensic figures (~17,600 actions in ~6,280 clusters); the contents and signature block of the 15-AG letter of 3 August; the AI Kill Switch Act’s thresholds, penalties, and red-teaming carve-out — closed 21 Aug at ⚠ R, the bill text having since been opened directly; see the census. Each has a named, fetchable primary: JFrog’s release notes, Hugging Face’s technical timeline, the letter itself or the Pennsylvania Office of Attorney General page, and the bill text or Rep. Lieu’s press release.

The gap, stated plainly. The sweep applied the project’s tier discipline (primary beats press beats reconstruction) but graded on who wrote the source, not on what the sweep opened. Those come apart exactly where it matters. A first-party document quoted accurately in three outlets is very probably accurate; it is still not a document this project has read, and the register’s whole premise is that the difference is not rhetorical. The 17 August pass proved the cost when three summary-era details fell against primary. This is the same failure with a better class of secondary.

Why it is graded higher than its size. It arrived on the same day as E14, which adopted the locator rule, and it broke that rule in the first substantial pass after adoption. A rule that does not survive its own first day of use is not yet a rule. Two of the regraded details are already in the working tree via the corrections patch, so the overstatement is live, not merely drafted.

The fix. Two parts, neither of which requires withdrawing the sweep’s substance. First, the corrections patch carries an inline grade note on the two affected details until each primary is opened. Second, the standing rule is extended: a ✅ requires that this project opened the source, not merely that a first party wrote it. Where the primary is quoted in a secondary and not yet fetched, the grade is ⚠ with the fetchable locator named. The fetch queue above is the work; none of it is difficult, and the substance is unlikely to move. What moves is whether the file can say it checked.

Outside catch, in the useful sense: the question that produced this entry was whether research conducted by companies about themselves can be trusted. The first honest answer turned out to be about this project’s own reading, not theirs. Status: open; closes item by item as each primary is fetched.

E16 — 21 August 2026 (internal catch; a coverage failure, not a false statement). The standing watch of 20 August, and the finding it carries at § 7(5), say “across four frontier regimes.” At that date there were at least six. Two were missed.

Connecticut, SB 5, enacted 27 May 2026 — twelve weeks before the sweep. It is a frontier statute on the same threshold this Act uses: a frontier developer is one training a foundation model on computing power “greater than ten to the twenty-sixth power integer or floating-point operations,” with a large-developer tier at $500,000,000 in annual revenue. The word “Connecticut” appears nowhere in this repository except as the 1991 due-process case at audit/record.md. The front page describes the interim standards as borrowed from “three enacted state laws.” There are four.

H.R. 9917, the AI Kill Switch Act, introduced 23 July 2026 — four weeks before the sweep, bipartisan, and covered that week by Roll Call, CNBC, Fox News, Al Jazeera and Tom’s Hardware. The watch tracked H.R. 9925, introduced the same day, and did not see the bill beside it.

What survives. The finding itself, undamaged and wider. Neither missed instrument reaches a natural person. Connecticut writes officers and directors into a frontier provision — quarterly anonymous reports of catastrophic risk “shall be shared with the officers and directors of the large frontier developer” — and attaches to that knowledge no duty, no response obligation, no signature and no liability. H.R. 9917 mandates a shutdown capability and civil penalties to $20,000,000 a day, and contains no officer, no director, no natural person, no certify and no criminal provision at all; its only human signature is the sponsor’s own on the introduction line. Six regimes, one vacancy.

What does not survive. The implied completeness. A project whose entire credential is a negative finding must be able to say how hard it looked, and a watch that misses an enacted statute in the family it claims to track — and a bill in the national press — is not a watch. The number four was not wrong about the four. It was wrong about the world.

Cause, stated so it can be fixed rather than regretted. The sweep was organised around the instruments already adopted at SEC. 3(c)(4) plus the federal vehicles already tracked. It re-checked what it knew instead of searching for what it did not. A watch assembled from its own prior list will keep returning its own prior list, and will report the absence of everything it never looked for.

Read E15 and E16 together; they are one failure at two scales. E15 is this project grading a citation on who wrote the source rather than on what the project opened. E16 is this project grading a field on what it had already adopted rather than on what exists. Both mistake the boundary of our own effort for the boundary of the world — the precise error the whole register was built to catch, arriving twice in one day, at the level of a footnote and at the level of a finding. Neither produced a false statement. Both produced a true statement standing on a claim of thoroughness it had not earned. That is the failure mode this project should expect to keep having, and the reason the register is the credential rather than the statute.

Cure. The bill census, opened 21 August 2026, which starts from external bill lists rather than from this project’s own adoptions, grades every row’s confidence, records the source lists’ own errors, and never states a total above the number of bills actually read. H.R. 9917 and Connecticut SB 5 are its first two completed rows. Three consequential text fixes follow: the front page’s “three enacted state laws” becomes four; the standing watch’s § 7(5) is restated at six with both misses named; and SEC. 3(c)(4)’s concordance acquires a Connecticut line or an express note saying why it does not.

Credit. The Kill Switch cluster was pointed out by a reader. The Connecticut miss was found in the check that followed it. Both are recorded here rather than quietly patched, on the register’s standing rule: the correction is the product. Status: open; closes when the three text fixes are live.

Closed 21 August 2026, and the cure needed correcting on the way. All three are now live — with one departure from what this entry prescribed, recorded because the entry was wrong about it. E16 called for the front page’s “three enacted state laws” to “become four.” Executing that literally would have introduced a fresh error, because those words describe what SEC. 3(c)(4) adopts, and it adopts three. The defect was never the count; it was the implication of completeness — a reader meeting “three enacted state laws” with no further word concludes there are three. So the cure applied is the accurate one: the count stands wherever it describes what the Act adopts, and the adopted texts now carries a paragraph naming Connecticut, its threshold, and the two reasons it is not adopted. The standing watch § 7(5) is restated at six with both misses named and the cause attached. Whether SEC. 3(c)(4) should adopt Connecticut at v3.5 is a drafting decision, not housekeeping, and is held as Open Question 1 in the open cure queue — the tagged v3.4 text is untouched, per E10. audit/record.md is a frozen drafting record and its “three enacted states” was true on the day it was written; frozen files are not retrofitted.

E17 — 21 August 2026 (outside catch; an overstated disanalogy, and a second overstatement in the cure for it). The same conduct, prosecuted, as first published, said of the five computer-crime defendants it gathers: “Every defendant below acted intentionally and knew they lacked authorisation.”

That is not true of those cases, and the error was the load-bearing one in the section. Whether authorisation had been exceeded at all was the central contested question in at least two of them. The Third Circuit’s own description of Auernheimer’s script is that it “accessed the publicly facing portion of the login screen and scraped information that AT&T unintentionally published.” Swartz was on a network he was entitled to use. The intent that looks obvious in hindsight was disputed at the time — which is what the prosecutions were about.

Why it mattered. The sentence appeared in a paragraph headed “the honest disanalogies, stated before anyone else states them” — a passage whose entire purpose was to concede the weaknesses of the comparison. An overstatement inside a concession is worse than one inside an argument, because a reader who checks it loses confidence in the concessions too, and the concessions are what make the file credible.

The fix. The claim is withdrawn and named as withdrawn in place, rather than quietly deleted. Three disanalogies that do survive are stated instead — the fraud-statute point, the intended-release point, and the CFAA-overbreadth point, the last of which is left as a live alternative conclusion a reader may reasonably reach against this project.

And a second correction caught in the same exchange, recorded here because it is the same species. The scope note added to the table of authorities described the omission of the standards/ citations as “deliberate.” It was not. The table was compiled on 20 August; the files citing that law were written on the 21st. Nobody decided anything — the omission was an accident given a policy’s clothing after the fact. The note now says so, and takes the decision openly instead.

Credit: both were caught by the maintainer on reading the drafts, not by the drafter. Status: cured; both files carry the corrected text.

E14 — 21 August 2026 (locator failure, primary-source conflict, and forward repair). The dossier’s entry on Agents of Chaos (arXiv 2602.20021) presented detailed case-study claims under a general citation to the paper, with no case-level locators, marked ✅. Checking the details against both primary versions found that the citation practice, not the underlying reading, was the central defect — and that one detail is version-sensitive in a way the entry had concealed.

The conflict. The entry said “a nine-day agent-to-agent loop.” That is supported by arXiv v1, which reports a mutual relay lasting at least nine days and ending after owner intervention. The authors’ current official case-study report describes the same case as an approximately one-hour relay that the agents terminated autonomously. Both are primary and both are the authors’ own. The earlier wording was not fabricated and is not withdrawn as false; what was wrong was presenting a version-sensitive figure as settled.

Disposition. CS4 now states the conflict on its face, cites both versions, and asserts no duration. It carries the case for what neither version disputes: a non-owner inducing conversational loops and persistent background processes with no designed termination condition. CS1 is restored and made more precise: the agent reported the deletion complete while the email remained in the Proton mailbox, untouched by the local deletion — the report-versus-reality finding occurring concretely inside a case study, not only in the abstract. The case-count wording is also resolved without relying on an aggregate total: arXiv v1 presents eleven principal numbered case studies and separately numbers five failed or hypothetical experiments CS12–CS16, while the current site presents all sixteen as incidents. The entry relies on identified cases and asserts no total.

The standing rule. A ✅ on a detail not stated in the abstract must identify the supporting case study, section, page, or equivalent primary locator; otherwise the detail carries ⚠. Where a cited work has more than one published version, an entry may not assert a figure that differs between them without naming which version it follows or stating the conflict.

The correction record. Two temporary upload commits on 21 August — 75832e3 and 899ea95 — corrected parts of the source reading but reused the number E13, displaced the existing punctuation erratum, and regressed unrelated front-page, contact, citation, and publication text. This forward repair restores the pre-upload versions of those unrelated files, restores the original E13 unchanged, and records the source correction here as E14. The temporary commits remain in the public history rather than being rewritten. Nothing about the statute, its version, its citation metadata, E11, E12, or E13 is changed. Status: cured.

E18 — 21 August 2026 (internal catch). The Sarbanes-Oxley analogy was published in a structure that implies a prosecution record it does not have.

The claim (standards/why_a_signature_works.md § 2). The file establishes in § 1 that a signed false document produced twenty-eight years for Stewart Parnell. It then lists the SOX certification penalties — ”$1,000,000 and ten years if knowing, $5,000,000 and twenty years if wilful” — and closes: “Twenty-four years later every public company in America has someone who signs. They did not run out of chief financial officers.”

Every sentence is true. The sequence implies a fourth one that is not: that the certification requirement has put executives in prison.

The record the file omits. Richard Scrushy, founder and former chief executive of HealthSouth, was “the first chief executive charged with violating the 2002 Sarbanes-Oxley Act.” On 28 June 2005 a federal jury found him not guilty on all thirty-six counts. Michael Zuppone, a former SEC regional office head, had said before the verdict that prosecutors could “wave that personal certification in front the jury to show that the defense claim — that their head was stuck in the sand — doesn’t hold water,” and said after it that “the utility of the criminal certification statute will be very much undermined.”

Why it is graded above its size. This project’s case for SEC. 8 rests on the claim that a signature is the instrument by which American law reaches an executive. The flagship modern example of a signature statute produced an acquittal in its first test. An opponent gets to say the mechanism is decorative using this project’s own exhibit, and the file does not mention the case anyone who knows the area thinks of first. Same species as E12 and E16: a true statement standing on an implication it had not earned.

What survives, and it is the stronger reading. SOX is strong evidence that a certification requirement changes conduct upstream — every public company now has a named person who has to ask, and the 2002 objections did not materialise. It is weak evidence that certification statutes are charged and won. The file was running two claims together and needed only the first.

And the honest mechanism is the one § 1 already states. Parnell was not convicted under food-safety law. He was convicted under fraud statutes that finally had a document to attach to. A certification does not create the offence; it makes existing offences provable. 18 U.S.C. § 1001, § 1519 and the ordinary fraud statutes already exist and already reach individuals. At the compute frontier they have nothing to attach to, because nobody signs anything. Stated that way § 1 and § 2 become one argument instead of two, and the file is more coherent, not less.

The fix (copy). § 2 gains the Scrushy record and states the claim at evidentiary strength. § 5’s conclusion is unchanged and is where the argument was always correctly put: a signature is “the ordinary instrument by which American law reaches an executive at all.” No statutory amendment. The Illinois follow-up email, unsent at the date of this entry, is corrected before sending.

Not asserted, and named rather than assumed. Whether prosecutions under 18 U.S.C. § 1350 are rare in general, and whether DOJ practice is to charge securities fraud or § 1001 with the certification as evidence, appear nowhere above: this project has not opened a source for either. Both are checkable and the fetch targets are named — the Federalist Society memorandum on § 906’s criminal penalties, and the Seattle University Law Review treatment of the certification provisions. Until then the entry rests on the one case it opened. E15’s rule, applied on purpose this time.

Sources: CNN/Money, 28 June 2005; NBC News, “Anti-fraud law fails first major court test”. ⚠ R — both opened 21 August 2026 via automated retrieval, not yet human-read.

Credit: raised by the maintainer asking whether the project’s own mechanism would actually reach anyone. Status: cured 21 August 2026. § 2 carries the Scrushy record and states the claim at evidentiary strength; § 3 and § 5 carry the related corrections described in E19. The two outstanding questions named above — the general prosecution rate under 18 U.S.C. § 1350, and whether DOJ practice is to charge fraud with the certification as evidence — are not closed by this and remain unasserted anywhere in the repository.

E19 — 21 August 2026 (internal catch; a live repository served the wrong document, and the cause was a tool this project built). For part of the evening of 21 August, the public Illinois repository served, as its SPONSOR_MEMO.md, the sponsor memorandum for the New York RAISE Act draft. Its CITATION.cff simultaneously described a third work — the general Frontier Artificial Intelligence Responsible Officer Act, at version 0.1, with both url and repository-code pointing at that other repository.

What that meant in practice. Anyone opening the Illinois repository for its sponsor memo was handed a memo about a different state’s statute. Anyone using GitHub’s Cite this repository control was handed a citation to a different document at a version and URL that were not this one. The repository was, for that window, internally inconsistent about which of three drafts it contained — in a project whose only authority is that its claims survive being opened.

Sequence. A commit at 22:31 replaced ninety-one lines of the Illinois memo with twenty lines of the New York memo. Its message said the opposite of what it did: “cite the enacted provisions at (C) and (G); rule the auditor out expressly; correct CITATION.cff.” Neither happened. The CITATION.cff in that commit was not modified at all. A repair commit restored the Illinois memo and corrected the citation file, and both were verified against the remote afterwards.

Cause, stated as far as it is known, and no further. The commit was produced by a shell script this project generated to move edited files onto the maintainer’s machine, because file transfer had failed repeatedly by other routes. The script wrote files and committed them without verifying afterwards what was actually on disk. How the New York memo came to be in that working tree is not established, and this entry does not guess — what is established is that nothing in the process would have noticed if it had been anything else, and nothing did.

Why it is graded with E15 and E16 rather than below them. Those two entries record this project mistaking the boundary of its own effort for the boundary of the world. This is the same error moved one step outward: a tool was trusted to have done what it was told, and its output was published without being read. The project’s standing rule is that a ✅ requires that we opened the source. It had no equivalent rule for artefacts we generate ourselves, and this is what that gap costs.

Two corrections follow from it, both applied. The repair script was rebuilt to verify file contents on disk and refuse to commit on mismatch — it now checks that the memo names Illinois, carries the (d)(2)(G) citation, and contains no reference to RAISE. And the standing rule is extended: a generated artefact is unverified until its output has been read, exactly as a source is. Committing is publishing.

One thing does not survive and is recorded as unfixed. The repair commit carries the same message as the commit that caused the damage, so the public history now shows two identically worded commits, the first of which broke the file and the second of which restored it. The history is not rewritten — E14 established that this project leaves bad commits in the public record rather than tidying them — so this note is the only thing that distinguishes them.

Credit: caught by the maintainer asking whether the push had actually worked, and confirmed by reading the deployed files rather than the local ones. The push had reported success. Status: cured; both files verified against the remote on 21 August 2026.

E20 — 21 August 2026 (internal catch, same day). “The only instrument naming officers and directors” was not the only one.

The claim (standards/frontier_bill_census.md, tally, and the Connecticut row’s verdict). The census reported 1 under “naming officers and directors in the operative text”, attributing it to Connecticut SB 5 alone, and the Connecticut row was written as the singular case: a frontier statute that routes quarterly catastrophic-risk reports to “the officers and directors of the large frontier developer” and asks nothing of them in return.

The text it missed. California SB 53, chapter 138 of 2025, adds Labor Code § 1107.1(e)(2)(A): “the disclosures and responses of the process required by this subdivision shall be shared with officers and directors of the large frontier developer at least once each quarter.” Subparagraph (B) disapplies it to an officer or director accused of the wrongdoing. Same recipients, same quarterly cadence, same absence of any consequent duty.

Cause. SB 53 had been word-tested for audit, signature, certify and certification — the terms the lineage work was chasing — and the nil result was correct. The word test the census’s own method prescribes was not run, and that list contains officer and director. A partial search was recorded as though it were the file’s standard one.

Why it is graded rather than fixed silently. The census’s whole value is that its counts are floors it can defend. A count of one, published as the only case, is a stronger claim than a count of two — it says a thing is anomalous. Two states doing something identically is a template, which is a different and better finding. The error made the evidence look thinner than it is while making the claim sound sharper than it was.

The fix (copy, applied 21 August 2026). The tally reads 2 and names both. The new SB 53 section states the parallel and draws the conclusion the correction supports: one state routing risk reports to named officers and asking nothing of them is a drafting choice; two states doing it identically is a template. No statutory change.

Credit: caught when the maintainer read the SB 53 chaptered text in full and pasted it, after the file had recorded a single automated retrieval of that same statute. The retrieval was right about what it was asked and blind to everything else — which is the standing argument for reading the primary. Status: cured.

E21 — Dates converted from “N days ago” run a day late, systematically

Filed 22 August 2026. Severity: low individually, method-level in aggregate.

The error. House language § 10a collected ten headlines from a search-results page. Dates displayed as “1 day ago”, “3 weeks ago” and so on were converted by counting back from 22 August 2026 and marked ⚠ approximate.

Two of those approximations became checkable when the underlying articles were opened. Both were wrong. Both were wrong by one day, in the same direction.

item	§ 10a estimate	actual	error
The Record	⚠ ~18 Aug 2026	17 Aug 2026	one day late
Reuters	⚠ ~21 Aug 2026	20 Aug 2026	one day late

Cause. A relative timestamp is a floor, not a point. “3 days ago” on a page rendered 22 August covers anything from the 19th back into the 18th, and rounding to the nearest whole day lands late. The ⚠ mark recorded that the dates were approximate. It did not record that the approximation had a known direction, which is the part that makes it correctable.

Why it matters more than two days. This project’s argument in that section is that exact wording carries legal weight. A file making that argument cannot publish dates it has not pinned — and a one-day error is enough to misorder a sequence, which is exactly what who has to tell you § 4a now turns on.

The fix (copy, applied 22 August 2026). § 10a’s grading section states the bias and instructs that every remaining ⚠ date be read as “probably a day earlier than shown” and published as none. Confirmed dates now come from the articles. No statutory change.

Credit: surfaced by the maintainer supplying the article texts. Status: cured for the checked items; six unopened headlines remain ⚠.

E22 — A quotation held in a working summary is not a quotation

Filed 22 August 2026. Severity: high — this one nearly published.

The error. A research file, press corpus, was written carrying four quotations attributed to four named people at four outlets. They came from a working summary of an earlier reading session, not from article text the project could still put its hands on. They were graded on the strength of “a human read the article in full” — true of the reading, and irrelevant to the transcription. They were withdrawn within minutes, before the file left the working copy, when the primary texts arrived and did not contain them.

Cause, stated precisely because it will recur. The confidence rubric grades how a source was obtained. It has no column for how the words travelled from the source into the file. A quotation that has passed through a summary is a paraphrase wearing quotation marks, and it passes every check the rubric currently runs.

Why it is the most dangerous entry in this register so far. E15 was about grading a claim on who wrote a source. This is worse: the source was real, the reading was real, the grade was honest, and the words were still not safe. One of the four — an expert account with no agency in it, sitting inside an article whose headline gave the verb to the model — would have been the single strongest item in § 10a. The best-sounding quotation in the batch was the one with no text behind it. That is the shape of the failure and it is not a coincidence: a remembered quotation drifts toward whatever the file needed it to say.

The fix (copy and method, applied 22 August 2026).

All four are quarantined in § 6 of the press corpus file, marked withdrawn, with the articles listed as leads to re-open.
New standing rule. A quotation enters a repository file only from text open at the moment of writing. If the text has to be remembered, the sentence is written as a claim about what the source says — never as words the source said.
Rubric note. ✅ describes acquisition, not transcription. A ✅ source can carry a ⚠ quotation, and the two must be graded separately.

No statutory change.

Credit: caught by the arrival of the primary texts — which is luck. Rule 2 exists because luck is not a control.

Addendum, same day: the first lead was re-opened, and the quarantine was justified

The BBC’s Meta article was opened hours later. The Hulme quotation exists. It is not what was remembered, and the three differences all run the same way:

held	published
“they’re not deliberately doing something devious…”	“are not conscious* — they’re not deliberately doing something devious”*
“…”	a whole elided sentence: models “coming up with very sophisticated strategies… to achieve the goal that they’ve been given”
“it will find a way.”	“it will find a way to achieve a goal that you haven’t thought about.”

The third is the finding. “It will find a way” is a sentence about a model. “It will find a way to achieve a goal that you haven’t thought about” is a sentence about the person who set the goal. The remembered version terminated precisely where the human being entered.

The clipped quotation was more agentive than the real one. House language § 10a argues that the corpus systematically clips toward agency. This project’s own working summary did the same thing, to a quotation it was going to use as proof.

The attribution was also wrong in a way that mattered. Daniel Hulme is global chief AI officer of the advertising firm WPP — a commercially interested executive, not the neutral expert the summary implied. That has to be stated every time the quotation is used.

Consequence for the rule. E22’s rule 2 was written as a discipline about texts. It is really a discipline about direction: a remembered quotation does not decay randomly, it decays toward what the person remembering it needed it to say. A summary is not a lossy copy. It is an interested one.

Status: cured. One of four leads released with corrections attached; three remain open.

E23 — Precision note: three quotations re-verified against the live page before publication

Filed 22 August 2026. Not an error. Recorded because the register should show the rule working, not only the rule being broken.

Immediately after E22, Chapter 3 of the UK Government Cyber Action Plan was added to why a signature works § 2a — a section whose entire weight rests on three strings: “personal accountability”, “a senior, capable individual with authority”, and “novel technologies, such as generative AI”.

All three were re-fetched from the live gov.uk page and confirmed verbatim before the section was written, along with the publication date (6 January 2026) and last-updated date (20 March 2026), neither of which was in the supplied text.

The re-fetch also produced a finding the reading had not: the phrase “personal accountability” occurs exactly once in the document. That is now the sharpest sentence in the section, and it exists because the check was run. E22’s rule paid for itself the same day it was written.

E24 — the front page conflated four different asks into one “get involved”

Filed 22 August 2026. A presentation defect in published material, caught by a private outreach audit, corrected the same day.

The README’s contribution section presented four distinct relationships — a formal council seat, a single bounded expert question, a talk-or-referral, and an anonymous citation catch — as one blended invitation, so a visitor could not tell which commitment was being asked of them, and the page appeared to ask everyone for everything. This is not a factual error in a claim; it is a defect in how the asks were shown, which matters because a review-recruitment page whose terms are unclear misstates, in effect, what a reviewer is agreeing to.

Fix: the section is recast as three labelled doors — check one thing, review one lane, talk or refer — with legislative and sponsor contact routed to its own track. The five-seat council terms are unchanged. Status: cured on the front page.

E25 — the two-twelves note miscounted the overlap it existed to clarify

Filed 22 August 2026. Internal catch, same day, by the repository consistency audit.

The frontier enterprises introduced a twelve-company coverage set alongside the existing framework twelve — the companies METR records as having published a frontier safety framework — and added a note headed “Two twelves, disambiguated” to stop a reader merging them. That note said “Five companies appear on both.” The overlap is eight: the five developers (OpenAI, Anthropic, xAI, Google DeepMind, Meta) plus Microsoft, Amazon and NVIDIA, each of which appears on the framework inventory this repository itself pins and in the coverage set’s compute layer. The one sentence whose whole function was disambiguation carried the arithmetic error.

Why it matters beyond the number. An undercount understates the very point the coverage set makes — that the compute layer self-designates as frontier just as the developers do — and it would have been checkable in ten seconds by any reader with both lists open.

Fix: corrected to eight, with the three named, in the enterprises file; the same disambiguation added to the models file, which introduces the framework twelve and previously carried no cross-reference at all. Status: cured.

E26 — H.R. 9917 was labelled the FRONTIER vehicle; it is the AI Kill Switch Act

Filed 22 August 2026. Internal catch, same day, by the repository consistency audit.

The frontier models file tabled other legal definitions of a covered developer and gave the row “Federal FRONTIER vehicle (H.R. 9917) | Over $100,000,000 of training-compute cost.” The threshold and the bill number are right and belong together; the name is wrong. H.R. 9917 is the AI Kill Switch Act (as the bill census records, and as E16 already established); the FRONTIER Act is H.R. 9925, cited correctly elsewhere in the same repository, including in CURE 4’s comparative note.

Why it matters. A misnamed federal bill in a comparison table is the kind of error that costs a reader’s trust in every other row, and this project’s whole claim on a reviewer’s time is that its rows can be checked. The two vehicles were already distinguished correctly in two other files, so this was drift within one table, not a misunderstanding.

Fix: the row now reads “Federal — AI Kill Switch Act (H.R. 9917)”. Status: cured.

E27 — the incident count was the BBC’s noun, in the file that exists to stop that

Filed 22 August 2026. Internal catch, same day, by the repository consistency audit. The correction of a correction.

The press corpus § 7 held an owed item — item 10, the count — precisely because the project was relying on an outlet’s arithmetic. Its discharge, on 22 August, then recorded: “Three developers disclosed — OpenAI, Anthropic, Meta — across four incidents.” Four is the BBC’s figure (“a fourth recent incident”), counted at the developer-event level. This file’s own timeline disproves it eleven lines earlier: “30 Jul — Anthropic discloses three incidents in its cybersecurity evaluations”, matching Anthropic’s own post (three incidents across six runs; three organisations compromised) and the dossier’s pin. One plus three plus one is five.

Why it matters more than a digit. Item 10 was opened for exactly this reason — so that “the census should hold the answer rather than a news outlet” — and the discharge closed it by importing the news outlet’s answer. The wrong number then propagated to four other files: the case, the frontier enterprises, known objections, and OPEN QUESTION 3 in the v3.5 queue. A count that travels is worse than a count that sits still.

Fix: the corpus states five incidents across three developers, names the composition, and explains that the outlets are each consistent about their own noun; the four dependent files are conformed. Status: cured.

E28 — “all self-disclosed,” in the same repository that argues the victim disclosed first

Filed 22 August 2026. Internal catch, same day, by the repository consistency audit.

The case introduced the summer’s incidents with the words “The incidents, all self-disclosed.” On the same day, known objections published the opposite as a load-bearing argument — that in the most consequential 2026 incident the victim disclosed first, Hugging Face having detected, contained and published its own forensic reconstruction before the developer said anything — and the press corpus graded that notification order as confirmed at the source, with Hugging Face disclosing on 16 July and OpenAI on 21 July.

Why it matters. The sentence gave away the strongest fact in the record. “All self-disclosed” concedes that voluntary disclosure worked; the truth is that it did not work in the first and largest case, which is the whole reason SEC. 9’s reporting clocks exist. A hostile reader who found the contradiction would have been entitled to ask which page the project believed.

Fix: the case now opens “The incidents — disclosed, but not all by the developers,” states the 16 July / 21 July order in the text, and carries a new paragraph on what the disclosure sequence proves. Status: cured.

E29 — an evaluator was placed behind an incident a prior correction had already removed it from

Filed 22 August 2026. Internal catch, same day. A 17 August correction overwritten by 22 August drafting.

OPEN QUESTION 3 in the v3.5 queue stated: “One testing vendor, Irregular, sits behind three of the four disclosed 2026 incidents … The escapes ran through the vendor’s misconfigured environment, not the developer’s own.” Known objections carried the same claim. But the dossier § A.4, expressly corrected on 17 August 2026, records the opposite of the OpenAI limb: “incident (1) did not — OpenAI’s chain ran through its own sandbox and a Modal customer’s harness, with Irregular named only in the reporting.”

Why it matters. CURE 7 Operation 4 — naming the auditor and evaluator into SEC. 4(c)’s non-shield list — is argued from this fact, so an overstated version of it weakens the drafting it supports. And the failure mode is the one this register is least willing to tolerate: a correction already made, in public, on the record, silently undone five days later by a new file written without reading it.

Fix: both files now state the tie precisely — Irregular’s environment is common to two of the three disclosing developers and four of the five disclosed incidents (Anthropic’s three, Meta’s one), while OpenAI’s chain ran through its own sandbox and a Modal customer’s harness — and cite the dossier correction. Status: cured. Standing consequence: a new file that restates a fact already graded elsewhere must cite the file that owns it, which is what the ownership rule now published in the verification record exists to enforce.

E31 — the register used one number twice

Filed 22 August 2026. Internal catch, by the repository consistency audit. A defect in the register itself, which is the one place this project cannot afford them.

E8 was assigned to two unrelated entries. The first, filed 20 August, is the “In one paragraph” correction — the front-page summary said the base duty is “not strict liability” while SEC. 6(c) and SEC. 10(a) make entity liability strict. The second, filed 19 August, is the Sacramento-scorecard correction — the front page claimed a table was “preserved verbatim in the diary” when the diary never carried it. Two different claims, two different dates, one number.

Why it matters. This register is the project’s only credential: it has no institution behind it, and it asks reviewers to trust a text on the strength of the fact that its mistakes are published with their fixes attached. A register that cannot count its own entries invites the obvious question about everything else it counts. E14 recorded and repaired exactly this failure mode for E13 and did not catch this instance, which means the check that found it — a full read of the register in sequence — was not being run.

Fix. The strict-liability entry keeps E8, because E12 cites it by that number in published text and a citation that has travelled should not be broken to tidy a sequence. The Sacramento entry is renumbered E30, carrying a note of its original number so a reader following an old link lands correctly. Nothing is deleted.

Two standing consequences. First, the next free number is now E32, and the register’s numbers are to be read as identifiers rather than as an ordering — filing order in this file is already non-monotonic for reasons several entries explain, and that is a feature of an append-only record, not a defect. Second, a sequence check joins the push routine: before any push that adds an entry, read the register’s numbers in order and confirm no collision and no gap. Status: cured.

E32 — E22 extended from the repository to correspondence

Filed 22 August 2026. Severity: rule change. The incident itself is outside this register’s scope and is not recorded here.

What happened, in the only terms this file can carry. In project correspondence, the maintainer described a named scholar’s published article as supporting a position the article does not take. The error was not one of transcription. The characterisation was drafted from a recollection of the article rather than from the article, and it inverted three of its load-bearing features: the body of law it belongs to, the defendant it runs against, and the thing its standard measures. It was retracted in writing to the same recipient, with the correction placed in the opening paragraph rather than a footnote.

Why it is filed here at all, given that nothing published carried it. A check of the tracked files was run and returned clean — no repository file cites the work, and no published claim rested on the mischaracterisation. So there is no published erratum to file, and the individual is not named, because the project’s standing rule keeps the names of people approached out of the repository. What is filed is the rule the incident produced, because that rule now governs work this register does cover.

The rule. E22 established that a quotation held in a working summary is not a quotation — that the confidence rubric grades how a source was obtained and has no column for how the words travelled from the source into the file. E22 was written about quotations in research files. The same defect operates on characterisations, and it operates outside the repository as readily as inside it. A remembered argument decays in the same interested direction a remembered quotation does: toward what the person remembering it needed it to say. Describing a scholar’s position as adjacent to your own is exactly the circumstance in which that pressure is strongest, and it is also the circumstance in which the error is most visible to the one reader who cannot miss it.

Accordingly, E22’s discipline is extended and restated in general form:

No text produced by this project — published, drafted, or sent — may characterise another person’s work unless that work is in hand at the time of writing. Not a summary of it, not a memory of having read it, not a citation to it in a third source. The work.

Consequence for the register. This is the second entry (E23 was the first) recorded not because something published was wrong but because the register should show the rules moving. The distinction is worth keeping visible: a register that only ever grows by failure teaches nothing about what the project does when it is working. Status: rule adopted; no repository text affected.

E33 — a filename is not a source either

Filed 23 August 2026. Severity: low — live for minutes, corrected the same hour, and filed because the register does not price errors by how briefly they were visible.

The error. The chapter outline published at commit a4aa8d4 described the FDA submission as “a real federal filing, published as filed.” It is a draft, published for criticism before filing; the docket closes 19 October 2026; the front page said so all along.

Cause. The description was written from the filename and the folder’s reputation, not from the file. That is the E32 failure operating on the project’s own repository within a day of the rule being written — which is worth recording precisely because it shows where the rule’s edge is.

Fix. Corrected in the consolidation that moved the chapters into the map; the outline’s standalone path was withdrawn from tracking minutes after creation, never having been mailed, with every link retargeted.

Rule sharpened. The work-in-hand rule extends inward: describe a file’s status only from the file’s own text — never from its name, its folder, or the author’s memory of writing it.

E34 — three of four, and the register number a page promised

Filed 24 August 2026. Severity: low — corrected in place the night it was found; this entry supplies the number the correction has carried as “candidate erratum, maintainer to number”.

The error. Comparative officer liability § 5 first mapped Lyness’s four-goal agenda (64 B.C. L. Rev. 253) onto the Act as a four-for-four convergence. It is three of four: Lyness argues for individual civil liability and “only civil liability” (at 297–99), expressly excluding the criminal form — the road this Act takes, with its due-care floor.

Cause. The mapping paragraph was committed before Part IV of the article had been read in full — reliance running ahead of reading, caught by the later sitting.

Fix. The page has carried the precision in its own § 5 addendum since the same night, convergence restated at three of four with the divergence argued, not assumed away; this entry numbers it. Maintainer ruling, 24 August 2026.

E35 — the Act’s own word, loosened in the Act’s own voice

Filed 24 August 2026. Severity: medium — the project’s central defined term, used loosely on the front page, live for two days.

The error. Three files — the front page’s dated record, the expanded timeline, and the press corpus — called the source of the 8 August admissions “a senior officer of the developer.” His displayed role at retrieval is head of strategic futures: the advice layer, which the front page itself excludes four paragraphs earlier (“technical work, access, advice … does not create personal liability”), and not an officer under the Act’s own final-material-authority test.

Cause. The defined term drifted toward the rhetoric: “a senior figure” lands softer than “a senior officer,” so the noun was upgraded to match the quote’s force — the project’s central vocabulary loosened in the project’s own voice, on its own front page.

Fix. All three files conformed to “the developer’s head of strategic futures” — which costs the argument nothing, since the quote’s weight is that it is the developer’s senior staff speaking, not that the speaker holds authority. And the accurate label is the stronger exhibit: the most candid account of the incident came from the layer the Act would not reach, while the layer with authority to halt said nothing on the record.

Rule sharpened. Defined terms may not be used in project prose in any sense looser than their definition — one owner per fact, extended to vocabulary. Caught in-house the same day, during a design review of the front page.

E36 — the first price in the genre, carried at double

Filed 24 August 2026, evening. Severity: medium — a dollar figure on the front page and in a reviewer packet, wrong by roughly a factor of two, for part of one day.

The error. Four public surfaces carried Colorado’s SB 26-189 fiscal note as “$100,403 General Fund, 0.8 FTE, $120,596 total,” dated 4 May 2026. The final revised note (6 May 2026), now in hand, states $46,190, 0.4 FTE (“in FY 2026-27 only”), and $56,286.

Cause. The figures entered ⚠ from reporting of the initial note and were relied on before the primary was retrieved — the ⚠ discipline working as designed, and the reliance running one day ahead of it. Whether the initial 4 May note carried the larger figures is unverified; the final revised note controls.

Fix. Front page, dated record, paths to enactment, the fiscal note § 6b, and the fiscal packet (via its builder) all conformed the same evening the primary arrived. The corrected floor is half the reported one — the argument the number serves got stronger.

Rule kept. Reliance follows retrieval; where it must precede it, the ⚠ travels with the figure and the primary is fetched before the figure is repeated a second time.

E37 — the same six days used for two different intervals, one of them wrong

The error. This project’s own pinned chronology for New York’s RAISE Act is PRINT NUMBER 6953A on 3 June 2025, PRINT NUMBER 6953B on 9 June 2025, passed both houses 12 June 2025. From those dates, the audit-and-signature provision at § 1421(4) existed for six days (3 to 9 June) and was struck three days before passage (9 to 12 June).

Several surfaces said instead that it was “struck at the B amendment, six days before passage” — the census twice, house language, the changelog, and the retrieval programme added the same day. One further surface said the bill “six days before it passed” carried the provision, which is a third interval again: 3 to 12 June is nine days. One number was doing the work of three, and only its original use was right.

Cause. The correct sentence, “it existed for six days”, was written first and is accurate. Later paraphrases reached for the striking number rather than the pinned dates, and nobody subtracted. The dates were in the same file the whole time, four hundred lines above.

Caught by. Preparing to pull the Senate floor transcript for the passage date. Establishing which date to search made the arithmetic unavoidable.

Why it mattered more than three days usually would. The wrong figure was in a drafted letter to Senator Gounardes, the bill’s surviving sponsor, who was present for all three dates. A cold approach that misstates the chronology of the recipient’s own bill does not get a second reading.

Fix. Every occurrence conformed to the pinned dates: the provision existed for six days and was struck three days before passage. The letter corrected before sending.

Rule kept, and it is a new one. An interval is a claim. Where this project states a number of days, the two dates it runs between are named in the same sentence or the sentence is rewritten until they are. A duration with no endpoints attached cannot be checked, and this one was repeated five times without anyone being able to.

E38 — the packet that promised the whole lane, and left out the only published criticism of it

The error. The criminal-law packet says of itself that it is “the lane’s whole apparatus inlined in reading order.” Since it was first built on 24 August 2026 it has not carried the one criticism of this lane that exists in print, by name, from a scholar this project cites: Sean Lyness, Revitalizing the State Environmental Responsible Corporate Officer Doctrine, 64 B.C. L. Rev. 253 (2023), argues the state doctrine should carry “individual civil liability—and only civil liability,” on the ground that Dotterweich and Park are misdemeanour authority decided “during a time when the immediate and collateral consequences were different” (at 297-98).

Where it was, and where it was not. The point was not suppressed. It has been in the table of authorities and in the comparative page since 25 August, and E34 already corrected the related overstatement. It was missing from the one page a criminal-law reviewer is told to read.

Why that is a defect and not a formatting slip. The packet exists so a reviewer does not have to hunt the repository. A packet that inlines the sweep’s own findings, inlines the drafted repairs, and omits the strongest published objection is a packet that flatters the lane. The seat was advertised to Sean Lyness himself — his letter has been drafted and is unsent — which would have put a man’s own argument in front of him with his own argument left out of the reading copy.

Caught by. Preparing the outreach to him, and re-reading what the packet actually contains before the letter went.

Fix. The objection is now stated in the sweep in the criminal section, in the project’s own words and against the project’s own text, and the packet is regenerated from it, so it appears in Part I where a reviewer meets it before the repairs. It is also added to the packet’s question menu as question 7, with the honest statement of where the lane’s answer stops: Park holds at the base tier and the sweep says so; nothing in this repository argues that the same authority reaches the felony tier at SEC. 6(b).

Rule kept. A packet that claims to be the whole lane owes the lane’s best objection, not only the lane’s own findings. Before any packet is sent to a named person, the repository is searched for that person’s own published position and, if it cuts against the Act, it goes in the packet first.

E39 — the same sentence twice, in two packets, for a day

The error. The criminal-law and enforcement packets ended with the filing instruction printed twice: “Or, if you were contacted by the maintainer through a different channel, reply on that channel. Or, if you were contacted by the maintainer through a different channel, reply on the channel you were contacted on.”

Cause. A sentence was added to the builders’ inline closing text rather than replacing the sentence already there. The two builders that carry the longer closing block took the duplicate; the six that carry the short one did not.

Caught by. Reading packets/build_criminal_packet.py end to end before editing it for E38.

Fix. One sentence, in both builders; all eight packets regenerated so the generated pages match their sources.

Rule kept. Generated pages are proofread as pages, not as diffs. A defect that only exists in the output of a script will not be found by reading the script’s change.

E40 — the council was described as five seats after it had grown to eight

The error. Two live surfaces described the review council as having five seats. The front page, at “Door two — review one lane”: “one of the five seats.” And the dossier’s recruitment paragraph, which went further and named the five roles: a criminal-law specialist, a former prosecutor or regulator, a frontier-security engineer, an open-source and academia reviewer, and someone who has administered a real budget.

The truth. The council runs to eight lanes and has since 23 August 2026: criminal law, enforcement, frontier security, fiscal and administration, federalism and preemption, proportionality and sentencing, torts and design, and open source and academia. There are eight packets, one per lane, built by eight committed scripts, and the reviewer page has said “eight lanes” on two separate lines throughout.

Cause, and it is the ordinary one. The count was raised where the work happens, on the reviewer page and in the packets directory, and the two places that only mention the council in passing were not swept. One of them, the dossier, is inside a folder whose sealed chapters are never edited, which makes it easy to skip; but the paragraph in question is in the folder’s live front matter, not the sealed text, so nothing prevented the fix except nobody looking.

Why it matters more than a number usually would. Both surfaces are recruitment copy. A prospective reviewer reading either one is being told, wrongly, how many seats exist and which five they are. Three of the eight lanes — federalism and preemption, proportionality and sentencing, torts and design — did not appear in that list at all, so a person qualified for one of those seats was being told there was no seat for them.

Caught by. The first full repository link-and-consistency sweep, 25 August 2026, run with a new committed tool, check_links.py.

Fix. Both corrected. The dossier paragraph now names all eight lanes and says in place that it named five when the chapters were sealed, with a pointer here. The sweep’s other findings are in the changelog.

Rule kept. A count that appears in recruitment copy is checked wherever it appears, not wherever it is maintained. The places that state a number in passing are exactly the places that go stale, because nobody edits them when the number changes.

E41 — three packets linked to a path the project’s own checker already knew was dead

The error. The criminal-law, enforcement and frontier-security packets each carried, in their opening orientation line, “the index of packets is one level up“. There is no packets/README.md. The index is packets/index.md. Three of the eight reading copies sent to reviewers therefore opened with a dead link in their second sentence.

What makes this worse than an ordinary broken link. The project already knew that path was dead. check_emails.py, the pre-send audit for correspondence, carries packets/README in its banned list under the label “packets/README (dead)”, so no outgoing email could contain it. The same string sat unnoticed in the repository’s own pages, because the email checker checks emails and nothing was checking the repository.

Cause. The three affected packets are the three built by the older “extraction” family of builder scripts, which share a longer orientation block. The path was correct when that block was written and was not revisited when the index moved.

Caught by. The same sweep as E40.

Fix. All three builders now point at ./index.md, and all eight packets were regenerated. Two further navigation defects found by the same sweep are fixed and recorded in the changelog rather than here, because they are omissions rather than false statements: the eight audit chunks were reachable from nowhere in the repository, and research/canon_check_2026-08-24.md was absent from the map that claims to record which file owns which question.

Rule kept. The repository gets the same pre-flight the post does. The link sweep is now a committed tool, check_links.py, and it is run before a push in the same breath as check_claims.py.

Corrections to the project contact; they enter the errata register with the fix attached and permanent credit.

E42 — the doctrine was said never to have left food and drug; it left decades ago, by act of Congress

Filed 25 August 2026. Internal catch, from a vocabulary audit of the library against the repository. The claim had stood on the front page since June.

What was published. Three pages carried a version of the same sentence. The README’s Overview: personal criminal exposure under the public-welfare doctrine “has never been extended past the food-and-drug frontier.” Questions and answers: the Park line “has simply never been extended past the food-and-drug frontier.” The glossary: “Eighty years old, never extended past the food-and-drug frontier.”

Why it is wrong, from a source this project already relied on. Lyness, Revitalizing the State Environmental Responsible Corporate Officer Doctrine, 64 B.C. L. Rev. 253 — the article the whole state-RCO argument is built on — records at n.33 that the Clean Water Act provides: “For the purpose of this subsection, the term ‘person’ means . . . any responsible corporate officer” (33 U.S.C. § 1319(c)(6)). At n.32 it quotes Copeland: “In the twenty years following the Park decision, the overwhelming majority of responsible corporate officer prosecutions were based on violations of environmental laws rather than the [Food, Drug, and Cosmetic Act].” And at n.118 it quotes United States v. Iverson, 162 F.3d 1015, 1024 (9th Cir. 1998): “In 1987, after the Supreme Court decided Park, Congress revised and replaced the criminal provisions of the CWA. . . . Congress made no changes to its ‘responsible corporate officer’ provision. That being so, we can presume that Congress intended for Park’s refinement of the ‘responsible corporate officer’ doctrine to apply.”

So the doctrine did not merely drift past food and drug. Congress legislated it into the Clean Water Act, and the courts have applied it there since at least 1998. The repository already cited § 1319(c)(6) in four files and Iverson in three. Nobody had read those against the front page.

Why it matters. It is the kind of overclaim a criminal-law reviewer catches in the first minute, on the most-read paragraph in the project, and finding it would have cost that reviewer nothing and cost this project its credibility on everything downstream. It also gave away a better argument than the one it made: a doctrine that has already moved once, by act of Congress, is a doctrine that travels. The honest claim is that it has never reached software — narrower, true, and stronger.

Fix: all three pages restated. The glossary’s entry now carries the CWA provision, the Copeland figure and the Iverson reasoning, so the correction is where a reader searching the doctrine will land rather than only in this register. Status: cured.

⚠ Read-status, stated so it is not mistaken for more than it is. The § 1319(c)(6) text, the Copeland sentence and the Iverson passage are all quoted from Lyness’s footnotes, not from the United States Code or the Federal Reporter. They are on the retrieval list. E22 governs: no outreach may describe them as verified until the primaries are read.

Method note. The audit that found this was not looking for it. It was counting which words a specialist reader would search for and failing to find. That is the second finding this month that came from asking what is absent rather than checking what is present.

E43 — a provision cited to a paragraph that does not exist, and an exception described as the prohibition

Filed 25 August 2026. Internal catch, same day, before the next chunk was written. The error was introduced today and published today.

What was published. The glossary’s new entry on exculpation and indemnification, added this afternoon, said: “SEC. 6(b)(5) bars indemnification and insurance for what the tagged text still calls ‘a knowing or wilful violation,’ once finally adjudicated.” The same wrong citation was carried in the changelog’s spelling entry earlier the same day.

Two errors, and the second is the worse one.

First, the citation. SEC. 6(b) has paragraphs (1) and (2). There is no SEC. 6(b)(5). The indemnification and insurance provision is SEC. 7. PERSONAL ECONOMIC CONSEQUENCES, subsection (b), and the language quoted sits at SEC. 7(b)(5). The phrase “under SEC. 6(b)” appears inside it as a cross-reference, which is how the mistake was made: a cross-reference read as a location.

Second, the substance. SEC. 7(b)(5) is not a bar. The bar is SEC. 7(b)(1)–(2), which prohibits insuring or reimbursing an individual penalty. SEC. 7(b)(5) is the carve-out from that bar: it permits insurance and advancement for “reasonable costs of defense, provided that amounts advanced or indemnified shall be repaid by a person finally adjudicated to have committed a knowing or wilful violation under SEC. 6(b).” The entry described the exception as the prohibition — a reader would have taken the Act to forbid the very thing that paragraph allows.

Why it matters, beyond accuracy. The entry was making a comparative argument: that this Act’s treatment of defence costs tracks a line Delaware already draws. With the provision read correctly the argument is stronger and more exact, because SEC. 7(b)(5) and 8 Del. C. § 145(e) share a mechanism rather than merely a mood — advance the costs, then claw them back on an adverse final adjudication. § 145(e) permits advancement “upon receipt of an undertaking . . . to repay such amount if it shall ultimately be determined that such person is not entitled to be indemnified.” The mistake cost the point its best form.

Fix: the glossary entry restated with the correct citation and the correct operation, and the comparison rewritten to name the shared mechanism. The changelog’s citation corrected. A held draft in the working library carried the same error and was corrected before it could enter the queue. Status: cured.

Method note, since this is the third finding today from the same habit. The wrong paragraph number was never checked against the statute; it was carried from an earlier reading of a passage that quoted the cross-reference. The project’s standing rule says a grep finds the owning file and never replaces reading it. The rule was written for exactly this and was not followed.

E44 — the tool built to protect quotations falsified fourteen of them

Filed 25 August 2026. Internal catch, same day, found while reading a packet for an unrelated reason. The damage was published for roughly four hours.

What happened. check_spelling.py, added this morning to hold the repository to American spelling, carries a rule stated in its own docstring: “A quotation from a British source must keep its own spelling.” It masked quoted spans so the sweep could not touch them. The mask matched only quotations that fit on a single line. Markdown wraps at ninety-odd characters, so most quotations of any length wrap, and the mask silently did not apply to them.

Fourteen quotations were altered. Among them:

UK Health and Safety at Work etc. Act 1974, s.37, reproduced in comparative officer liability: “If the offence is proved to have been committed with the consent or connivance of—” became offense. A British statute, quoted verbatim, rewritten into American spelling.
Four passages from the UK AI Safety Institute’s incident report, including “there is good reason to think near-impossible tasks push models towards more ‘creative’, and more transgressive, problem-solving” — where towards became toward. This one is the sharpest rebuke available: the tool’s own docstring names AISI’s towards as the example of why the mask exists, and the mask then failed on that exact sentence.
The tagged statute itself, quoted in the torts and design packet: “finally adjudicated to have committed a knowing or wilful violation under SEC. 6(b)” became willful, so the packet quoted the Act as saying something the Act does not say.
Two quoted UK government passages, a quoted objection, and a quoted definition of specification gaming.

Why the design was wrong, not just the regex. The sweep ran line by line. A per-line mask can never see a span that crosses a line, whatever pattern it uses, so this was not a tuning error that a wider regex would have fixed. The tool has been rewritten to compute protected ranges over the whole file before touching anything.

Fix: all fourteen restored from a pre-sweep snapshot, verified by re-running the detector that found them, which now reports none altered. The masking rewritten whole-file. Status: cured.

A limitation accepted on purpose. Quote marks alternate, so text sitting between two quotations now reads as quoted and is protected too. That leaves a little of the project’s own prose British. Under-changing is recoverable; falsifying a source is not, so the error is left pointing that way, and it is written into the script beside the rule.

Why this entry is longer than the defect. Every claim in this repository is checkable only because quotations are exact. A tool that rewrites them, silently, in bulk, while its own documentation promises it will not, is the most damaging class of failure the project has had — worse than a wrong number, because a wrong number announces itself to anyone who checks and a quietly corrected quotation does not. It was found by accident. There is now a detector for it, and it should be run after any bulk edit.

E45 — the glossary said the Act was silent on willful blindness; the Act uses it, once, against the wrong person

Filed 25 August 2026. Internal catch, four hours after publication, on the first end-to-end reading of the statute this maintainer had done all day.

What was published. The glossary’s willful blindness entry, added this afternoon: “The tagged text does not mention it. That is a gap, not a position.”

What the tagged text says. SEC. 2(b), the reliance path for non-modifying deployers: reliance “is unavailable to a deployer that knows, or consciously avoids knowing, of a material nonconformity in the adopted validation or in the deployed configuration.”

The doctrine is in the Act. It appears exactly once. And it appears against the smallest actor the Act reaches. A downstream deployer forfeits its safe course for deliberate ignorance; a controlling person of the developer faces nothing of the kind, because SEC. 6 says nothing about it.

Why the correction is worth more than the error cost. “The Act is silent” invited a cure that adds something new. “The Act uses it once, against the deployer and not the developer” is a different and better finding: an asymmetry that runs the wrong way on any reading, sitting in the tagged text, which CURE 22 now corrects rather than merely supplements. The entry has been restated on that footing.

Fix: the glossary entry rewritten to quote SEC. 2(b) and state the asymmetry. Status: cured.

Method note, and it is the third entry today from one cause. E43 cited a paragraph that does not exist because a cross-reference was read as a location. E44 let a tool rewrite fourteen quotations because a mask was built line by line. This entry claimed the Act was silent on a doctrine it uses.

All three were written by someone who had read this statute in pieces and never in one sitting. Six hundred and eleven lines, and the sections quoted most confidently today — SEC. 4, 5, 7, 8 — were read closely while SEC. 0, 1, 2, 6, 10, 11, 12 and 13 were grepped. The project’s standing rule is that a grep finds the owning file and never replaces reading it. That rule was being applied to files and not to the instrument itself.

The reading is now done. It produced this erratum in its first ten minutes.

E46 — two quotations published this afternoon are not in the opinions they were attributed to

Filed 25 August 2026. Found by doing the thing the ⚠ flags said had not been done: reading the primaries. Published for roughly five hours.

What was published

Three sections added this afternoon — the glossary’s corporate entries, the “Corporate law already answers this” objection, and the shelf beneath both — carried ten Delaware quotations. Every one was marked unverified, and the read-status blocks said in terms that they came from a retrieval reply rather than from the reporters. That discipline is the only reason this entry is a correction and not a disaster.

Ten quotations, read in the opinions on 25 August 2026:

Quotation	Result
Stone v. Ritter, the Caremark conditions predicate	✅ verbatim, slip op. 17
Stone, “known duty to act… conscious disregard”	✅ verbatim, slip op. 17
Stone, “failure to act in good faith may be shown”	✅ verbatim, slip op. 15
Massey, “Delaware law does not charter law breakers”	✅ verbatim, slip op. 46
Massey, “lawful business” by “lawful acts”	✅ verbatim, slip op. 46
Marchand, “mission critical”	✅ verbatim, slip op. 36
Marchand, the three things that did not exist	✅ verbatim, slip op. 32
In re McDonald’s, officers owe a duty of oversight	⚠ misquoted
In re TransUnion, “make a business judgment to… break the law”	❌ not in the opinion
Walton, “conscious decision to violate the law”	❌ not in the opinion

The two that do not exist

We published: In re TransUnion Derivative Stockholder Litigation, 324 A.3d 869, 887 (Del. Ch. 2024): “What a corporate fiduciary cannot do, however, is make a business judgment to cause or allow the corporation to break the law.”

The sentence does not appear in that opinion. What the opinion does contain, at slip op. 30, is the Massey language — “Delaware law allows corporations to pursue diverse means to make a profit, subject to a critical statutory floor” — which the source had separately attributed to Massey. So the same passage was given twice, once correctly and once as a sentence nobody wrote.

⚠ Half withdrawn, 26 August 2026. The sentence is not a fabrication. It is Vice Chancellor Laster’s, at slip op. 76 of the Walton demand-futility opinion of 26 April 2023: “What a corporate fiduciary cannot do, however, is make a business judgment to cause or allow the corporation to break the law.” What survives is that it was attributed to the wrong case — and whether it also appears in In re TransUnion needs TransUnion, which this project does not hold. A misattribution is a different and lesser error than a sentence nobody wrote. See E62.

We published: Ontario Provincial Council of Carpenters’ Pension Trust Fund v. Walton, 294 A.3d 65, 90, 92 (Del. Ch. 2023), that the rule protects a decision that “carries legal risk, but which otherwise involves legally compliant conduct,” and that proceeding unlawfully “would constitute a conscious decision to violate the law.”

Neither sentence appears. The opinion discusses “a conscious decision to prioritize profits over compliance,” which is a different proposition.

⛔ THIS ENTRY’S WALTON FINDING IS WITHDRAWN, 26 August 2026. Both quotations it declared fabricated are verbatim in the opinion. They are in the 26 April 2023 demand-futility opinion, at slip op. 76 and slip op. 77–78; this entry checked the 12 April 2023 laches opinion, which is a different document under the same case number. See E60. An earlier note added here the same morning, reading the laches opinion and explaining how the “fabrication” had been assembled from a sentence at slip op. 2, is withdrawn with it: it was a plausible reconstruction of something that never happened. The In re TransUnion finding in this entry is untouched and stands — it has not been re-checked and nothing here casts doubt on it.

The one that was misquoted

We published: “This decision confirms that officers owe a duty of oversight,” at 289 A.3d 343, 361.

The opinion says, at slip op. 2: “This decision clarifies that corporate officers owe a duty of oversight.” Two words wrong and the page wrong.

Also corrected: the Marchand monitoring sentence was published as “a good faith effort to put in place a reasonable system of monitoring and reporting about the corporation’s central compliance risks.” The opinion says, at slip op. 31: “the board must make a good faith effort — i.e., try — to put in place a reasonable board-level system of monitoring and reporting.” The published version dropped board-level, which is the qualifier that makes In re McDonald’s significant.

What this cost the argument, and it is not nothing

The two missing quotations were the centrepiece of the answer. The section asserted that “Delaware forecloses it in its own words” and that “the business judgment rule cannot be a defense to this Act, and it is Delaware that says so.” That claim rested on two sentences that do not exist.

Massey survives and is verified, so the answer survives in a weaker and more honest form: the business judgment rule sits above a statutory floor, and a statute sets the floor. What Delaware has not squarely said is that the rule can never be raised against a duty imposed by a statute outside the DGCL. The section now says that, and says it is open.

Fix: both fabricated quotations removed; the McDonald’s and Marchand quotations restated from the opinions with slip pages; the seven verified quotations marked ✅ with the date they were read; the section’s claim narrowed to what the surviving authority supports. Status: cured.

The rule this proves, and it was already written down

E22 requires a quote in hand. A quote in a reply is not a quote in hand.

The retrieval reply that supplied these was the best of three received today: it carried four explicit CANNOT VERIFY rows, it marked every other row VERIFIED or SECONDARY, and it gave pincites to the page. It marked both fabricated quotations VERIFIED, sourced to “opinion itself.” The calibration signals were real and the calibration was wrong, which is the harder case to catch.

Seven of ten were exact. That is a good hit rate and it is the reason the failure is dangerous: a source that is right most of the time trains you to stop checking. The ⚠ blocks are what made this recoverable, and every claim added today that has not yet been read still carries one.

E47 — three page numbers, from a source that has no page numbers, caught before publication

Status: caught, not published. Recorded because the near-miss is the finding.

What nearly went in. The retrieval debt on United States v. MacDonald & Watson Waste Oil Co., 933 F.2d 35 (1st Cir. 1991) was to be paid by reading the opinion. A first pass over it reported all three passages present and reported them at different pages than this repository publishes — 51, 45 and 50, against the published 55, 51 and 52 n.15. The next step was to correct five files to the new numbers.

What the second pass found. The source carries no star pagination at all. It reproduces the opinion as continuous text. The three page numbers were not read out of the document; they were produced by the process that read it. Had the correction gone in, this project would have replaced three sourced pincites with three invented ones and marked the row verified while doing it — strictly worse than the debt it was paying.

What is actually now known. Both operative sentences are verbatim in the opinion. Footnote 15 is the willful-blindness footnote. The pages are not confirmed and remain the secondary source’s. The row keeps its ⚠ for that reason.

The rule, which E22 did not already cover

E22 says a quote must be in hand. It says nothing about the pincite, because until today the two always arrived together. They do not.

E47 — text and page are two claims and are verified separately. A source without star pagination can confirm a quotation and can never confirm a pincite. Where a page cannot be confirmed, the published pincite stays as it was, attributed to whoever supplied it, and the ⚠ stays with it.

And the second-order rule. The first pass returned pages that did not match the repository’s, which read as a correction and therefore as evidence of careful reading. A reported discrepancy is not evidence of retrieval. It is the single most persuasive form an invention can take, because agreeing with what you already hold is what a lazy answer looks like — so disagreement gets trusted by default. Ask what the page marker looked like, in its own characters, before believing a page.

E48 — a published quotation with two words the court did not write, and both were ours

Status: found and corrected the same day, by reading the opinion.

What we published. United States v. Iverson, 162 F.3d 1015 (9th Cir. 1998), in CURE 22 and in the criminal packet, as a block quotation:

“The government still had to prove that the discharges violated the [CWA] and that defendant knew that the discharges were pol[lutants].”

What the opinion says:

“The government still had to prove that the discharges violated the law and that defendant knew that the discharges were pollutants.”

Two alterations, and they are different in kind. “pol[lutants]” is nonsense — a bracket around nothing, of the sort produced when a passage is reassembled rather than transcribed. “[CWA]” is worse: the court wrote a general word, the law, and the published version replaced it with a specific statute. Square brackets are a promise that the substitution is the editor’s and is faithful. Here they narrowed the court’s language and then advertised the narrowing as an editorial courtesy.

Nothing in the argument turns on either word, and that is precisely why it went unnoticed through the sweep, the cure, the packet build and two rounds of review. A quotation that supports your point is not checked for the words that do not.

Two more defects in the same passage, found by the same read

The Park ratification paragraph was cut before its most useful sentence. We published the paragraph as ending “…to apply.” The opinion ends it “…to apply under the CWA,” and, earlier in the same paragraph, carries a parenthetical we had elided entirely: “(Most importantly, Congress made a violation of the CWA a felony, rather than a misdemeanor.)”

That parenthetical is the bridge to the objection the criminal lane most fears. Ahmad argues CWA discharges cannot be public welfare offenses because they are felonies. Iverson states the felony grade in a parenthetical and applies Park’s responsible-officer refinement anyway. The project had both cases and could not see they touched, because the sentence joining them had been cut out of our own copy.

And the whole of Iverson’s stated test was missing. “[A] person is a ‘responsible corporate officer’ if the person has authority to exercise control over the corporation’s activity that is causing the discharges. There is no requirement that the officer in fact exercise such authority or that the corporation expressly vest a duty in the officer to oversee the activity.” Read against SEC. 4(a), this Act is narrower than the federal standard a circuit approved — an answer to the overbreadth objection that was sitting inside a case we had cited eleven times and never read.

The rule

E48 — an elision is an edit. What a quotation leaves out is published too.

E22 asks whether the words in the quotation marks are the court’s. It does not ask what stood either side of them. Every passage carried at second hand in this repository has been through someone else’s choice about where to stop, and that choice was made for their argument, not ours. When a quotation is finally read in the source, read the paragraph around it: in this instance the sentence before the quotation and the clause after it were each worth more to this project than the quotation itself.

E49 — a “finding” the repository had already made three days earlier, caught at the door

Status: caught, not published. Recorded because it is the fourth of its kind today.

What was about to be written. Reading the Guidelight control assessment sent the maintainer back into SEC. 8 and SEC. 9, and two things looked like discoveries: that SEC. 9(b)’s constructive-notice limb is inert for the first [180] days because SEC. 9 commences on the effective date while SEC. 8 commences at day 180; and that measuring the counterfactual against “the monitoring the entity certified it maintains” rewards an entity for certifying thin monitoring. Both were about to be opened as a new OPEN QUESTION.

Both are CURE 14, opened by the lane sweep on 22 August, which states them as defects (ii) and (iii) of three, in those words, and drafts the repair. The second one is in the cure’s title: “a detection clock that cannot be gamed by certifying less monitoring.”

How it was missed. The statute was searched. The cure register was not. v3_5_cure_language.md carries a numeric index of every cure by name in its own second paragraph, and the answer was sitting in a title.

The rule

E49 — the register is part of the text. Before opening a question, read the index of what is already open.

This project’s three instruments — the statute, the sweep and the cure register — are searched as if the first were the source and the other two commentary. They are not. A defect this repository has already found is not findable by reading the statute, because the statute still contains it.

What survived the check, and it is smaller and better than what was nearly claimed. CURE 14 frames the certification incentive as gaming. CURE 15 frames the nonconformity report as punishing candour. Nobody had placed them side by side: through SEC. 8 and SEC. 9 the Act punishes honesty about the system and rewards honesty about the watching, and repairing either alone leaves the asymmetry standing. That observation is now in both cures and on the flow page — a paragraph of synthesis rather than the new finding it was going to be announced as.

E50 — the spelling sweep erased the erratum that records the spelling sweep erasing things

Status: applied to the working tree, caught in the diff, reverted before commit.

How it became possible. ledger/errata.md opens at line 7 with  and no closing END marker anywhere in the file. The sweep’s sealed-region pattern ran a seal to the matching END or to end of file, so a single unterminated marker sealed 1,599 of this register’s 1,605 lines, including every entry written since 19 August. The register had never once been swept. The same defect sealed 160 lines of the diary.

What happened when the hole was closed. With the seal made honest, the sweep proposed 43 substitutions across eight files. Fifty-six of the changed lines were in this register, and one of them was E44:

We had written: “…where towards became toward. This one is the sharpest rebuke available: the tool’s own docstring names AISI’s towards as the example of why the mask exists.”

The sweep produced: “…where toward became toward.”

The tool deleted the record of its own previous failure, and misquoted AISI a second time inside the sentence that exists to explain why AISI must not be misquoted. E44 was recoverable in August because the erratum named the words. After this run it named nothing.

The rule, and it is a category the project did not have

E50 — an errata register is a document about words. Much of it quotes words as specimens rather than using them, and no normaliser can tell mention from use. It is corrected by hand or not at all.

ledger/errata.md and ledger/diary.md are now in the sweep’s SKIP_FILES beside the drafting record and the dossier. They are excluded for a different reason from the other two: those are sealed because they are frozen, these because their content is evidence.

And the second-order lesson, which is the more useful one. The seal hole was not found by reading the file. It was found by asking why a checker’s own machinery had a branch nobody had tested. A silent skip and a clean pass are indistinguishable in the output, and this one printed 0 substitutions for six days while a fifth of the repository went unread. The sweep now prints *** UNTERMINATED SEAL *** with the file, the line, and the number of lines left unprotected, and protects nothing on a broken marker’s say-so.

Not fixed, and it is a maintainer question. The BEGIN marker at line 7 is still unterminated. Closing it requires deciding where the merged content ends, and the honest answer may be that a register appended to daily was never sealed content in the first place and the marker should go.

E51 — a pincite copied from a citing case, from the wrong one of its four citations

Status: published in audit/v3_5_cure_language.md and its packet since 26 August 2026; corrected the same day on reading the cited case.

The claim. CURE 24 wrote that Hanousek holds the CWA’s criminal provisions to be public welfare legislation, “resting on United States v. Weitzenhoff, 35 F.3d 1275, 1283 (9th Cir. 1993).”

The source. Hanousek cites Weitzenhoff four times, at three different pages. At 1283, for the standard of review: “presents a question of statutory interpretation, which we review de novo. See United States v. Weitzenhoff, 35 F.3d 1275, 1283 (9th Cir.1993).” At 1282-83, for the facts of the case. At 1286, for the sentence actually wanted: “The criminal provisions of the CWA constitute public welfare legislation. See Weitzenhoff, 35 F.3d at 1286.” And at 1286 n.7, for the penalty argument.

What went wrong. The pincite was lifted from the first of the four, because it was the first one a search returned, and attached to the proposition carried by the third. 1283 is a real page of a real case cited for a real proposition. It is not this proposition’s page. Nothing in the form of the citation showed the defect: the reporter is right, the volume is right, the case is right, the year is right, and the number is a page the citing court genuinely wrote down.

The fix. 1283 corrected to 1286 in the cure language and in packets/criminal_law.md, with the Weitzenhoff row rewritten from the opinion itself.

The rule

E51 — a pincite borrowed from a citing case is a claim about that case’s citation, not about the source. Before taking it, find the citing court’s own sentence and confirm it states the proposition you are about to attach the number to. A case cited more than once has more than one page, and only one of them is yours.

This is E47’s sibling. E47 says a source without pagination can confirm a quotation and never a pincite. E51 says a source with pagination can confirm the wrong one.

And the second-order defect, which is the more expensive of the two

The file that would have settled this was on the shelf, correctly retrieved, and labeled as useless. It was saved on 26 August as RECORD_9Cir_US-v-Weitzenhoff_35-F3d-1275_amended-opinion_PARTIAL-star-pagination.pdf. The copy in fact carries continuous star pagination from *1279 to *1299. The label came from a check that found the first marker, \*1279, and did not ask whether there were more.

Then the label traveled. A retrieval brief issued that morning recorded Weitzenhoff as “only one star marker (*1279) - not enough to settle pincites at 1283 or 1286,” and the repository proceeded on that basis for a day, including in the sentence this erratum corrects. A warning label is a claim. It was written by the same process that writes the claims it warns about, and it was never checked by the process that checks them.

Both defects have the same shape: a search returned a first hit and the first hit was treated as the whole answer. E51’s checkable form is therefore one question, asked twice - is this the only one?

What the corrected read produced, and it is larger than the correction

Weitzenhoff at 1286 n.7 holds that Staples “refrains from holding that public welfare offenses may not be punished as felonies,” which is the answer to Ahmad, the objection the lane sweep had recorded as unanswerable. The finding was one page away from a citation this repository had published, and the wrong page number is what kept it there.

E52 — a quotation made a row stop being a row, and the sweep counted it as absent

Status: found the same hour it was created, on 26 August 2026. Both checkers repaired.

What happened. The Weitzenhoff row was written into standards/table_of_authorities.md carrying a citation exactly as the reporter prints it: "511 U.S. at ----, 114 S.Ct. at 1804". The four hyphens are the West convention for a U.S. Reports page that had not issued when the opinion went to print, and E48 requires them to be reproduced rather than tidied away.

check_citations.py decided the row was a table separator and skipped it. The test was:

elif line.startswith("| ") and "---" not in line and not line.startswith("| Authority"):

Three hyphens anywhere in a row disqualified it. The row was therefore not counted among the rows, not listed, and its flags were not counted in the standing debt. The sweep printed rows in the table ...... 150 while the file held 151, and reported a debt of 22 against a true 23. check_claims.py carried the same idiom for the state-of-play table.

How it was caught, and this is the only reason it was. The row was expected to appear in the debt list and did not. Nothing in the output said a row had been dropped; the numbers were simply smaller, and a smaller debt figure looks like progress. Had the row not been one whose absence was being watched for, this would have been a silent undercount of the repository’s own reading debt.

The repair. A separator row is now identified by what it actually is - a row whose every cell consists only of hyphens, colons and space:

_SEP_CELL = re.compile(r"^[\s:-]+$")

def is_separator_row(line):
    cells = [c for c in line.strip().strip("|").split("|")]
    return bool(cells) and all(_SEP_CELL.match(c) for c in cells)

Applied in both checkers. The sweep now reads 151 rows and 22 flagged, the second figure having also been corrected by the second defect below.

The rule

E52 — a structural test written against a substring is a test against content. Markdown structure is positional, and a checker that infers it from characters will be defeated the first time a quotation contains those characters. Test the shape, not the spelling.

This is the third member of the family E50 opened: a silent skip and a clean pass are indistinguishable in the output. E50’s seal ran to end of file, E51’s label was believed because nobody asked whether the first hit was the only one, and E52’s row vanished into a count that went down. All three were invisible in a report that said nothing was wrong.

And the second defect in the same row, which is a repeat offense

The Weitzenhoff row was written with three ⚠ markers used as annotation - one flagging an unavailable pincite, one introducing the dissent, one introducing a correction - on a row whose reading is complete. check_citations.py counts any row containing ⚠ as unread, so the row would have declared itself unread the moment it became visible.

This is the same mistake made three times in two days. It was made on the § 3663A row, caught; made again on both interim-standard rows, and reported to the maintainer as a debt of 22 when the truth was 19; and made a third time here. The markers are removed and the row’s residual debt now lives where it belongs - on the Staples row, which is genuinely unread.

⚠ in this table is a read-status flag and has no other meaning. It is not emphasis, not a caution, and not a way to point at something. If a read row needs to say something urgent, it says it in bold.

E53 — two rows for one authority, twice, added by the person who had just written the rule against it

Status: found within the hour by a check written for the purpose. Both pairs merged.

What happened. On 26 August 2026 rows were added to standards/table_of_authorities.md for Staples v. United States and for Liu v. SEC, each described in the row itself as “cited in the repository’s prose and never graded until now.” Both already had rows — Staples under “Culpability and elements,” Liu under “Penalties and proportionality” — and neither was checked for first.

Why it matters more than tidiness. A row in this table is a read-status. Two rows for one document are two read-statuses that drift, and nothing on either says which is current. Staples briefly stood as both “unread, the pivot of the whole felony question” and “the modern presumption of scienter where penalties are severe,” with no flag at all.

This is E49 again, on the same day E49 was cited in another correction. E49’s rule is that the register is part of the text and the index of what is already open is read before a question is opened. A table of authorities is exactly such an index. Knowing the rule and citing the rule did not produce the two-second grep that the rule asks for.

The repair, and it is the useful part. check_citations.py now detects two rows carrying the same authority, and distinguishes the accident from the intended case: a pair is allowed when one of the two rows points at the other in words — the pattern already used for Iverson, where one row carries the reading and the other says where the read-status lives. A pair with no pointer is reported. The two pre-existing pairs it found on its first run, Iverson and 33 U.S.C. § 1319(c)(6), were both real: the § 1319(c)(6) pair had no pointer and its two rows had already drifted, one recording the statute’s “means” verbatim and the other paraphrasing it as “includes.”

The rule

E53 — a table of authorities is an index of read-statuses, and a second row for one authority is a second read-status. Before adding a row, grep for the party name. Where two rows are wanted, one of them must say where the reading lives.

And a second checker defect found in the same pass

check_citations.py reported Liu v. SEC as cited-with-no-row while its row sat three lines away. Its matcher tests whether either party name appears in the table and skips any party shorter than five characters to avoid noise. “Liu” and “SEC” are both shorter than five characters, so that case could never match, no matter how many rows it had. Any case with two short party names was permanently unmatchable and permanently reported as debt. A guard against false positives had manufactured a false positive that could not be cleared by doing the work it demanded. The matcher now falls back to the whole caption when every party name is short.

E54 — a case caption written in the form such captions usually take, from a source that gave no caption

Status: published in standards/table_of_authorities.md on 26 August 2026; corrected the same day.

The claim. A row reading In re Alibaba Group Holding Ltd. Securities Litigation (S.D.N.Y., filed 4 Aug 2026), flagged ⚠ unread, “known from The D&O Diary.”

The source, which this project holds and which was re-read to write this entry. The D&O Diary piece names no case. It says “A securities class action complaint filed on August 4, 2026, in the Southern District of New York against Alibaba Group Holding Limited (Alibaba) and the company’s CEO,” and thereafter calls it “the Alibaba SCA.”

The truth. Wistisen v. Alibaba Group Holding Limited, No. 1:26-cv-06654 (S.D.N.Y., filed 4 Aug. 2026) — caption, docket number, court and filing date confirmed against the CourtListener RECAP index on 26 August 2026.

How the wrong caption got written. The court and the date were in the source. The caption slot was empty, and it was filled with the form such captions usually take: In re [Company] Securities Litigation. That form is correct for a consolidated securities class action. This one is a single named plaintiff and is not consolidated, so the invented caption is wrong in the one respect that would matter to anyone trying to find it.

Nothing marked it. The ⚠ said unread, which was true. A read-status flag cannot say “this caption was never in a source,” and the row carried an attribution — “known from The D&O Diary” — that made the caption look sourced when only the surrounding facts were.

The rule

E54 — a citation has parts, and they are sourced separately. Court, date and docket may come from a report; the caption comes from the docket or it does not exist. Where a slot is empty, the entry says it is empty. A conventional form is not a source, and filling a slot from convention is E46 with better manners.

E55 — a citation taken from the first of two footnotes that contradict each other twelve lines apart

Status: published since 25 August 2026; corrected 26 August 2026.

The claim. Kadrey v. Meta Platforms, Inc., No. 23-CV-03217-VC, 2025 WL 1752484 (N.D. Cal. June 25, 2025), taken from Maxwell V. Pritt’s Written Answers to Questions for the Record at printed p. 93 of S. Hrg. 119-202. The row was honest about the chain: quoted at second hand, read by OCR, “the pin cite is his, not ours.”

What is actually on that page. Both of Pritt’s first two footnotes cite the same case, and they give different numbers:

¹ Kadrey, et al. v. Meta Platforms, Inc., No. 23-CV-03217-VC, 2025 WL 1752484, at *2 (N.D. Cal., June 25, 2025).

² Kadrey, et al. v. Meta Platforms, Inc., No. 23-CV-03417-VC, Dkt. 574 (Pls’ Mot. for Partial Summary Judgment) at 8.

Read at 200 dpi, re-read at 450 dpi, and finally looked at as an image rather than as text, because a 2 and a 4 are exactly what an OCR pass gets wrong and the whole question was one digit. The two footnotes genuinely differ. CourtListener’s RECAP index resolves it: 3:23-cv-03417, N.D. Cal., filed 7 July 2023.

So this is not an OCR error, ours or anyone’s. It is a typo in a document submitted to the Senate, reproduced faithfully by a project that took the first citation it found and did not read twelve lines further down the same page — where the source corrected itself.

The rule

E55 — when a citation is taken at second hand, read every place the source cites that authority, not the first. A source that cites something twice has checked itself, and the check is free.

Second-order, and it is the reason this was findable at all. The correcting footnote was discovered only because an outside retrieval returned a different docket number and the difference had to be explained. A disagreement with an outside source is worth more than an agreement, and this project’s response to one should always be to open the page rather than to decide which side is more likely right.

E56 — a holding published inside out: the court’s confirming reason reported as its ratio

Status: published in the cure register, the lane sweep and both packets since 25 August 2026; corrected 26 August 2026 on reading the opinion.

The claim, in the form it appeared in four files. “United States v. Ahmad, 101 F.3d 386, 391 (5th Cir. 1996) holds that illegal discharges under the CWA are not public welfare offenses, because they are ‘felonies punishable by years in federal prison’.”

The opinion, read 26 August 2026 in a law.resource.org reporter capture. The sentence the “because” was built from reads:

“The fact that violations of § 1319(c)(2)(A) are felonies punishable by years in federal prison confirms our view that they do not fall within the public welfare offense exception.”

A reason offered as confirming a view already reached is not the reason the view was reached. The court’s ground is mistake of fact: whether “knowingly” attaches to the nature of the substance discharged, so that “one who honestly and reasonably believes he is discharging water may find himself guilty of a felony if the substance turns out to be something else.”

And a second thing the reading corrected. This project had been describing Ahmad and Weitzenhoff as a circuit split — including in E51, written the same morning. Ahmad says the opposite about its own relationship to that case: the Ninth Circuit “was concerned almost exclusively with whether the language of the CWA creates a mistake-of-law defense. Both cases are easily distinguishable, for neither directly addresses mistake of fact or the statutory construction issues raised by Ahmad.” A narrower disagreement does survive — over what Staples decided at 618 — and that one is real.

The rule

E56 — “because” is a claim about a court’s reasoning, and it is verified in the opinion or not made. A case known only through a summary may be reported as holding something; it may not be reported as holding it for a reason.

This is E32 applied to the inside of a case rather than its outside: no characterization until read, and the ordering of a court’s reasons is a characterization. The published version was sharper than the opinion, which is the tell. A secondary summary compresses toward the memorable sentence, and the memorable sentence in Ahmad is the one about felonies.

Cost, stated plainly. For a day the criminal lane treated its own hardest objection as resting on a ground the case does not rest on, and built two cures around answering it. The answers survive — Weitzenhoff at 1286 n.7 still refuses Ahmad’s reading of Staples — but the objection they were answering was not quite the one in the reports, and a criminal-law reviewer would have said so in the first paragraph of a disposition.

E57 — a new provenance grade, because three sources arrived by a route this register had no name for

Not a correction. A rule entry, filed the day the category appeared, so that nothing enters the repository through it unlabeled.

What happened. Three sources that repeated retrieval runs could not reach — Ind. Const. art. 1, § 16, Walton, and Philip Morris — were reached on 26 August 2026 by a fetch tool that downloads a page and has a small language model answer a question against it. It returned the Indiana clause, Walton’s full caption and court, and the Philip Morris respondeat superior passage: all three had defeated a script, and none of them defeated this.

The problem is that the register has two grades and this is neither. A source is held — the document is on the shelf, and someone can open it — or it is not held. This is a third thing. A model read the document and told us about it. That is better than a secondary summary, because the model had the primary in front of it. It is worse than reading, because the thing that reaches the repository is the model’s rendering, and nobody has seen the page.

E22 already decides the hard part. A quotation held in a working summary is not a quotation, and a model’s answer is a working summary — it is generated text about a document, which is the exact category E22 was written to exclude. So nothing arriving this way may be published as a quotation. What may be taken is the kind of fact that survives rendering: a caption, a court, a docket number, a date, a page count, whether a copy carries star pagination.

The rule

E57 — a model-mediated fetch is a lead, not a reading. It may fix metadata and it may not supply a quotation. It is marked ◐, it always carries a ⚠ beside it, and the file it produces says MODEL-MEDIATED-FETCH-NOT-THE-DOCUMENT in its own name.

The mark is now in this table’s legend, and check_citations.py reports any ◐ row that does not also carry a ⚠, because a row graded half way and counted as graded is E52 in a new costume.

Two things worth keeping from the three fetches

One. The route matters, and it is not the same route a script takes. Philip Morris was reached through CourtListener, which an earlier run that same morning reported behind a bot challenge, and the Indiana clause came from a legislature PDF whose sibling URL returns an empty application shell. “Blocked” is a property of a request, not of a site, and this register should stop recording it as though it were the second.

Two. A collision found while pinning a clause, recorded before it bites. The companion cites Or. Const. art. I, § 16 and Ind. Const. art. 1, § 16 in a single sentence. Same section number, same opening words, three words apart at the end: the companion gives Oregon as “all penalties shall be proportioned to the offense,” and Indiana reads “all penalties shall be proportioned to the nature of the offense.” Oregon’s text is still unpinned. Whoever pins it must pin it against Oregon’s own publication, because the near-identity is precisely the shape of mistake this shelf keeps making with surnames, and it would be invisible in a proofread.

Added the same day: the clearest demonstration this rule will ever get

Hours after E57 was written, two fetches of the same Walton opinion, through two different renderings of it, disagreed about the document’s own front page. The Delaware courts’ own service returned 26 April 2023, 58 pages. A mirror’s PDF of the same opinion returned 12 April 2023, 56 pages, with a confident summary of what the case was about. Two further pages settle it — the opinion’s caption block reads “Date Submitted: January 13, 2023 Date Decided: April 26, 2023” — so the second rendering was simply wrong.

Nothing in either answer looked uncertain. Both gave a date, a page count and a subject, in the same register of confidence. The disagreement was only visible because the same document was fetched twice, which is not something a retrieval does by default and is not something a reader would think to do.

So the rule tightens. Where a ◐ source supplies a fact that will be published — a date, a docket, a page count — fetch it twice, from two renderings, and record both. Where they disagree, neither is the answer; the document is.

What the three fetches did not do. None of them put a document on the shelf. The Indiana constitution PDF, the Walton opinion and a clean Philip Morris print are each one deliberate download away, and until someone makes it, the strongest thing this project can say about all three is that it knows where they are. E47’s discipline applies unchanged: the Philip Morris pincite at 1118 is no closer to confirmed than it was yesterday.

E58 — a citation that exists only in a URL and a page title is not a citation, and it nearly overwrote a correct one

Status: caught before any change was made. Filed because the near miss was one keystroke wide.

What happened. While looking for a clean print of United States v. Philip Morris USA Inc., a search result appeared whose page title and URL both carried the case at 556 F.3d 1095. This repository publishes it at 566 F.3d 1095. One digit, and the outside source is the kind that usually wins an argument: a public-health litigation tracker, a specialist body with no reason to be careless.

A structural argument pointed the same way and was not enough. Volume 556 of the Federal Reporter covers early 2009 and volume 566 covers late spring; the opinion was decided 22 May 2009, which fits

That is a rule of thumb about publication schedules, not a source, and this register does not correct citations from rules of thumb in either direction.

What settled it. The Solicitor General’s own petition in the case, which states in its OPINIONS BELOW section: “The opinions of the court of appeals (Pet. App. 1a-98a, 99a-176a) are reported at 566 F.3d 1095 and 396 F.3d 1190.” The repository’s citation was right.

Where the 556 came from is the interesting part. It appears in the tracker’s page title and URL slug — the parts of a web page that are written once, by hand, usually years ago, and never proofread against anything, because nothing renders them next to the text they describe. The body of that page does not repeat the citation at all.

The rule

E58 — a citation appearing only in a URL, a page title, a filename or a link label is not a citation. Those are labels, written once and never checked against the document they name. Before changing a citation this project publishes, find it in a document that a court, a party or a publisher would be embarrassed to get wrong.

And the second-order habit this reinforces, because it is the third one-digit dispute in a single day. E51 was a page number taken from the wrong one of four citations; E55 was a docket number taken from the first of two footnotes that disagreed; this was a volume number taken from a URL. All three were single digits, all three looked exactly like every other digit on the page, and none of the three would have been caught by reading more carefully. They were caught by noticing that two sources disagreed and going to a third. Disagreement is the instrument. Care is not.

E59 — two opinions, two weeks apart, under one case number, and a date I confirmed from three sources that the document contradicts

Status: published this morning in standards/table_of_authorities.md; corrected the same day when the document arrived.

What was published. That the Walton opinion’s own caption block reads “Date Submitted: January 13, 2023 Date Decided: April 26, 2023,” and that the subject is the Rule 23.1 demand-futility ruling. It was written as a confirmation, on the strength of three sources agreeing: the Delaware courts’ own opinion service, a Delaware firm’s case note, and Justia’s case page, which carries the date as a field.

What the document says. The PDF, downloaded and opened:

OPINION ADDRESSING DEFENDANTS’ MOTION TO DISMISS ON THE BASIS OF LACHES Date Submitted: January 13, 2023 Date Decided: April 12, 2023

Sixty-four pages, not the 56 or 58 that two separate model-mediated fetches reported.

Both things are true, and that is the finding. There are two opinions in C.A. No. 2021-0827-JTL, two weeks apart: a laches opinion on 12 April 2023, which is the one now on the shelf, and a Rule 23.1 demand-futility opinion on 26 April 2023, which is the one the firm note describes. Justia’s case page carries the later date and serves the earlier opinion’s PDF. Nothing in any of the three sources said “one of two.”

What this costs, and it reaches backwards

This repository cites Walton as 294 A.3d 65, 90, 92. Which of the two opinions is reported at 294 A.3d 65 is now unsettled, and so is the question of which one E46 read when it found two quotations fabricated. E46’s finding survives — the invented sentences are absent from the opinion now held, and the reading also showed exactly which real sentence they were assembled from — but an erratum that says “not in the opinion” has to name the opinion, and until today nobody knew there was a choice to make.

The rule

E59 — a case number is not a document. Where a matter has produced more than one opinion, a citation that names only the case names nothing, and agreement among secondary sources about “the” opinion is agreement about a thing that may not exist. Before citing, ask how many opinions there are.

And the third strike against confirming anything by consensus

Three sources agreed on 26 April and the document says 12 April. They did not agree because they had checked; they agreed because they were describing the other opinion, and nothing in their phrasing distinguished the two. This is the same shape as E58, where a citation lived only in a page title, and the same shape as E55, where a witness’s two footnotes disagreed. Agreement among sources that are all downstream of one another is not corroboration. The document is the only thing that ends the question, and this register has now recorded that lesson four times in one day, which suggests the lesson is not the problem.

One thing that went the other way, recorded because the register should not only collect failures

The Philip Morris passage that the same model-mediated route returned — corporations liable for specific intent on the “knowledge and intent” of their employees, because “a corporation only acts and wills by virtue of its employees” — is in the opinion, word for word, confirmed on reading the document. E57’s caution is not that these fetches are wrong. It is that nothing in the answer tells you which kind you have got.

And a false negative worth admitting, because it nearly produced a fifth erratum in the wrong direction: a first search of the document for that passage returned zero hits and briefly looked like proof the quotation was invented. The passage was there. The search failed because the text wraps mid-phrase across two lines and the document uses curly quotation marks. A grep that returns nothing is evidence about the grep.

E60 — the register’s most-cited fabrication finding was itself wrong, and a real quotation was deleted on the strength of it

Status: found 26 August 2026 on reading the second of two opinions. The correction runs the opposite way from every other entry in this register: something true was removed as false.

What E46 said. That two sentences published in docs/known_objections.md and attributed to Ontario Provincial Council of Carpenters’ Pension Trust Fund v. Walton, 294 A.3d 65, 90, 92 (Del. Ch. 2023) “do not appear” in the opinion. The passage was removed and replaced with a sentence saying Delaware has not squarely decided the question.

What the opinion says. Both sentences are there, word for word.

At slip op. 76:

“When directors make a business decision that carries legal risk, but which otherwise involves legally compliant conduct, then the business judgment rule protects that decision.”

At slip op. 77–78, across a page break:

“In the former case, the directors can make a business judgment to pursue the project. In the latter case, the decision to pursue the project would constitute a conscious decision to violate the law, the business judgment rule would not apply, and the directors would be acting in bad faith.”

The published sentence quoted both correctly. Nothing was invented.

How a correct quotation got deleted as a fabrication

C.A. No. 2021-0827-JTL produced two opinions two weeks apart — a laches opinion on 12 April 2023, 64 pages, and a demand-futility memorandum opinion on 26 April 2023, 123 pages. E46’s check was run against the first. Neither sentence is in the laches opinion, and both are in the other one. Nothing in the retrieval, in the case name, in the citation, or in three secondary sources that describe “the” opinion said there was a choice to be made. E59 records the discovery that there are two; this entry is what that discovery cost going backwards.

What it cost

A published answer to the objection a governance lawyer raises first — can the business judgment rule be raised against this Act? — was replaced with “that question is open” and a note blaming the project’s own fabricated citation. For a day, this repository told reviewers it had asserted something on the strength of quotations nobody wrote, when it had quoted a Delaware Vice Chancellor accurately and pincited him correctly.

And the deterrent effect is the worse half. Walton was flagged in the table of authorities as the case carrying two fabricated quotations, with a standing instruction that nothing be attributed to it. That instruction was wrong and it was the loudest instruction in the row.

The rule

E60 — a finding that a quotation is absent is a claim about a document, and it names the document or it means nothing. “Not in the opinion” requires knowing which opinion, and how many there are. Absence is only ever provable against a specific text; it is never provable against a case.

The asymmetry that makes this rule matter. A false positive — publishing a quotation that turns out to be invented — is caught by the next person who reads the source. A false negative is not caught by anything, because the quotation has been deleted and nobody re-checks a sentence that is no longer there. This one surfaced only because an unrelated question about a date sent someone to the other opinion.

Two searches that returned nothing and meant nothing

Both quotations were initially reported absent from the 26 April opinion too, by grep, before being found by eye. “carries legal risk” matched only because the phrase happens not to wrap; the longer sentence returned zero hits because it spans a page break, with a page number and two blank lines sitting inside it. Earlier the same day the Philip Morris passage returned zero hits because it wraps mid-phrase and uses curly quotation marks.

Three false negatives from three different causes in one session. E47 says a source without pagination cannot confirm a page. This is its inverse: pagination inside a text file breaks the sentences it paginates, and a search over extracted text is searching a document that has had furniture inserted into it.

What is not fixed

The reporter pincites 90 and 92 are still unconfirmed. The copy now held carries slip pages, not Atlantic Reporter pages, and slip-to-reporter offsets are not computable — the two data points available happen to differ by sixteen, which is a coincidence and not a mapping. Under E47 this copy confirms the quotations and cannot confirm the pages.

Part I(b) — The negative-findings register

Opened 26 August 2026 after E60, in which a correct quotation was deleted from a public page because a search for it returned nothing.

Why this is a separate register. Every other entry above is additive: it says “we published X, X is wrong, here is the fix,” and it quotes X, so the entry carries its own evidence and any later reader can re-test it. A finding that something is absent carries no evidence. The text is gone, nothing points at it, and nobody re-checks a sentence that is no longer there. That asymmetry is why E46 stood for a day and could have stood for years: a false positive is caught by the next reader; a false negative is caught by nothing.

So negative findings are listed, not merely recorded, and check_quotations.py --negatives re-tests every line here against the shelf on every run. When a new source lands, the register answers for itself.

# status | quoted string (verbatim, as published) | attributed to | erratum
MISATTRIB| What a corporate fiduciary cannot do, however, is make a business judgment to cause or allow the corporation to break the law | Walton slip op. 76 (real); credited to In re TransUnion, 324 A.3d 869, 887 | E46 -> E62
WITHDRAWN| carries legal risk, but which otherwise involves legally compliant conduct | Walton, C.A. 2021-0827-JTL (Del. Ch. 26 Apr. 2023) | E46 -> E60
WITHDRAWN| would constitute a conscious decision to violate the law, the business judgment rule would not apply, and the directors would be acting in bad faith | Walton, C.A. 2021-0827-JTL (Del. Ch. 26 Apr. 2023) | E46 -> E60

A WITHDRAWN line stays here forever. It is the record of a finding this project got wrong in the direction nothing else catches, and deleting it would be deleting the only evidence that the failure mode is real.

E61 — how a negative finding is made

A finding that a quotation is absent is made by reading the cited location, not by searching.

Enumerate the documents first. A case number is not a document; one matter can produce several opinions (E59). State how many candidates exist and which was read.
Go to the pincite, not to the string. “I read page 90 and it says X” is checkable by anyone. “I searched and found nothing” is checkable by no one.
Where the location cannot be resolved — no pincite, or a copy whose pagination cannot settle one (E47) — the finding is “unverifiable from this copy.” That is a weaker and different claim from “absent,” and it does not justify deleting anything.
Name the file. The shelf filename, so the check can be re-run against the same bytes.
A search that returns nothing is evidence about the search. Extracted text has furniture inserted into it: page numbers land inside sentences, words hyphenate across line breaks, quotation marks curl. All three defeated a search on 26 August alone.

And the tool obeys the same rule. check_quotations.py can conclude that a quotation is on the shelf. It can never conclude that one is absent — a miss is printed as a prompt to read, and the report says so on every run. A quotation-checker that reported absences would recreate E60 at scale, which is the one thing this project cannot afford to automate.

E62 — the third “fabrication” was a misattribution, and the instrument built to catch this found it on its first run

Status: found 26 August 2026, about twenty minutes after check_quotations.py was written, by the tool’s own re-test of the negative-findings register.

What E46 said. That this sentence, published attributed to In re TransUnion Derivative Stockholder Litigation, 324 A.3d 869, 887 (Del. Ch. 2024), “does not appear in that opinion.”

Where the sentence actually is. Walton, demand-futility opinion of 26 April 2023, slip op. 76:

“What a corporate fiduciary cannot do, however, is make a business judgment to cause or allow the corporation to break the law. ‘Delaware law does not charter law breakers.’ In re Massey Energy Co., 2011 WL 2176479 (Del. Ch. May 31, 2011).”

E46 was right that it is not in the opinion it was credited to, and wrong about what that means. The sentence is Laster’s own, in a case this project cites, on the proposition it was wanted for. A misattribution is not a fabrication. One is a citation error; the other says no court wrote the thing. This register called the first the second and let it stand.

Still open: whether the sentence also appears in TransUnion. That needs TransUnion, which is not held. It is a retrieval now, not a conclusion.

The rule

E62 — “not in the opinion it was credited to” and “nobody wrote it” are different findings, and the second is far larger. Before recording a fabrication, search the whole shelf for the sentence. A real sentence in the wrong coat is the commonest citation error there is, and it is repaired by fixing the attribution, not by deleting the argument.

What found it, which is the part worth keeping

check_quotations.py was written this afternoon in answer to E60, on one design rule: it can conclude that a quotation is on the shelf and can never conclude that one is absent. Its --negatives mode re-tests every recorded negative finding against the shelf on every run, because what E60 exposed is that nobody re-checks a sentence that has been deleted.

Three findings were in the register when it first ran. It printed:

*** WAS RECORDED ABSENT AND IS ON THE SHELF ***
    What a corporate fiduciary cannot do, however, is make a business judgment  (E46)
    found in: RECORD_DelCh_Ontario-Carpenters-v-Walton_...DEMAND-FUTILITY-2023-04-26...
    Read the document. This is how E60 was found.

An hour earlier, having withdrawn two of E46’s three findings, this project explicitly declined to touch the third — “it has not been re-checked, and nothing in the Walton correction bears on it.” That was the right call on the evidence then available and it was wrong on the facts. The instrument, which knows nothing and cannot read, caught in one second what careful restraint got wrong. That is not an argument against restraint. It is the argument for putting negative findings somewhere a machine re-tests them, because judgment does not re-examine what it has already deliberately set aside.

E63 — “every case cited in prose now has a read-status” was false, and the checker that said so was matching on the words “United States”

Status: reported to the maintainer three times on 26 August 2026 and written into research/verification_record.md and a retrieval brief. Corrected the same day.

The claim. That check_citations.py showed zero case captions cited in the repository’s prose without a row in the table of authorities — announced as the closing of Tier 4, the nineteen captions that had never been graded.

The defect. The matcher asked whether either party name of a caption appeared anywhere in the table’s text. So Johnson v. United States matched, because dozens of rows say “United States.” Lambert v. California matched on “California.” Any case with one common party name was unmatchable-as-missing, permanently, no matter how absent its row.

What it hid. Lambert v. California and Johnson v. United States — the two leading vagueness cases, cited on the objections page against this Act’s own duty language, with no row and no read-status. The vagueness objection is the one that most often kills a bill in committee, and its two authorities were invisible to the instrument built to find exactly this.

The repair. A party name now counts as evidence only where some single row carries both parties. And trim_second_party stops at a sentence boundary, because captions were bleeding into the following sentence — “Ulster County v. Allen. Extend the” — and then matching no row, which reported real rows as missing in the other direction.

The rule

E63 — a test that asks whether a token appears somewhere in a corpus is not a test about the row that should contain it. Membership questions are answered against the record, not against the file.

This is E52’s family again — a structural question answered by substring — and the third time in one day that a checker reported a clean number while not looking at part of what it claimed to cover.

What was deliberately not done

The sweep now reports three captions with no row that do have rows: abbreviation mismatches and one caption joined across an “and”. Those are false positives and they are being left in.

Tuning a checker until it reports zero is the failure this register spent the day recording. A false positive costs somebody a look. A false negative cost this project a published claim that was not true, repeated three times, in two documents. The error is left pointing toward noise.

E64 — a view the Supreme Court declines to adopt, published as the rule it adopted

Status: published in standards/table_of_authorities.md on 26 August 2026, hours before the case was read. Corrected on reading it.

What was published. That Staples v. United States, 511 U.S. 600, holds at 618 that serious felonies fall outside the public welfare offense exception “absent a clear statement from Congress that mens rea is not required.” The line was taken from Ahmad, which quotes it that way, and recorded here as one of two pincites the case supplies.

What page 618 actually says, read in the Library of Congress U.S. Reports print, in one continuous passage:

“Close adherence to the early cases described above might suggest that punishing a violation as a felony is simply incompatible with the theory of the public welfare offense. In this view, absent a clear statement from Congress that mens rea is not required, we should not apply the public welfare offense rationale to interpret any statute defining a felony offense as dispensing with mens rea. But see United States v. Balint, 258 U. S. 250 (1922).

We need not adopt such a definitive rule of construction to decide this case, however. Instead, we note only that where, as here, dispensing with mens rea would require the defendant to have knowledge only of traditionally lawful conduct, a severe penalty is a further factor tending to suggest that Congress did not intend to eliminate a mens rea requirement.”

“In this view” opens it and “we need not adopt such a definitive rule” closes it. The sentence is the antecedent of the reservation — the Court states a possible rule in order to decline it, and attaches a But see to its own counter-authority in the same breath.

What this does to the leading objection

Ahmad, 101 F.3d 386, 391 quotes the declined view as though it were the holding, and this project repeated it. The two sentences Weitzenhoff and Ahmad fight over turn out to be the same passage on the same page, and read whole it favours Weitzenhoff:

The Court declines the felony-incompatibility rule in terms.
It calls a severe penalty “a further factor”, not a bar. Ahmad treats it as decisive.
It cites its own counter-authorities: Balint in the text, and at 617 n.14 State v. Lindberg, 125 Wash. 51 (1923), “applying the public welfare offense rationale to a felony.”
And Balint is the case Hanousek rests SEC. 6(a)’s due-process answer on, already read and held here. The Supreme Court flagged it as the answer to the felony objection thirty-two years before this project independently arrived at it.

The operative trigger is not the penalty at all. It is “where, as here, dispensing with mens rea would require the defendant to have knowledge only of traditionally lawful conduct.” Owning a gun is traditionally lawful. So the real question for this Act was never “can a felony be a public welfare offense.” It is whether training and deploying a frontier model is traditionally lawful conduct — a narrower question, and one a reviewer can actually answer.

The rule

E64 — a proposition a court states in order to reject it reads exactly like a proposition a court holds. Before quoting a sentence from an opinion, read the sentence after it. Signals like “in this view”, “it might be suggested”, and a But see attached to the court’s own contrary authority are the tell, and they are invisible in a quotation taken at second hand.

And the second-order point, which this register has now made four times in one day. The error entered because the sentence arrived through Ahmad rather than from the reports. E22 says a quotation held in a working summary is not a quotation; a quotation held in an adversary’s brief is not a holding, and an adversary has every reason to stop reading at the sentence that helps.

E65 — Global-Tech called a constitutional ceiling in three files, and it decides no constitutional question

Status: published in standards/what_these_words_mean.md, packets/criminal_law.md and audit/v3_5_cure_language.md. Corrected on reading the opinion, 26 August 2026.

What was published, in identical words in all three files:

“The constitutional ceiling is Global-Tech Appliances, Inc. v. SEB S.A., 563 U.S. 754, 769 (2011).”

The pincite is right and the characterization is wrong. Read in the govinfo U.S. Reports print, which carries real reporter pagination, page 769 carries the two-part test:

“While the Courts of Appeals articulate the doctrine of willful blindness in slightly different ways, all appear to agree on two basic requirements: (1) The defendant must subjectively believe that there is a high probability that a fact exists and (2) the defendant must take deliberate actions to avoid learning of that fact.”

That is not a constitutional holding. Global-Tech is a civil patent case under 35 U.S.C. § 271(b). The passage is the Court’s distillation of what the Courts of Appeals already agree on — “all appear to agree” — supported by a footnote collecting one case from each circuit. No constitutional question is presented, argued or decided. The word “constitutional” appears in this project’s sentence and nowhere in the Court’s.

The one constitutional touch in the opinion is a different doctrine. At 767 the Court notes that it has used the Model Penal Code’s definition of knowledge “as a guide in analyzing whether certain statutory presumptions of knowledge comported with due process,” citing Turner v. United States, 396 U.S. 398, 416–417 (1970) and Leary v. United States, 395 U.S. 6, 46–47 and n.93 (1969). A due-process limit on a statutory presumption is not a constitutional ceiling on a judicially administered doctrine, and the two were merged somewhere between reading about the case and writing about it.

What is true, and is worth more to this Act than what was claimed

The test is a ceiling in substance, by ordinary stare decisis rather than by the Constitution, and it binds in the direction SEC. 6(b) needs. At 770 the Court rejects the Federal Circuit’s looser standard in terms — it “departs from the proper willful blindness standard in two important respects”: it allowed knowledge on “merely a ‘known risk’”, and demanding only “deliberate indifference” it “does not require active efforts by an inducer to avoid knowing.” The Court fixes willful blindness above recklessness and negligence: a willfully blind defendant “can almost be said to have actually known the critical facts,” while “a reckless defendant is one who merely knows of a substantial and unjustified risk.”

And the answer to the criminal-versus-civil question runs the other way

The question asked of this case was whether the Court limits willful blindness in criminal as against civil contexts. It does the reverse. At 766: “The doctrine of willful blindness is well established in criminal law.” At 768, having traced that history: “we can see no reason why the doctrine should not apply in civil lawsuits for induced patent infringement under 35 U.S.C. § 271(b).” The movement is criminal outward into civil, and the opinion contains no sentence narrowing the doctrine for criminal cases.

Which leaves a smaller problem in place of the one that was claimed. The authority this Act cites for the criminal knowledge element of SEC. 6(b) is a civil case, describing criminal practice accurately but deciding a civil question. That is a fair citation and it should be made in those words, not dressed as a constitutional limit.

The rule

E65 — “constitutional” is a claim about what a court decided, not an intensifier for how firmly it said it. A rule can bind without being constitutional, and describing it as constitutional attributes to a court a question it was never asked. Before writing that an authority sets a constitutional limit, name the constitutional provision and the party who raised it.

The pincite survived the characterization, and that is the part to notice. E47 separates text from page as two claims verified separately. This entry adds a third: the proposition a pincite is offered for is a claim too, and a correct page number lends it a borrowed credibility it has not earned.

E66 — half of Johnson & Towers published as the whole of it, and the missing half is the half that bears on SEC. 6(d)

Status: published in standards/table_of_authorities.md and audit/v3_5_cure_language.md (CURE 22). Corrected on reading the opinion, 26 August 2026.

What was published. That United States v. Johnson & Towers, Inc., 741 F.2d 662 (3d Cir. 1984) holds, as “the Third Circuit’s outlier rule”, that the jury must find the defendant “knew a permit was required and that none was held”. CURE 22 carries it as the Third Circuit’s side of a split.

That much is verbatim in the opinion, in the body of Part III.B:

“in light of our interpretation of section 6928(d)(2)(A), it is evident that the district court will be required to instruct the jury, inter alia, that in order to convict each defendant the jury must find that each knew that Johnson & Towers was required to have a permit, and knew that Johnson & Towers did not have a permit.”

The next sentence was not published, and it is the court’s own qualification:

“Depending on the evidence, the district court may also instruct the jury that such knowledge may be inferred.”

And Part IV states the holding with the qualification built into it:

“we conclude that the individual defendants are ‘persons’ within section 6928(d)(2)(A), that all the elements of that offense must be shown to have been knowing, but that such knowledge, including that of the permit requirement, may be inferred by the jury as to those individuals who hold the requisite responsible positions with the corporate defendant.”

The knowledge requirement and the route around it are one sentence, and this project published the first clause. The court says as much itself at the head of Part III.B: its conclusion “does not impose on the government as difficult a burden as it fears.”

What this does to CURE 22

CURE 22’s new text for SEC. 6(b) ends with a sentence written as a concession against the Act’s own convenience:

“Responsibility and authority under SEC. 6(d), standing alone, do not establish knowledge.”

On the sentence Johnson & Towers actually holds, the Third Circuit permits the inference that sentence declines to permit — knowledge “may be inferred by the jury as to those individuals who hold the requisite responsible positions with the corporate defendant.” The two are not in contradiction: a permissive inference a jury may draw is not the same as a matter established as a matter of law, and SEC. 6(b) may be drafted more narrowly than RCRA if that is the choice. But the concession was drafted as though the authority compelled it, and it does not. It is a policy choice, and it should be defended as one.

The second half of the question, which the row does not answer

The rule is confined to the subsection and is not stated generally. The court’s own words tie it to the construction it had just performed — “in light of our interpretation of section 6928(d)(2)(A)” — and Part IV repeats the confinement. Carrying it as “the Third Circuit’s rule” on knowledge of a legal requirement, unqualified, is broader than the opinion.

The pincite, and a third copy the shelf did not know it had

669 remains unconfirmed, and the reason is not the one the row gives. Three copies are held:

Copy	Pagination
law.resource.org reporter capture	none — continuous text, paragraph-numbered
FindLaw capture	none — the reporter pages are absent
OpenJuris capture, filed as `_second-copy`	*star pagination, `664`** — and one page only, “Page 1 of 1”

The only copy on the shelf that carries star pagination is a one-page capture that stops five pages short of the pincite. Under E47 669 stays the secondary source’s and the ⚠ stays with it.

And the filename is the finding. _second-copy says nothing about what distinguishes this copy from the other two — that it is a truncated single page, and that it is the only one with real pagination. CLAUDE.md records that a filename is a claim written by the same process that writes the claims it warns about, and that nothing checks them. This one understated a capability and concealed a truncation in the same word.

The rule

E66 — a holding that states a requirement and then states how it may be satisfied is one holding. Quoting the requirement and stopping is not an elision of detail; it reverses the balance the court struck, because the second clause exists to answer the objection the first clause invites.

This is E64’s neighbour rather than its repeat. E64 is a court stating a proposition in order to reject it. Here the court states the rule and means it — and qualifies it in the following breath, so the quotation is accurate, the attribution is correct, and the reader is still misled about what the case does. Reading the sentence after is not only a test for whether the court believed it.

E67 — Bank of New England misdated by three years, and its most-quoted sentence belongs to a district court in West Virginia

Status: published in standards/what_these_words_mean.md, docs/known_objections.md and standards/table_of_authorities.md. Corrected on reading the opinion, 26 August 2026.

What was published, in the glossary and in the same form in the other two files:

United States v. Bank of New England, N.A., 821 F.2d 844, 856 (1st Cir. 1984): “a corporation cannot plead innocence by asserting that the information obtained by several employees was not acquired by any one individual who then would have comprehended its full import. Rather the corporation is considered to have acquired the collective knowledge of its employees and is held responsible for their failure to act accordingly.”

Three things are wrong with that, and the quotation is not one of them

First, the year. The opinion was argued March 4, 1987 and decided June 10, 1987. It is 821 F.2d 844 (1st Cir. 1987). The library filename has carried 1987 since the file arrived; three published files carried 1984. The date most likely migrated from the conduct — the charged transactions run from May 1983 to July 1984 — and no one asked which was which.

Second, and this is the substance: the sentence is not the First Circuit’s. It is a block quotation, and the line immediately following it in the opinion gives the source:

“United States v. T.I.M.E.-D.C., Inc., 381 F. Supp. at 738.”

The words are those of the United States District Court for the Southern District of West Virginia, 1974, quoted with approval by the First Circuit. This project published a district court’s sentence as a court of appeals holding, in a glossary entry defining the doctrine.

Third, “at 856” cannot be checked. The copy held is a law.resource.org capture with no star pagination, as its filename says. Under E47 the pincite stays the secondary source’s and the ⚠ stays with it.

What the First Circuit does say in its own words, which is stronger than what was quoted

“A collective knowledge instruction is entirely appropriate in the context of corporate criminal liability. … Corporations compartmentalize knowledge, subdividing the elements of specific duties and operations into smaller components. The aggregate of those components constitutes the corporation’s knowledge of a particular operation.”

And, upholding the charge: “Since the Bank had the compartmentalized structure common to all large corporations, the court’s collective knowledge instruction was not only proper but necessary.”

The instruction the court approved is quoted in the opinion and this project has never carried it:

“In addition, however, you have to look at the bank as an institution. As such, its knowledge is the sum of the knowledge of all of the employees. … So, if Employee A knows one facet of the currency reporting requirement, B knows another facet of it, and C a third facet of it, the bank knows them all.”

The tension this was expected to expose is not there, and a different one is

The question put to this case was whether a broadly stated collective-knowledge doctrine sits badly with SEC. 4, which locates authority in one natural person. It does not, and the reason is in the first line the court wrote about it. Collective knowledge is a rule of corporate criminal liability: knowledge of employees “is imputed to the corporation,” and the aggregate “constitutes the corporation’s knowledge.” The doctrine has no operation against a natural person, and no case cited in the string supporting it aggregates the knowledge of several people onto one defendant.

So docs/known_objections.md and the glossary both overstate the Act’s own modesty. They say the Act “declines to aggregate”, which reads as a choice to give up an available tool. The tool was never available against an individual. The Act is not declining collective knowledge; it is operating in the register where the doctrine does not reach. That is a smaller claim and a true one, and it survives a reader who knows the case.

And a theory in the same instruction that the repository has never mentioned

The trial judge gave the jury a second route to willfulness, which the First Circuit also upheld:

“the bank as an institution has certain responsibilities … you will have to determine whether the bank as an organization consciously avoided learning about and observing CTR requirements. The Government to prove the bank guilty on this theory, has to show that its failure to file was the result of some flagrant organizational indifference.”

That is willful blindness at the level of the organization, with the jury directed to weigh “the bank’s effort, if any, to inform its employees of the law; its effort to check on their compliance; its response to various bits of information”, and its policies against “how it carried out its stated policies.” It is the entity-level analogue of what SEC. 6(b) does to a person, it is approved appellate authority, and this repository has cited Bank of New England five times without it.

The rule

E67 — a quotation inside a block quote belongs to whoever the citation under it names. An appellate court quoting a district court with approval makes that sentence persuasive, not its own, and the difference is the difference between a First Circuit holding and a 1974 district court in West Virginia. Where a quotation is followed by a bare citation rather than by more of the court’s prose, that citation is the attribution.

And the smaller rule, which cost less but recurs. A case’s date is a fact about the opinion, not about the conduct. Where a published year sits inside the span of years the facts cover, check it — 1984 was in the indictment, and that is exactly why it looked right.

E68 — Caremark can settle its own pincite after all, and the sentence before the famous one answers the objection the famous one was raising

Status: the pincite question is resolved in the repository’s favour; two elisions and one characterization are corrected. Read 26 August 2026.

The provenance caveat was right about the document and wrong about the pagination

standards/table_of_authorities.md and research/verification_record.md both said of the copy held — a Thomson Reuters/Westlaw reprint with KeyCite headers, hosted by Penn Carey Law — that “whether it may confirm 971 is a question for whoever reads it.”

It confirms it. The reprint carries star pagination throughout, *960 to *972, and the oversight passage falls between the *971 and *972 markers. 971 is confirmed from the copy on the shelf.

The distinction worth keeping. “Not an official court print” is a claim about whose text this is — the reprint is Thomson Reuters’ rendering, not the Atlantic Reporter’s own — and it remains true. It is not a claim about whether page boundaries are marked, and the filename’s caveat was read as though it were both. A provenance limit and a pagination limit are separate properties of a copy, and this shelf has been treating “unofficial” as implying “unpaginated” since the file arrived.

The quotation is verbatim, and two things were dropped from it

standards/what_these_words_mean.md carries it exactly as the opinion has it. The opinion opens the clause with three words this project drops without an ellipsis:

“Generally where a claim of directorial liability for corporate loss is predicated upon ignorance of liability creating activities within the corporation, as in Graham or in this case, in my opinion only a sustained or systematic failure of the board to exercise oversight . . . .”

“In my opinion” is Chancellor Allen marking the standard as his own view, in the paragraph that became the most-cited passage in Delaware oversight law. Under E48 an elision is an edit, and this one removes the author’s own hedge from a sentence this project offers as settled doctrine.

And the standard is confined, in the same sentence, to the branch of the case it decides — claims “predicated upon ignorance of liability creating activities”. It is not a general standard for director oversight, and both files present it as one.

The reservation two sentences earlier, which the objection never had to survive

docs/known_objections.md builds the corporate-law objection on this passage: Delaware “set the bar high on purpose”, so a state criminalizing the same conduct at a lower threshold “has overridden a considered judgment about how much protection a decision-maker needs.”

Chancellor Allen expressly declines to decide the case this Act is about. At 971, in the paragraph immediately preceding:

“Thus, this case presents no occasion to apply a principle to the effect that knowingly causing the corporation to violate a criminal statute constitutes a breach of a director’s fiduciary duty. See Roth v. Robertson, 64 Misc. 343, 118 N.Y.S. 351 (N.Y. Sup. Ct. 1909); Miller v. American Tel. & Tel. Co., 507 F.2d 759 (3d Cir. 1974).”

The considered judgment Delaware made was about ignorance. It was not made about knowing violation, and the court said so while making it — and cited two authorities going the other way on the reserved question. SEC. 6(b) is a knowing-conduct offense. The objection, at the strength this page gives it, does not reach SEC. 6(b) at all, and the answer has been sitting two sentences above the quotation since the objection was written.

That does not dispose of the objection against SEC. 6(a), whose floor is a failure of due care and which really does sit below the Caremark bar. The page should make that division rather than answering for the whole Act.

The rule

E68 — a caveat in a filename names one limit, and a reader will generalize it to every limit that sounds like it. “Not an official print” was allowed to mean “cannot settle a page” for as long as the file sat unopened. Check the property you actually need against the document, because the label warns about a different one.

And the reservation is part of the holding. E64 covers a proposition a court states in order to reject it. This is a court stating what it is not deciding, next to what it is — and the sentence a project quotes is worth less than the sentence saying which cases it governs.

E69 — the instrument built to prevent E60 could not see the shelf, and said so in a line nobody read

Status: found and fixed 26 August 2026, while running the checkers before a hand-off. Nothing was published wrong; the guarantee simply was not running.

What happened. check_quotations.py, written this afternoon in answer to E60, resolves the shelf from a single hardcoded path:

LIB = os.path.expanduser("~/mnt/faap/library/_text")

~/mnt does not exist on this machine. The library is at ../library/_text. Every run of the tool produced exactly one line —

*** SHELF NOT REACHABLE at <the configured path> — this run proves nothing ***

— and exited. No quotation was checked, and --negatives re-tested nothing.

Why this is worse than an ordinary broken script

CLAUDE.md records the guarantee this tool exists to provide: negative findings “are re-tested on every run by check_quotations.py --negatives”, because “a false positive is caught by the next reader; a false negative is caught by nothing.” That re-test is the only mechanism standing behind a deleted sentence. It has not run since the tool was written.

The banner was honest and that is the whole problem. The tool did not claim a clean pass — it said in terms that the run proved nothing, which is the correct behaviour and the reason this entry records no wrong publication. An honest failure message is still a silent failure if the number it replaces is the number anyone looks at. CLAUDE.md warns that “a silent skip and a clean pass look identical in output”. This is the neighbouring case: a loud skip and a clean pass look different, and are read the same, because both end the run without a complaint to act on.

The fix, and what it now shows

The path resolves against three candidates in order — the maintainer’s device mount, ../library beside the repository, and a local working copy — taking the first that exists, and falling back to the mount so the existing banner still fires when none does. The post-mortem is in the file, per the standing rule that these scripts carry their defects beside them.

First real run: 245 shelf files, 1,108 published quotations of 60 characters or more, 100 found on the shelf. The three recorded negative findings all re-test as confirmed present, which is what E62 established by hand and nothing had re-established since.

One shelf file has no extractable text — BILL_CO-SB25B-004_signed-act_2025-08.txt — so a miss against that source means nothing at all, and the tool now says so on every run.

The rule

E69 — a tool that reports its own failure has not thereby reported it to anyone. Before relying on a checker’s guarantee, run it once and confirm it can reach what it checks. A guarantee that depends on a path is a guarantee about a machine, and this repository is worked on from more than one.

And the caution it repeats. CLAUDE.local.md records that tool-building is where this assistant over-produces, and that several of 26 August’s errata were caused by checkers written that same afternoon. This is the fourth. The instrument was correct in design, complete in its reasoning about false negatives, and pointed at a directory that was not there.

E70 — Cedar Point’s “sine qua non” is a law professor’s phrase in a see also parenthetical, and the case says nothing about intangible property

Status: published in docs/known_objections.md and standards/table_of_authorities.md. Corrected on reading the opinion, 26 August 2026.

What was published, in the takings objection stated at its strongest:

“the right to exclude is the ‘sine qua non’ of the property interest (Cedar Point Nursery v. Hassid, 594 U.S. 139, 150 (2021)).”

Read in the opinion, the phrase is there and it is not the Court’s. It appears once, at the end of a string citation, in the weakest signal the Bluebook has:

“see also Merrill, Property and the Right to Exclude, 77 Neb. L. Rev. 730 (1998) (calling the right to exclude the ‘sine qua non’ of property).”

A parenthetical characterizing an academic article is not a holding, and “calling” is the Court’s own word for what Merrill does with it. The table row carried it as the proposition the case is cited for.

The Court’s own words for the same idea, in the same paragraph, are weaker and are themselves quotations: the right to exclude is “universally held to be a fundamental element of the property right” and “one of the most essential sticks in the bundle of rights that are commonly characterized as property” — both quoted from Kaiser Aetna v. United States, 444 U.S. 164, 176, 179–180 (1979). The pincite this project needs is to Kaiser Aetna, and the sentence it wants is forty-two years older than the case it credited.

The per se limb, verbatim, and it is narrower than the objection assumed

“The essential question is not, as the Ninth Circuit seemed to think, whether the government action at issue comes garbed as a regulation (or statute, or ordinance, or miscellaneous decree). It is whether the government has physically taken property for itself or someone else—by whatever means—or has instead restricted a property owner’s ability to use his own property. . . . Whenever a regulation results in a physical appropriation of property, a per se taking has occurred, and Penn Central has no place.”

The dividing line is physical appropriation against restriction on use, and it is stated as the essential question. Cedar Point concerns a right to “physically enter and occupy the growers’ land for three hours per day, 120 days per year.”

What the case does not contain, which is the finding

The words “trade secret” and “intangible” do not appear in the opinion. Neither does Ruckelshaus. Cedar Point is about physical entry onto land, and its companion authority Horne is about raisins — tangible personal property physically set aside for the government.

docs/known_objections.md builds the objection by pairing Ruckelshaus (trade secrets are property) with Cedar Point (the right to exclude is the property itself), and then reasons at point Two that “the per se limb may not care about publication at all”, so “a compelled handover to the State is a handover whether or not the State prints it.”

That step needs a bridge Cedar Point does not build. Extending a per se physical-appropriation rule from occupying land to compelling the production of records is the whole of the argument, and the case cited for it never reaches intangible property, never cites the case that does, and frames its own rule around whether property was physically taken. The objection may still be good — but it is good on Ruckelshaus and on an extension nobody has briefed here, not on Cedar Point.

The prediction in the reading brief was right and this is the second instance. The takings section was built reading Ruckelshaus and Cedar Point through the xAI plaintiff’s brief. Ruckelshaus was read and came out narrower and more favourable to this Act than the brief implied. So does this.

The pincite cannot be settled, and the second copy is a label

Two copies are held and neither carries U.S. Reports pagination. The supremecourt.gov slip opinion carries slip pages, as its filename says. The second copy — filed as 594-US-139_2021_Justia — carries no internal reporter pagination at all, only the print pagination of the capture (“Page 22 of 26”). The reporter citation lives in the filename and the URL.

Under E58 a citation that exists only in a URL or a filename is a label, not a citation. This filename promises reporter pagination and delivers none, and it was the more promising of the two copies on its name alone. 150 stays the plaintiff’s brief’s pincite (E47).

The rule

E70 — a citation signal is part of the citation. See also introducing a parenthetical that reports what an academic “call[s]” something is the furthest a court can stand from adopting a phrase while still printing it. Before attributing a quoted phrase to a court, find what introduces it, and if the answer is a signal rather than the court’s own sentence, the phrase belongs to whoever the citation names.

And the corollary this project keeps meeting. E67 found a district court’s sentence published as a court of appeals’. This finds a law review’s phrase published as the Supreme Court’s. Both arrived the same way: from a source quoting the case rather than from the case, and in both the real author was named on the page all along.

E71 — which Walton opinion is 294 A.3d 65 cannot be settled from the shelf, and the only source that answered it is one E59 already discredited

Status: the question E59 opened is examined and stays open. Recorded because “still open” is a finding when the reason is known, and because one source that looked like an answer is not one.

The question, narrowly. C.A. No. 2021-0827-JTL produced two opinions two weeks apart (E59). This repository cites Walton as 294 A.3d 65, 90, 92. Which opinion carries that citation?

Both are on the shelf and read. Neither answers it.

	Laches opinion	Demand-futility opinion
Caption	“OPINION ADDRESSING DEFENDANTS’ MOTION TO DISMISS ON THE BASIS OF LACHES”	“MEMORANDUM OPINION”
Submitted	13 January 2023	13 January 2023
Decided	12 April 2023	26 April 2023
Last slip page	62	121

Neither PDF carries an Atlantic Reporter stamp anywhere, which is ordinary — a slip opinion is published before the reporter citation exists. The shelf cannot settle this, and no further reading of these two documents will change that.

The source that appeared to answer it, and why it does not

research/verification_record.md carries a ◐ model-mediated row asserting: “Laster V.C., 26 Apr. 2023, C.A. 2021-0827-JTL, 58 pp., reported 294 A.3d 65.”

That row is wrong about the page count in a way that is checkable, and it is now checked. The 26 April opinion runs to slip page 121, not 58. E59 already recorded that “two separate model-mediated fetches” reported 56 or 58 pages for a document that has 64. This is the same family of fetch, wrong by the same margin, and its reporter citation carries exactly the credibility of its page count. Under E57 a model-mediated fetch is a lead, not a reading, and this one has now failed the only part of itself that could be tested.

What reading the two opinions did establish

The 26 April opinion cites the 12 April opinion, and does it by Westlaw number:

“Ontario Provincial Council of Carpenters’ Pension Tr. Fund v. Walton (Walmart Laches), 2023 WL 2904946, at *18 (Del. Ch. Apr. 12, 2023).”

Two things follow. As of 26 April 2023 the laches opinion had no Atlantic Reporter citation to give — which is consistent with either eventually taking 294 A.3d 65 and settles nothing between them.

And the court supplies the short form this repository needs. Vice Chancellor Laster calls the earlier one Walmart Laches, and relies on it at *18 and *21 of the later one. E59’s rule is that a case number is not a document; the court had already solved the naming problem, and adopting Walmart Laches for the April 12 opinion and Walmart Demand Futility for the April 26 opinion makes the collision hard to repeat.

The temptation that was declined, and it is E47’s

The published pincites are 90 and 92; the confirmed quotations sit at slip op. 76 and 77–78 of the 26 April opinion. A ratio can be computed from those numbers that makes 90 and 92 look right, and a 62-page laches opinion beginning at 65 would end well before 90, which makes the answer look obvious.

That reasoning is not permitted here and would not be worth much if it were. E47’s rule is that a page produced by a process is not a page read from a document, and its second-order rule is that the most persuasive form an invention takes is one that agrees with the arithmetic. The pincites 90 and 92 come from the same secondary source whose account of this case has now been wrong three times; using them to identify the opinion and then reporting them as confirmed against it would be circular.

What would settle it

One look at 294 A.3d 65 in the Atlantic Reporter, or any reporter-paginated copy of either opinion. Until then the row keeps its ⚠ on 90 and 92, the citation stays as published with its source named, and neither opinion may be described as “the” reported one.

The rule

E71 — when a source is shown wrong on a fact that can be checked, its other facts do not survive on their own. The ◐ row was believed for its reporter citation while being disbelieved for its page count, and those arrived together from one fetch. A retrieval is credited or discredited whole, unless some part of it has been independently confirmed.

E72 — the “no-fault” claim at n.18 is good, the authority for it is now read, and the “split” attached to it is the law firm’s word and not the court’s

Status: a debt paid rather than an error published. One characterization in standards/table_of_authorities.md is corrected. Read 26 August 2026.

What was owed. model_act_v3_4_companion.md n.18 asserts that the clawback “keeps Sarbanes-Oxley § 304’s no-fault severity”. The table row recorded that this was “[t]he appellate authority this project’s ‘no-fault clawback’ characterization rests on and does not cite”, known only from a law-firm alert, and that the claim “remains uncited until someone reads it.”

It is read, and the claim holds. SEC v. Jensen, 835 F.3d 1100 (9th Cir. 2016), holding verbatim:

“In accordance with its text and legislative history, we hold that SOX 304 allows the SEC to seek disgorgement from CEOs and CFOs even if the triggering restatement did not result from misconduct on the part of those officers.”

The reasoning is textual and the court states it plainly: the clause “as a result of misconduct” modifies “the material noncompliance of the issuer”, so “it is the issuer’s misconduct that matters, and not the personal misconduct of the CEO or CFO.” The court adds that Congress “knew how to draft a statute that would limit the disgorgement remedy to cases of officer or director misconduct, and chose not to do so”, contrasting the enacted text with a rejected House version.

First in the courts of appeals, in the court’s own words: “While we are aware of no circuit court that has addressed this issue, most district courts to have examined it have concluded that SOX 304 does not require CEOs or CFOs to have personally engaged in misconduct.”

The one thing the row had wrong

The row reports that “district courts had split for fourteen years.”

The opinion describes no split. It says “most district courts to have examined it have concluded” against a personal-misconduct requirement, and the string it gives runs one way — Jenkins (D. Ariz. 2010), Baker (W.D. Tex. 2012), Geswein (N.D. Ohio 2011), Life Partners Holdings (W.D. Tex. 2014). No contrary district decision is cited anywhere in the discussion. “Split” came from the law-firm alert, and this project reproduced it as though it were the court’s account of the landscape.

That matters more than it looks. A holding that resolves a split and a holding that ratifies a settled district consensus carry different weight, and the second is what happened here.

What the copy can and cannot do

The pincite cannot be settled. The copy is the Ninth Circuit’s own PDF, and its page markers are the slip pages of that issue (26, 27, 28, 29 through the SOX 304 discussion). No F.3d page appears anywhere in the document, so 835 F.3d 1100 and any pincite into it stay the secondary source’s under E47.

And the disambiguation the row recorded in advance held up. This is not the Jensen at n.22 — the DOJ food-safety prosecution in D. Colo. Two cases, one surname, opposite subjects.

The rule

E72 — a secondary source’s account of the landscape is a claim, and a separate one from its account of the holding. The alert was right that Jensen is first in the circuits and right about what it held. It was wrong that the district courts were split, and that word travelled into this repository attached to two facts that were true. Check the background characterization against the opinion’s own description of the authorities, which is usually a paragraph away from the holding.

E73 — Jewell and Cincotta confirmed verbatim, the footnote number settled where the page cannot be, and Cincotta’s next sentence widens what this project cites it for narrowing

Status: two debts paid, one characterization corrected. Read 26 August 2026.

United States v. Jewell, 532 F.2d 697 (9th Cir. 1976) (en banc)

The glossary’s quotation is verbatim, and it is the court’s own recapitulation:

“In the language of the instruction in this case, the government must prove, ‘beyond a reasonable doubt, that if the defendant was not actually aware . . . his ignorance in that regard was solely and entirely a result of . . . a conscious purpose to avoid learning the truth.’”

The words are the trial court’s instruction, set out at length earlier in the opinion and adopted by the en banc court as the standard. The glossary calls this “the classic formulation is United States v. Jewell”, which is right about whose standard it became and imprecise about whose sentence it is. The bracketed “[of the crime]” the glossary inserts stands in for the instruction’s actual words — “that there was marijuana in the vehicle he was driving when he entered the United States” — and the brackets disclose the substitution, which is what E48 requires.

704 stays unconfirmed. The copy is a law.resource.org capture with no star pagination, as its filename says (E47). A lead, recorded as one: Cincotta cites this discussion as “United States v. Jewell, 532 F.2d 697, 699–704”, which puts 704 inside the range a first-order source assigns to it. That is corroboration from another opinion, not a page read from this one, and it does not retire the ⚠.

And the two authorities in this glossary entry share a source, which the entry does not say. Jewell quotes Glanville Williams — “A court can properly find wilful blindness only where it can almost be said that the defendant actually knew” — and that is the same sentence Global-Tech quotes at 770 to fix willful blindness above recklessness. The Supreme Court’s ceiling and the Ninth Circuit’s origin rest on one line of a 1961 textbook.

United States v. Cincotta, 689 F.2d 238 (1st Cir. 1982)

The quotation is verbatim, and the footnote number is confirmed even though the page is not. In the capture held the passage sits in the numbered note beginning “2”, between the note quoting the indictment and note 3. “n.2” is settled; “243” is not, the copy carrying no star pagination.

What the glossary omits is the sentence immediately after, and it runs the other way. The glossary offers Cincotta as a narrowing of Jewell — “Narrowed in United States v. Cincotta” — on the strength of “The conscious avoidance principle means only that specific knowledge may be inferred when a person knows other facts that would induce most people to acquire the specific knowledge in question.”

The court continues:

“Thus, if someone refuses to investigate an issue that cries out for investigation, we may presume that he already ‘knows’ the answer an investigation would reveal, whether or not he is ‘certain’.”

A presumption of knowledge from a refusal to investigate is wider than the sentence quoted, not narrower, and it is the operative half for a prosecutor. The “only” limits what conscious avoidance is — circumstantial evidence of knowledge rather than a substitute for it, as the note goes on to say — and it does not limit how far the inference reaches.

Which compounds a correction already made. E65 records that this glossary entry reads as a descent from broad to narrow — Jewell, then Cincotta “narrowed”, then Global-Tech as the ceiling — and that the sequence is backwards, because Cincotta’s “would induce most people” is an objective test while Global-Tech requires a subjective belief in a high probability plus deliberate avoidance. This entry adds that Cincotta is not a narrowing of Jewell either. The word “Narrowed” was doing work no authority in the entry supports.

The rule

E73 — where a page cannot be confirmed, check whether some other coordinate in the citation can be. A footnote number, a part heading, a paragraph number and a docket entry are all locators, and a source without star pagination may still fix three of them. “Unconfirmable” is a property of the page, not of the whole pincite, and this repository has been retiring the entire locator whenever the page failed.

E74 — Liu confirmed as n.18 states it, and read beside Jensen it puts a question to the sentence next to it

Status: a debt paid; one assumption in n.18 identified as an assumption. Read 26 August 2026.

n.18 states that “Liu v. SEC, 591 U.S. 71 (2020), confined equitable disgorgement to net profits applied for victims — a statutory clawback is not so confined, but the section adopts Liu’s destination logic by choice (restitution first, fund second).”

The first clause is exact. The holding, verbatim:

“The Court holds today that a disgorgement award that does not exceed a wrongdoer’s net profits and is awarded for victims is equitable relief permissible under §78u(d)(5).”

And the rationale: “to avoid transforming an equitable remedy into a punitive sanction, courts restricted the remedy to an individual wrongdoer’s net profits to be awarded for victims.”

The assumption, which reading Jensen the same day exposed

n.18’s second clause — “a statutory clawback is not so confined” — is offered without authority and is doing real work: it is what lets SEC. 7 keep Liu’s destination logic as a choice rather than a constraint.

But Liu’s limit attaches to relief that is equitable, and the Ninth Circuit calls SOX 304’s reimbursement provision exactly that. In the same opinion this project read today to confirm n.18’s “no-fault” claim:

“This is consistent with our conclusion elsewhere that the reimbursement provision is an equitable and not a legal remedy.” SEC v. Jensen, citing SEC v. Jasper, 678 F.3d 1116, 1130 (9th Cir. 2012).

So the one statutory clawback in American law that this section is modelled on has been characterized by a court of appeals as equitable, and Liu confines equitable disgorgement to net profits for victims. Whether Liu’s ceiling reaches a statutory clawback that a court has called equitable is a live question, and n.18 answers it in a subordinate clause.

This is not a finding that n.18 is wrong. Liu construes §78u(d)(5) specifically, SOX 304 is a different statute with its own text, and Jensen’s characterization was made for a different purpose. It is a finding that the clause is an argument rather than a given, and that the two authorities n.18 cites in the same breath pull against each other in a way nobody here had noticed because they had not been read together.

The practical consequence is small and worth stating, because it cuts the way the section already goes: SEC. 7 adopting restitution-first “by choice” reaches the same destination Liu would compel if the ceiling does apply. The design is safe; the reasoning offered for it is not the reasoning that makes it safe.

Pagination

⚠ The copy is the supremecourt.gov slip opinion and it says so on every page — “Cite as: 591 U. S. __ (2020)”, the blank being the tell. No U.S. Reports pincite can be taken from it (E47). The filename is accurate.

The rule

E74 — two authorities cited in one sentence have to be read against each other, not only against the sentence. n.18 cites Liu and relies on SOX 304, and the proposition joining them — that a statutory clawback escapes Liu’s ceiling — survived because the two were verified separately and never set side by side. Where a note cites more than one case, ask what each says about the other.

E75 — the last six of the sixteen, read: five confirm what was published, and National Pork Producers did not remove a rule that never existed

Status: five debts paid with no correction owed; one characterization corrected. Read 26 August 2026.

Veeck v. Southern Building Code Congress Int’l, 293 F.3d 791 (5th Cir. 2002) (en banc)

docs/questions.md states it accurately — “model codes enacted into law enter the public domain as law” — and the court’s own summary is the source of that phrasing: “as law, the model codes enter the public domain and are not subject to the copyright holder’s exclusive prerogatives.”

The clause after it is the one this project should carry, because it is about this project. The same sentence continues: “As model codes, however, the organization’s works retain their protected status.” Veeck is a two-sided holding. An unenacted model act is not in the public domain under it, and this Act has been enacted nowhere.

So Veeck is not authority that this Act “is CC0”, as the table row’s cited-for column has it. The CC0 dedication is a choice, and Veeck tells you what happens after a legislature acts, not before. The choice is prudent on Veeck’s reasoning — the text goes into the public domain the moment anyone enacts it, so reserving rights buys a right that expires on success. That is a good argument and it is not the argument the row makes.

⚠ No star pagination, as the filename says; nothing may be pincited to it.

Kentucky v. Dennison, 65 U.S. (24 How.) 66 (Dec. Term 1860)

The quotation in audit/record.md is verbatim in the Library of Congress print, allowing for a scan whose OCR renders “offences” as “o[lYnces” in the same line:

“The word ‘crime’ of itself includes every offence, from the highest to the lowest in the grade of offences, and includes what are called ‘misdemeanors,’ as well as treason and felony.”

And the overruling note is precisely right. The syllabus separates the two holdings the way this repository does: point 4 gives the scope of “treason, felony, or other crime”, while points 8 and 9 are the mandamus holding — “Congress cannot coerce a State officer, as such, to perform any duty” and “upon that ground only, this motion for a mandamus was overruled.” It is points 8 and 9 that Puerto Rico v. Branstad, 483 U.S. 219 (1987) overruled; the scope holding stands, which is what the record says and what the Act relies on.

A note on this copy, recorded because the next reader will hit it. The scan is genuinely paginated, but the page numbers are OCR-damaged and irregular in the extracted text — markers jump 97 to 105 with the intermediate numbers lost to the running heads. The passage’s page was not computed and is not published, per E47. Nothing turns on it: audit/record.md cites the case without a pincite.

A search that failed, recorded because it failed the way the manual says it will. The first --find for this quotation returned NO MATCH. The string was in the document. The search omitted the quotation marks around ‘crime’ that the reporter prints, and the tool matched nothing. The banner held — a miss is not a finding — and reading the file settled it in one command.

National Pork Producers Council v. Ross, 598 U.S. 356 (2023)

n.17 says the case “removed the almost-per-se rule against state laws with extraterritorial practical effects”.

The Court’s position is that there was no such rule to remove. Its words:

“A close look at those cases reveals nothing like the ‘almost per se’ rule against laws that have the ‘practical effect’ of ‘controlling’ extraterritorial commerce that petitioners posit . . . . Baldwin, Brown-Forman, and Healy did not mean to do so much.”

The rule is attributed throughout to the petitioners — “Petitioners insist that Baldwin, Brown-Forman, and Healy taken together suggest an ‘almost per se’ rule” — and the Court declines to find it in them, explaining that the highlighted language “appeared in a particular context and did particular work.”

“Removed” and “declined to recognize” are different holdings, and the second is stronger for this Act. A rule abolished in 2023 invites the argument that it may be restored or narrowed. A rule the Court says never existed leaves nothing to restore. n.17 understated its own authority.

The rest of n.17 checks out: the line is re-read as being about discrimination — laws that “hoard” commerce “for the benefit of in-state merchants” — and what remains is Pike.

⚠ Slip opinion: “Cite as: 598 U. S. ____”. No U.S. Reports pincite from this copy.

Sveen v. Melin, 584 U.S. 811 (2018)

The two-step Contract Clause test is verbatim, and the threshold is as the record has it: whether the state law has “operated as a substantial impairment of a contractual relationship”, and “[i]f such factors show a substantial impairment, the inquiry turns to whether the state law is” drawn appropriately.

⚠ Slip opinion: “Cite as: 584 U. S. ____”. ✅ But 811 does not need this copy — audit/record.md records it confirmed against the preliminary print, “Volume 584 U.S. Part 2, Pages 811–836”. The table row’s ⚠ said no U.S. Reports pincite was available and another file in this repository already had the first page, which is E49’s rule turned on the shelf: the register is part of the text, and so is the drafting record.

Trump v. Slaughter, No. 25-332 (2026)

The holding, verbatim: “If anything more is left of Humphrey’s, we overrule it.” The Lawfare commentary’s account was right, including “the demise of Humphrey’s Executor”. What survives of the 1935 case is only “its observation that an agency that ‘exercises no part of the executive power’ need not fall within the rule of Presidential removal.”

And the answer to the question the table row asked is that SEC. 3’s Agency is untouched. The decision rests entirely on Article II and the President’s removal power over federal officers. The phrase “state agency” does not appear in the opinion, and no part of its reasoning reaches a State’s power to structure its own agencies, which is a matter for that State’s own constitution.

What is exposed is a rhetorical position, not a legal one. SEC. 3 is designed on the independent-commission model, and that model has just lost its federal exemplar. A legislator who asks “why build an independent commission when the Supreme Court has just dismantled the idea federally” is asking a fair question with a good answer — the answer being that the objection is about Article II and a State is not subject to it. That answer is not written anywhere in this repository, and the row was right that nobody had asked.

Desai & Riedl, Responsible AI Agents, arXiv:2502.18359

Verbatim, from the abstract: “no matter how much AI Agents seem like human agents, they need not, and should not, be given legal personhood status. In short, humans are responsible for AI Agents’ actions.”

And the liability-shield reasoning the row summarizes is in the text: “Anthropomorphizing software confuses issues and could lead to a world where software has legal personhood, related rights, and liability shields. If that happens, the power for people to use software would grow while also increasing the ability to avoid responsibility. That is the situation to avoid.”

The row’s caution stands and is confirmed by the same page: “Put simply, responsible AI Agents are about responsible human action” — a premise this Act shares, reached by authors who prefer design standards to personal criminal duties. Ally on the premise, not the mechanism, exactly as the row says.

The rule

E75 — check whether the authority says a rule was abolished or says it never existed. The two read alike in a summary and differ in what they leave standing. Where a court attributes a rule to a party — “petitioners insist” — and then declines to find it, a project that reports the rule as “removed” has credited the losing side’s premise while citing the winning side’s case.