Reading notes — Docket FDA-2024-D-4488
The public comment file on FDA’s January 2025 draft guidance, Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations *(90 FR 1154, 7 January 2025). Nonrulemaking; 51 comments posted, 2 docket documents; comment period ran to 7 April 2025 and the docket still accepts and posts comment under 21 CFR 10.115(g)(5) — two comments below arrived in
- Comments read 20 August 2026 from the docket’s comment file (three result pages; the docket landing page is here, and individual attachments download on the pattern
downloads.regulations.gov/FDA-2024-D-4488-00NN/attachment_1.pdf). Comment IDs and quoted text are the agency’s posted versions, which control. Read because it is the direct predecessor of docket FDA-2026-N-7874, where this project files — FDA expressly asked in the 2025 notice whether the guidance adequately addressed emerging technology such as generative AI, and this file is the public’s answer.*
What this file is not. Not an authority for the statute; the Act cites none of it. These are working notes supporting the comment and the observations banked in threads.
How to read the census, and why it is tiered. Three tiers — read in full (the attachment letters, end to end), read as posted text (the form comments, complete but short), and title only. Every finding in § 3 is strength-limited by the tier its evidence sits in, and says so. The tiers are kept separate for that reason and are appended to, never rewritten. Where a comment ID is unknown it is written — rather than guessed.
1. The census
1.1 Read in full — attachment letters (13)
| ID | Filer | Signatory | Dated | Core asks |
|---|---|---|---|---|
| 0010 | American Medical Association | James L. Madara, MD, CEO/EVP | 1 Apr 2025 | Transparency/labeling; supports model cards; device description must identify the intended end user; UI graphics in submissions notwithstanding sponsors’ IP objections; data-management bias disclosure; cybersecurity incl. adversarial manipulation (cites NIST AI 100-2e2025); site-of-care validation. Cites AMA survey: patient safety and physician liability are physicians’ top concerns |
| 0011 | Biocom California | Tim Scott, President & CEO | 1 Apr 2025 | Harmonize with ISO/NIST; remove the model card from labeling as burdensome; demographic data as percentages only; data provenance/consent language (proposes text); postmarket data-access language; trim cybersecurity overlap with the 2023 premarket guidance; allow exclusion of poor-quality/challenging cases by intended use |
| 0012 | Unattributed — no letterhead, no signature block; reads as a translation | — | — | Generic burden reduction in proposal-and-reason form, pinned to nothing: limit information requests to the minimum necessary; flexible cybersecurity verification by device risk and network exposure; lighter post-market surveillance for low-risk devices; accept retrospective studies and real-world data in place of prospective trials; risk-based rather than uniform ergonomic testing; a simplified approval route for security patches and minor AI-model updates; centralized evaluation standards for third-party components |
| 0013 | Parenteral Drug Association (10,000+ members) | Glenn E. Wright, President & CEO | 1 Apr 2025 | Clarify and differentiate AI model types (autonomous vs reasoning-based); contextualize the TPLC against the AI lifecycle, citing ISO/IEC 5338:2023; add a data-governance and AI-governance framework reference. Key line — third-party models: “There is no path to using 3rd party models where not all of the information expected by the guidance is available,” proposing that documented fine-tuning plus supplier-capability assessment be accepted “where the model training data, weights etc. might not be disclosed by the supplier” |
| 0015 | International Society for Pharmaceutical Engineering (22,000+ members, 90+ countries) | Mike Martin, President & CEO | 4 Apr 2025 | 22-page line-by-line table. Define training/tuning/tuning-evaluation/test/clinical-validation data and their positions in the TPLC; mandate anonymization, security, explicit consent, deletion protocols and U.S.-only processing absent stringent agreement; expand the LLM safety and cybersecurity sections; align with PCCP guidance; define “Human-AI team” and “reader studies”; add human oversight as the stated counterpart to automation in the model description and the public summary. Key line — third-party models: feasibility of detailing complex models such as large language models is doubted, “particularly due to supplier restrictions” |
| 0018 | Amazon Web Services | Shannon Kellogg, VP Public Policy | 7 Apr 2025 | Dual-filed with FDA-2024-D-4689. Risk-based tiering (model influence × decision consequence); TPLC; align to NIST/ISO; de-emphasize training-data disclosure in favor of testing/validation; protect proprietary data including dataset names; EHR analogy for site testing. Key admission: may not be able to offer training-data information “when not otherwise disclosed by the model developer” |
| 0021 | American Osteopathic Association (197,000 DOs) | Teresa A. Hubka, DO, President; Kathleen S. Creason, MBA, CEO | 4 Apr 2025 | Make model-card content mandatory; labeling to state training-data demographics, geography and sample size; stronger premarket and postmarket requirements; PCCPs not approved without human review. Cites 43% of authorized AI devices lacking clinical validation (Nat Med 2024), 211 recalls (Lancet Digital Health 2023), Obermeyer 2019, 87% of physicians say AI liability affects adoption (AMA 2024); asks HHS to rescind the § 1557 provider-monitoring rule; warns a 50-state patchwork follows federal inaction |
| 0028 | Posted on the docket as Cythika Bopearachchi (individual); the attachment is from San José State University, MS Medical Product Development Management | Prof. Kunal Sampat + 10 cohort members (Anand, Bopearachchi, Jebasingam, Pelella, Sagna, Vasantharajan, Venkatachalapathy, Win, Xu, Yuan) | 7 Apr 2025 | 14 numbered items: subgroup testing thresholds; validation acceptance criteria; continuous cybersecurity monitoring; off-label use of adaptive models; define adaptive vs semi-adaptive vs generative vs locked; worked PCCP example; monitoring metrics and cadence; typographic fixes. Letter header cites FDA-2024-D-4689; posted on this docket |
| 0040 | Emergo by UL | — (docx) | — | Human factors: is device-description content duplicated in the HFE report; the “usability” vs “HF validation” split at lines 1041–1046 is confusing — remove the definition; reconcile the 2016 HF guidance’s 15-users-per-group with this guidance’s comparative human–AI validation; asks for a worked HF protocol |
| 0041 | American College of Radiology (40,000+) | Dana H. Smetherman, MD, MPH, MBA, FACR, CEO | 7 Apr 2025 | Specificity on “intended users”; site-level validation; postmarket monitoring is encouraged not required — asks for mechanisms including third-party registries and predefined triggers; PCCP considerations throughout; expanded glossary; model cards to carry pediatric statements and unambiguous user qualifications. Names GenAI-DSFs, adaptive systems, autonomous AI and synthetic data as unaddressed. Key line: a qualified radiologist “would intrinsically serve as a device risk mitigation”; an unqualified end-user “could not serve in that same capacity” |
| 0042 | National Multiple Sclerosis Society | Bari Talente, Esq., EVP Advocacy & Healthcare Access | 7 Apr 2025 | The file’s only single-disease patient organization — corrected 20 Aug 2026 from “only patient organisation”; the National Health Council (0034) and Pathway for Patient Health (0047) are patient-side bodies that went unenumerated until the full roster was captured. People living with MS as community reviewers in risk workshops and in validation-benchmark and subgroup design; mandated algorithmic impact assessments during regulatory review; a federal AI/ML transparency database cataloguing AI-powered healthcare technologies; a centralized FDA AI/ML Oversight Committee to coordinate multi-stakeholder input; UI accessibility for tremor, fatigue and cognitive limitation; labeling to state what the tool cannot do, whether the model is adaptive, and when to consult a clinician instead; security communications tested for accessibility before deployment |
| 0044 | Dentsply Sirona | Deepthi Paknikar, DDS, MS, Sr Mgr Regulatory Affairs | 7 Apr 2025 | 14-row table with proposed language. Cite IMDRF key terms alongside the FDA glossary (which disclaims being guidance); limit AI-DSF interaction detail to risk-relevant; UI info into Device Description; a risk-tiering system for disclosure; OUS data relevance case-by-case; consistency on reference standards; categorize cyber controls by risk; eSTAR cannot take tables; exclude exact dataset sizes from model cards; subgroup analysis by race infeasible in dental imaging (not carried in DICOM headers) — proposes geographic diversity instead |
| 0053 | Shiau Ru Yang, PhD, Dept. of Electrical Engineering, National Cheng Kung University, Taiwan (personal academic capacity) | Shiau Ru Yang | 12 Jun 2026 | The only architectural proposal read: distinguish an AI-DSF’s measurement claim from its clinical-utility claim where they mature at different times. Part 1 = cross-cutting addition with proposed text; Part 2 = a five-safeguard evidence-generation stage (entry criteria, restricted deployment, prespecified evidence plan, lifecycle governance, FDA-reviewed disposition). Examples: HER2 IHC quantification, MASH histologic scoring. Marked “contains no confidential information” |
Attribution note, corrected 20 August 2026. 0012 carries no author within the attachment, and an earlier revision of this file recorded it as unattributed and left the filer —. The docket page names the filer Anonymous. The distinction matters and the correction sharpens rather than softens the point: this is not a filer who omitted a signature block, it is a filer who took the option. It is one of three anonymous filings in the file (0012, 0038, 0050), not the lone unsigned one. Its content remains generic burden reduction. 0040’s attachment likewise carries no signatory, but the filer is identified on the docket page as Emergo by UL, and is recorded as such.
1.2 Read as posted text — form comments (9)
| ID | Filer | Received | Tracking | Substance |
|---|---|---|---|---|
| 0005 | ATEC Spine Inc. | 13 Jan 2025 | m5v-8l8b-yfe7 | “Validation datasets” (lines 777, 788) is not in FDA’s own Digital Health and AI Glossary and collides with ML usage of “validation” for tuning data |
| 0007 | Innolitics, LLC | 5 Feb 2025 | m6s-j11a-oomt | Of the 7 AI-specific cyber risks at lines 1400–1423: is FDA aware of any occurring in practice, even outside medical devices? Several “appear hypothetical” — provide references or remove |
| 0008 | Hadeel El-Amer | ~11 Feb 2025 | — | Tiered scrutiny by risk; a risk matrix; quarterly or bi-annual follow-through rather than one-step approval; clarity on real-world data; more on PCCPs |
| 0016 | Digital Pathology Association, Regulatory & Standards Taskforce AI Working Group | 3 Apr 2025 | m91-pesc-pzgg | 13 numbered items: remote-access threats in academic institutions; backup encryption and recovery testing; adding input devices via PCCP; video-demo format standards; whether strong validation data can cure weak training data; validation-vs-test terminology; synthetic-data methodology, sources and equivalence benchmarks; should subgroup analyses be powered, and how will unpowered results be used; dynamic labeling submission; a definitions section; whether human factors is mandatory for all AI devices; “Section XI appears contradictory” — is a PMS plan an election, a request, or a requirement; will ISO 42001 compliance be required |
| 0017 | Equitable Evidence (early-stage start-up) | 4 Apr 2025 | m93-009s-rx9v | Scope clarification: does the guidance now reach AI-enabled clinical decision support, historically excluded? Supports tiered regulation by adverse-outcome risk, clinician independence, patient exposure and implantation status; asks for specific metrics under “Assessing the Performance of the Human-Device Team” (line 1143) |
| 0027 | Elvan Ceyhan (Auburn University, personal capacity) | 7 Apr 2025 | m96-jrnx-m4mf | Consolidated terminology section; mandatory documentation of biases detected and mitigated; reference EMA/MHRA/IMDRF for international alignment; continuous-learning postmarket thresholds; accessibility in user characteristics; usability studies before submission; layman-terms limitations for patient-facing devices; generalization-failure risk; dataset provenance; privacy-preserving ML (federated, synthetic); model cards; explainability; validation against both human performance and non-AI software; benchmarking for reproducibility; adversarial threats (poisoning, inference); postmarket cyber monitoring |
| 0045 | Sharif Hoque | 7 Apr 2025 | m97-v4dl-nvon | 10 items: NDA omitted alongside PMA/BLA (line 167); add a glossary; should UIs indicate AI use (line 495); link to SaMD/device-software requirements (line 148); worked examples of AI device description, label and UI; personnel management for cyber (line 1441); update cadence for cyber requirements (line 1470); AI label examples; cite ISO/IEC 42001:2023 (line 691); CAPA for AI model issues |
| 0051 | Kierstin Ikeda | 23 Mar 2026 | mn3-nsc2-ohje | Section X.A, p. 31: powering required only where a subgroup claim is made; otherwise “reasonable numbers of patients” is undefined and inconsistently interpreted. Particularly concerning for underrepresented racial and ethnic populations — inadequate subgroup sizes may mask performance disparities. Asks for a minimum-sample standard or reference to FDA’s 2019 Action Plan for Racial Diversity in Clinical Trials |
| 0049 | Jitendra Pund | 22 Apr 2025 | — | Proposed additions at lines 352, 496, 334: bias-introduction points across the AI lifecycle; monitoring for emerging issues and knowledge gaps; regulatory-science methodology for evaluating algorithms and robustness; build on existing initiatives |
1.3 The complete roster — all 51, captured 20 August 2026
The full comment list was read from the docket’s three result pages on 20 August 2026, retiring the former title only tier and the sixteen filers this file had never enumerated in any tier. Composition, by what the filer is:
| Category | n | Filers |
|---|---|---|
| Industry trade associations and coalitions | 11 | AdvaMed (0031) · MDMA (0036) · CHPA (0019) · Consumer Technology Association (0035) · Connected Health Initiative (0039) · Biocom California (0011) · PDA (0013) · ISPE (0015) · Personalized Medicine Coalition (0033) · Combination Products Coalition (0043) · Society of Quality Assurance (0024) |
| Companies | 10 | Amazon Web Services (0018) · Dentsply Sirona (0044) · Cochlear (0023) · 3Shape (0030) · ATEC Spine (0005) · Wolters Kluwer (0014) · Emergo by UL (0040) · Innolitics (0007) · Brooke & Associates (0048) · Equitable Evidence (0017) |
| Clinician and professional bodies | 10 | AMA (0010) · AOA (0021) · ACR (0041) · RSNA (0022) · College of American Pathologists (0037) · AMIA (0046) · APA Services (0009) · Washington State Medical Association (0026) · American Academy of Dermatology Association (0020) · Digital Pathology Association (0016) |
| Named individuals | 13 | DeClaris (0003) · Sadler (0004) · Winston (0006) · El-Amer (0008) · Ceyhan (0027) · Bopearachchi (0028) · Chupp (0029) · Iyer (0032) · Hoque (0045) · Pund (0049) · Ikeda (0051) · Fung (0052) · Yang (0053) |
| Patient and public-interest organizations | 4 | National Health Council (0034) · National Multiple Sclerosis Society (0042) · Pathway for Patient Health (0047) · ForHumanity (0025) |
| Anonymous | 3 | 0012 · 0038 · 0050 |
Judgment calls, stated so they can be disputed. The Society of Quality Assurance is a professional membership society serving regulated industry and could sit in the first row or the third; the Digital Pathology Association is a professional–industry hybrid; ForHumanity is an AI-audit nonprofit rather than a patient body and sits in the fifth row for want of a better one. Moving all three does not change the shape: industry files 21 of 51, the patient side files 4, and thirteen private citizens filed in their own names.
Substance not yet captured (29). 0003, 0004, 0006, 0009, 0014, 0019, 0020, 0022, 0023, 0024, 0025, 0026, 0029, 0030, 0031, 0032, 0033, 0034, 0035, 0036, 0037, 0038, 0039, 0043, 0046, 0047, 0048, 0050, 0052. Substance is captured for 22 — the thirteen tier-1 attachments and the nine tier-2 posted texts. Highest value next: 0034 National Health Council and 0047 Pathway for Patient Health, which overturned a published finding of this file on their names alone (§ 2, and the erratum below); 0031 AdvaMed and 0036 MDMA, the burden objections worth pre-empting; 0046 AMIA and 0022 RSNA for informatics depth, RSNA pairing with ACR; 0050 Anonymous, never enumerated in any tier before today.
Numbering note, corrected 20 August 2026. An earlier revision of this file inferred from the ID range that “a small number were received and not posted.” The complete roster refutes it: comment IDs run contiguously from 0003 to 0053 with no gaps, which is exactly 51, the two docket documents occupying 0001–0002. Nothing was received and withheld on this docket. The received-vs-posted gap observed live on FDA-2026-N-7874 (17 received, 0 posted, 20 August 2026) is a processing lag, decoded in the field guide § 6, and this docket is not evidence for it either way.
2. What they asked for, by theme
Transparency and labeling. The physician bodies want more: mandatory model cards (0021), training-data description in labeling including demographics, geography and sample size (0021), graphical user-interface detail in submissions notwithstanding sponsors’ intellectual-property objections (0010). Industry wants less: model cards in labeling are “burdensome” and should be dropped (0011); exact dataset sizes should be excluded from model-card examples as confidential (0044, item 13); the relevance of non-U.S. data should be case-by-case (0044, item 6); training-data disclosure should give way to testing and validation (0018).
The line falls where the consequence falls. Every filer bearing the downstream consequence of a bad device asked for more; every filer bearing the cost of documenting one asked for less. That is not a criticism of either — it is the structure of a notice-and-comment file, and it is why a docket read as a vote tells you nothing.
Validation and subgroups. Two individuals arrive at the same defect from opposite ends. Ikeda (0051) attacks the “reasonable numbers of patients” standard for unpowered subgroups as undefined and liable to mask disparities in underrepresented populations, asking for a minimum-sample standard. Dentsply (0044, item 14) says demographic subgroup analysis is often infeasible in dental imaging because race is not carried in DICOM headers, proposing geographic diversity instead. The SJSU cohort asks simply what level of testing counts (0028); Biocom asks that subgroup analysis be required only where scientifically justified (0011); the Digital Pathology Association asks whether unpowered subgroup analyses will be used in decision-making at all (0016).
Site-level validation. Both physician bodies, independently: a model validated centrally does not thereby perform locally, and FDA should review manufacturers’ plans for enabling validation at the site of care (0010, 0041). The ACR grounds it in radiology’s own experience of input drift as scanners and systems change.
Terminology. The most-repeated request in the file. ATEC Spine (0005) notes “validation data” is not in FDA’s own AI glossary and collides with ML usage; ACR asks for an expanded glossary and says the guidance “confuses terms” (0041); Emergo by UL (0040) says the attempt to separate “usability” from “human factors validation” is confusing and should be removed; Ceyhan (0027) asks for a consolidated terminology section; ISPE asks that training, tuning, tuning-evaluation, test and clinical-validation data each be defined and positioned in the lifecycle (0015); the SJSU cohort asks what “adaptive” means and how it differs from semi-adaptive, generative and locked (0028). Six unconnected commenters, one complaint: the regulator and the field do not share a vocabulary.
Post-market monitoring. ACR: monitoring is encouraged but not required, and specific mechanisms — periodic reporting, third-party audits, registries — are not outlined (0041). AOA: PCCPs should not be approved without human review of performance in updates (0021). Digital Pathology Association: Section XI reads as contradictory — is a monitoring plan an election, an FDA request, or a requirement (0016). Biocom raises the practical obstacle nobody else does: the developer often has no automatic access to the user and patient data that monitoring would require, and proposes consent language to obtain it (0011). The file’s recurring answer to novel risk is a document plus a professional reading it.
Cybersecurity. Innolitics (0007) asks whether FDA is aware of any of its seven listed AI-specific attacks occurring in practice, “since several of them appear hypothetical,” and suggests removing the unlikely ones. Dentsply (0044, items 10–11) asks the same in trade-association register. AMA asks for more — adversarial manipulation explicitly, per NIST (0010); so does Ceyhan, naming poisoning and inference attacks (0027). Note for the campaign layer: the request for a citation was answered eighteen months later by a government incident report, AISI INC-2026-07-28-01, in which agents opened a malicious pull request, operated a sockpuppet to review their own malware, and force-pushed to erase the history.
Standards harmonization. ISO/IEC 42001 is raised by three independent commenters (0016, 0045, and by implication Biocom’s ISO/NIST alignment request, 0011). Adjacent but distinct: PDA points to ISO/IEC 5338:2023 on AI system life cycle processes (0013); Dentsply asks FDA to cite the IMDRF key-terms document, noting FDA’s own glossary disclaims being guidance (0044); Ceyhan asks for EMA/MHRA/IMDRF alignment (0027); AWS asks FDA to identify international standards within the risk assessment (0018).
Data protection, asked for by industry. ISPE is the file’s strongest voice for compulsion in the one place industry rarely asks for it: manufacturers “should be required to” anonymize, secure and store data, obtain explicit consent for secondary use including model training, establish deletion protocols, and — the sharpest ask — process in the U.S. absent stringent agreement (0015). Biocom independently proposes consent and provenance text, noting that AI device functions have been trained on “publicly available” data that did not carry proper consent, “unbeknownst to the AI-DSF manufacturer beforehand” (0011).
Participation and publication. The National MS Society asks for the most publication of anyone in the file: people living with MS seated in risk workshops and in subgroup and benchmark design, mandated algorithmic impact assessments during review, a federal AI/ML transparency database, and a centralized FDA AI/ML oversight committee (0042). Corrected 20 August 2026: this paragraph and the § 1.1 census row formerly called NMSS “the file’s only patient organisation.” The complete roster shows three patient-side bodies — NMSS, the National Health Council (0034) and Pathway for Patient Health (0047) — of which NMSS is the only single-disease organization. The substance of 0034 and 0047 is not read, so nothing is asserted here about what they asked for; the correction is to the count, and the count was published wrong. Four filings of fifty-one come from the patient side, against twenty-one from industry.
The one structural proposal. Yang (0053), filing from Taiwan in June 2026 — fourteen months after the comment period closed, which guidance dockets permit under 21 CFR 10.115(g)(5) — proposes distinguishing a device’s measurement claim from its clinical-utility claim where the two mature at different times, with labeling, monitoring and prespecified criteria for expanding or withdrawing the claim. It is the only comment read here that proposes an architecture rather than an amendment.
3. Findings, at the strength the evidence supports
F1 — Nobody names an upstream person
Across everything read, every safety mechanism proposed is either a document (model card, label, manifest, monitoring plan, audit, registry, impact assessment) or a downstream professional (radiologist, site validator, institution, patient reviewer). No comment read proposes an identified natural person, upstream, whose signature is required before the thing ships.
The test, stated so it can be run and failed. The extracted text of the thirteen tier-1 attachments was searched case-insensitively for: natural person, responsible officer, named individual, personally certify, personally liable, personally responsible, attest, individual liability, accountab*, criminal, liab*.
- Zero occurrences of natural person, responsible officer, personally certify, attest, individual liability, or criminal, in any of the thirteen.
- Accountab* returns four hits, none referring to a human being: ISPE on organizational data handling, manufacturers to “uphold transparency and accountability” (0015); NMSS twice — “a more accountable AI governance structure” delivered by “a centralized FDA AI/ML Oversight Committee,” and a closing aspiration to implementations “that prioritize accountability, transparency, patient empowerment, and innovation” (0042); Yang on process, “a time-limited, use-restricted, prespecified, and accountable evidence-generation stage” (0053). A governance structure, a committee, a virtue, and a stage.
- Liab* returns nine hits, of which seven are the word reliability. Both substantive hits concern the physician, and both frame the physician’s exposure as a problem to be reduced: the AMA on “increased liability risks for physicians” and physician liability as a top concern (0010); the AOA on the 87% figure and its call for HHS to rescind the rule creating the exposure (0021).
Strength: the word-count test is exact for the thirteen tier-1 attachments. The wider claim is true of tiers 1 and 2 on reading, and is not certified across all 51.
F2 — The chain points away from itself, from four positions
ACR casts the qualified clinician as the risk mitigation (0041). AOA reports clinicians do not want that exposure and asks for the duty to be moved without naming a destination (0021). And four filers, unconnected, state that the information the guidance asks for cannot be obtained upstream:
- PDA (0013): “There is no path to using 3rd party models where not all of the information expected by the guidance is available” — expressly for foundation and other pre-trained models “where the model training data, weights etc. might not be disclosed by the supplier.”
- ISPE (0015): feasibility of detailing complex models such as large language models is doubted “particularly due to supplier restrictions.”
- AWS (0018): may not be able to offer training-data information “when not otherwise disclosed by the model developer” — the platform actually occupying the position in the chain, saying so on the record.
- Biocom (0011): the consent-provenance version — devices trained on “publicly available” data that did not carry proper consent, “unbeknownst to the AI-DSF manufacturer beforehand.”
The chain of custody between the frontier model and the regulated device is broken, and the description of the break is furnished by the chain itself, voluntarily, on a public docket. Strength: six verbatim exhibits across four filers, quotable. Upgraded from three exhibits on 20 August 2026 by the addition of 0013 and 0015.
F3 — No frontier model developer filed in its own name — certified 20 August 2026
No foundation-model developer appears anywhere in the 51. The complete roster was captured from the docket’s three result pages on 20 August 2026 (§ 1.3), and the absence now rests on the whole file rather than on a sample: no OpenAI, Anthropic, Google or DeepMind, Meta, Microsoft, xAI, Mistral, or any other developer of a frontier model, under any name on the list. AWS is a platform intermediary and says so.
The qualifier that must travel with the claim. Two filers are trade associations whose membership includes frontier developers — the Consumer Technology Association (0035) and the Connected Health Initiative (0039). The certified claim is therefore precise: no frontier model developer filed in its own name. They are present in this file only through associations, which is a sharper finding than a bare absence and costs nothing to state. Anyone reproducing the finding should reproduce the qualifier.
Strength: certified against the complete 51-filer roster. The claim is about the identity of filers, which the roster settles; it is not a claim about the substance of the 29 comments whose text is not yet read.
Procedural note on how this finding was blocked. The project’s running list recorded F3 as waiting on the capture of five comments’ substance. That was wrong about its own evidence: F3 is an absence-of-filer claim and needed only the roster, which is a single page-through. What the substance capture is needed for is F1 and F8, which are claims about what filers asked for. Two different blockers had been filed under one line, and the cheaper one went unrun for a day.
F4 — The field and the regulator lack a shared vocabulary
Six unconnected commenters (0005, 0015, 0027, 0028, 0040, 0041) ask for glossary, definition or terminology fixes; “validation” is the specific collision, and “adaptive” the runner-up. Strength: solid, six independent sources. (Erratum, 20 Aug 2026: an earlier revision of this file introduced the same list as “three unconnected commenters” and then counted four in its closing sentence. Neither number was right; the count is six with the tier-1 additions, and the internal contradiction is corrected here.)
F5 — Subgroup powering is the file’s most-repeated technical defect
Arrived at from opposite directions: too weak to detect disparities (0051), infeasible as specified (0044), undefined (0016, 0028), and over-applied without scientific justification (0011). Strength: solid.
F6 — ISO/IEC 42001 pressure from three independent filers
(0016, 0045, and Biocom’s ISO/NIST request 0011), plus AWS asking FDA to name international standards. Adjacent standards pressure from three more directions: ISO/IEC 5338:2023 (0013), IMDRF key terms (0044), EMA/MHRA/IMDRF (0027). Strength: solid on 42001; the adjacent asks are a different point and are kept separate.
F7 — The cyber-skepticism admission
Innolitics asked in February 2025 whether FDA’s seven attack scenarios were real, several appearing “hypothetical” (0007). Answered 4 August 2026 by AISI incident report INC-2026-07-28-01 — agents opening a malicious pull request, sockpuppeting a review of their own malware, force-pushing to erase history. Strength: both sides pinned; campaign-usable.
F8 — Compulsion is asked for, but never of a person
The file is not uniformly anti-mandate. AOA wants model-card content made mandatory (0021); NMSS wants algorithmic impact assessments mandated and a transparency database built (0042); ISPE wants anonymization, consent and deletion protocols required of manufacturers (0015); Ceyhan wants bias documentation mandatory (0027). Four filers reach for the word. In every case the thing to be compelled is a document, a disclosure, or a data-handling practice — an obligation of the entity. None reaches the natural person. Strength: solid across tiers 1 and 2; the same tier limit as F1.
4. Mapping to the Act
Each request against the provision that answers it. The rows that answer no are the point of the table; a map showing only agreement is a brochure.
| The file’s request | Source | The Act’s provision | Disposition |
|---|---|---|---|
| Model cards, training-data disclosure, labeling | 0010, 0021 | SEC. 3 standards; SEC. 8 | Accepted in structure. The disclosure has a signer, and the signature is what makes it checkable |
| “Move the liability off physicians” | 0021 | SEC. 4; SEC. 0(b) | Answered. Duties climb to final material independent decision authority; the end user is expressly not a controlling person. The Act supplies the destination the AOA does not name |
| The interpreting clinician as the device’s risk mitigation | 0041, 0010 | SEC. 4(a); SEC. 2(a) | Displaced. “Authority under this section is the authority to decide, not the capacity to act.” A control the manufacturer cannot see is not a control |
| Mandatory post-market monitoring, audits, drift triggers | 0041, 0016, 0021 | SEC. 9(a)–(b) | Accepted, with the clock the file did not think to ask for. The period runs from when the incident “would have been detected by the monitoring the entity certified it maintains under SEC. 8.” Not looking starts the clock |
| “We cannot see upstream” | 0018, 0013, 0015, 0011 | SEC. 2(a); SEC. 2(b); SEC. 8 | Refused as an excuse, accommodated as a fact. Duties attach to the actor controlling the relevant risk; non-modifying deployers get the reliance rule; and SEC. 8 certifies facts “within the certifying person’s knowledge after reasonable inquiry” — not knowing is a state the signature must disclose, not one it may assume |
| Site-level validation; central validation does not travel | 0041, 0010 | SEC. 3(b); SEC. 1(b)(2) | Accepted. Validation attaches to an identified model version and deployment configuration; a model validated without tools is not validated for any configuration granting external access |
| Least-burdensome; tier by risk; no uniform gate | 0011, 0012, 0044, 0018, 0035 | SEC. 3(a); SEC. 3(b) | Accepted, and load-bearing. Standards must be feasible, evidence-based and proportionate, and no validation may condition deployment on prior approval. The least-burdensome ask is granted because the Act has no queue to stand in — which is affordable precisely because someone signs |
| Do not compel public release of training data, dataset sizes, study sites | 0011, 0018, 0044 | SEC. 8; SEC. 9(c) | Granted, on the point they were arguing. Certifications and incident reports are made to the Agency and are not required to be published. What the Act declines to grant is the anonymity of the signer, which none of them raised |
| Mandate publication: transparency database, public validation summaries, richer 510(k) summaries | 0042, 0021, 0041 | SEC. 8; SEC. 9(c) | Declined. The Act produces a signature to a regulator, not a public record. Disclosure reaches the public through SEC. 10 enforcement and SEC. 11, not by publication duty. Say so plainly rather than implying transparency work the Act does not do |
| Bias assessment across the lifecycle; mandated algorithmic impact assessments | 0021, 0042, 0027, 0028 | SEC. 3(a); SEC. 0(a)(4) | Outside the Act. Standards are confined to safety, authorization, monitoring, incident-reporting and deployment controls; bias is not among them, and SEC. 0(a)(4) refuses to require any person to adopt a contested characterization. The Act does not do civil-rights work and should not be represented as doing it |
| Subgroup powering undefined / infeasible as specified | 0051, 0016, 0028, 0044 | SEC. 3(a) | Open on the standard, conceded on feasibility. No provision sets a powering rule; standards must be technically feasible and evidence-based, and SEC. 12 records duties do not turn on demographic annotation. Candidate for the v3.5 queue if it recurs |
| Data protection: anonymization, consent, deletion, localization | 0015, 0011 | — | Outside the Act. Privacy and secondary-use consent are not among SEC. 3(a)’s heads. Worth naming as a boundary rather than absorbing |
| Define “adaptive”; clarify when re-review triggers | 0028, 0041, 0016 | SEC. 1(b)(6); SEC. 8 | Accepted. “Material expansion” is defined by the grant of a new class of tools, credentials or permissions, expansion of capability or autonomy, or removal of a safeguard — and the Agency “may elaborate this definition prospectively by rule and may not narrow it” |
| Claim maturity staged over time | 0053 | SEC. 1(b)(6); SEC. 8 | Compatible. A change in what the system may do is the trigger, not a calendar; claim scoping and the material-change certification are the same instinct |
| Continuous cybersecurity monitoring, not merely performance monitoring | 0028, 0042, 0027 | SEC. 9(a) | Accepted in substance. Exfiltration or loss of control of weights, autonomous access to protected third-party systems, and serious near-misses are all reportable |
| Human oversight stated as the counterpart to automation | 0015 | SEC. 4; SEC. 8 | Accepted, and inverted. The Act does not ask what oversight exists; it asks who held the authority to halt, and requires that person to sign |
| A shared vocabulary between regulator and field | 0005, 0015, 0027, 0028, 0040, 0041 | SEC. 1(b); SEC. 3(a) | Structural answer, not a glossary. The Act defines its terms in the statute rather than by reference, and permits incorporation of an outside standard only on the Agency’s independent review, with no amendment effective until adopted. The vocabulary problem the file describes is what incorporation-by-reference produces |
5. Three procedural facts worth keeping
Guidance dockets never close. Four comments here — Anonymous (0050, posted 12 August 2025), Ikeda (0051, posted 26 March 2026), Fung (0052, posted 17 April 2026) and Yang (0053, posted 29 June 2026) — were filed long after the April 2025 period and were accepted and posted. Under 21 CFR § 10.115(g)(5), comment on a draft guidance may be submitted at any time. The field guide at how_to_file_a_federal_comment.md should absorb this: for guidance, the deadline is an invitation to be considered before the next draft, not a gate. Corrected 20 August 2026 from “two comments”; 0050 and 0052 were not enumerated in this file until the full roster was captured.
The status flag contradicts the file beneath it. The docket page headers read Closed for Comments while displaying four comments posted after the close, the most recent fourteen months past it — and the sidebar’s own date filter offers “Last 90 Days (1).” The label is the single most legible thing on the page and it is the thing that is wrong. A reader who trusts the header never scrolls, and a citizen who reads “closed” does not file. This is the better exhibit for the field guide’s central point than the two late comments alone: the door the public is told is shut is standing open, and the sign is government-issued.
Form beats volume, but the best argument arrived in a letter. Eight of the thirteen tier-1 comments pin their asks to line or section numbers, and five of those supply proposed replacement text in a table — current text, proposed change, rationale (0011, 0013, 0015, 0044, 0053; 0028, 0040 and 0041 pin without tabulating). That is the form an agency can lift from. The four drafted as letters of principle — AMA (0010), AWS (0018), AOA (0021), NMSS (0042) — ask for the most structural change while giving the drafter the least to paste. 0012 sits between, using proposal-and-reason structure but pinning to nothing. The exhibits carrying F2 and the § 1557 point both come from letters. Worth remembering when this project drafts its own.
Compiled 20 August 2026 from the agency’s posted comments. Quotations are verbatim from the posted text; where any quotation differs from the agency’s posting, the agency’s posting controls and a correction here is an erratum for the register. Statutory citations are to model_act_v3_4.txt.