AIxBio RegistryAI molecules in the clinic

Evidence-first re-review: protocol

Status: v1.8, 15 Nov 2026. v0.9 was written before any row was re-coded; v1.0 was frozen after the pilot (the pilot record (not published)); v1.1 changes the strict line by John E. Milad's decision after the full review (§ 10), v1.2 narrows physics_only (§ 11), v1.3 judges the method rather than the vocabulary (§ 12, which also sets the standing procedure), v1.4 settles delivery, lab-in-the-loop platforms and ML-guided optimisation (§ 13), v1.5 applies the method test to repurposing rows (§ 14), v1.6 retires the vaccine carve-out and judges model-selected antigens by the ordinary rules (§ 15), v1.7 writes down three readings for antigen products (§ 16), and v1.8 anchors § 13.1 to the active substance and makes delivery-only AI borderline (§ 17). All eight are recorded deviations, listed in §§ 9–17 with the reason. Any later change needs a new version and a re-code of the rows it affects.

This protocol applies METHODOLOGY § 1 under the stricter reading approved by John E. Milad on 20 Oct 2026 (the maintainer's decisions log (not published), "Evidence-first re-review…"). Where it is stricter than METHODOLOGY, it wins for the re-review; METHODOLOGY is updated to match once the re-review is signed off.

The acid test. If the developer itself does not claim, in its own words, that AI or machine learning did work for this molecule, the registry must not claim it either.

1. Unit of review

2. Evidence fields

Recorded for every strict and borderline row in molecules.csv. The best single claim is recorded: the one that comes closest to passing rule 1 and the step test.

Field Content
ai_claim_quote Verbatim, 40 words or fewer, copied from the page at ai_claim_url. Ellipses only to drop words, never to join sentences from different paragraphs.
ai_claim_url The page the quote was copied from. It must have been fetched during the review.
ai_claim_by developer (the company that discovered the molecule, or its successor or acquirer), partner (a co-developer, licensee or the AI-platform partner named in the deal), developer_paper (a paper or patent authored by the developer or partner), or third_party (press, analysts, trackers, reviews, conferences not by the developer). A verbatim quote of a named developer executive inside a news article counts as developer; a paraphrase does not.
ai_claim_date Publication date of the quoted source, ISO format.
ai_step target, design or both: the step the quote says the AI did. Blank when the quote names no step.
ai_step_stated Strict rows only: yes if the claim passes the materiality test (§ 3, step test), no if it passes only the acid test. A quality flag, not an inclusion test (v1.1).
borderline_reason Borderline rows only, exactly one code from § 4.

excluded.csv gains exclusion_code (§ 5). Every check is also logged in data/audit-log.csv (append-only; fields ai_claim_*, inclusion, borderline_reason, exclusion_code), with the source URL and a verbatim quote of 25 words or fewer.

3. Decision rules

A row is strict only if rules 1 and 3 pass and the row survives rule 4 (v1.1). The step test (formerly rule 2) no longer decides inclusion; it sets the ai_step_stated flag on every strict row.

Rule 1: attribution (acid test). The developer or a partner, in its own words, ties AI or machine learning to this molecule: - the source names the molecule, or is unambiguously about it (for example, a release announcing that molecule's first dosing); - the words are the developer's or partner's: its website, release, filing, presentation, or a paper or patent it authored (developer_paper counts here; it does not make the row attribution tier A, which needs independent documentation); - the AI is named as AI, machine learning, deep learning, generative or neural models, or a named tool the same developer describes as AI or ML (for example "Chemistry42"). "Computational", "in silico", "physics-based", "structure-based", "free-energy perturbation", "molecular dynamics" and "rational design" are not AI.

Fails rule 1: the claim is only about the company or platform and never connects to the molecule; or only third parties make it.

What counts as crediting AI (v1.1). In the statement about this molecule, the thing credited must be AI itself: AI, ML or generative models named in that statement; a platform that the statement itself calls AI ("our AI-powered platform", "our AI-enabled platform", "AI + robotics platform"); a named tool the developer describes as an AI or ML model (Chemistry42, PRISM, MALEXA); or the molecule called "AI-designed", "AI-discovered" or "AI-generated". Any verb of origin will do ("discovered using", "developed from", "derived from", "program from"). It is platform_only (borderline) when the statement credits a named platform or engine without calling it AI there, and the developer describes that platform elsewhere as combining AI with other methods (Monte Rosa's QuEEN, Black Diamond's MAP, the Benevolent Platform); or when AI appears only in company boilerplate next to the molecule.

Step test (rule 2 in v1.0; the ai_step_stated flag from v1.1). The claim says what the AI did for this molecule, in one of two steps: - target: AI or ML identified, nominated, predicted or prioritised the target, or the target–disease link, that this molecule acts on; - design: AI or ML generated, designed, optimised or engineered the molecule's structure or sequence, or selected the clinical compound or its chemical starting point from a set it scored (for example ML virtual screening that produced the series).

Wording rules: - Passes: "designed by/with/using [AI]", "AI-designed", "generated by [generative AI]", "optimised with [ML]", "[AI] identified the target", "AI-predicted target", "AI modeling revealed the target". - Fails (platform language, no step): "discovered using/with our AI platform", "developed using", "derived from", "emerged from", "came from", "powered by", "enabled by", "leveraging", "from our AI-enabled platform", "AI-discovered" with no step, "program from our AI platform". - The AI must be the instrument of the step in the quote. The quote itself calls it AI or ML ("our generative AI platform", "machine learning", "AI model"), or names a tool the developer describes as an AI or ML model (for example Chemistry42, PRISM). A multi-method platform that combines AI with another way of finding molecules (physics or docking, structure-based design, display or screening as the source of hits) passes only if the quote says the AI part did the step. Wet-lab testing inside an AI loop does not make a platform multi-method (§ 13). "Designed using our [multi-method] engine" or "rationally designed using [platform]" is platform_only. - ML used only for screening logistics, ADMET or property filtering among many gates, formulation, delivery vehicles, patient selection, biomarkers or trial operations fails. (From v1.8, AI credited for a delivery system or formulation alone is borderline delivery, § 17.2.)

In v1.0, failing this test made the row borderline platform_only. From v1.1, a strict row that fails it is flagged ai_step_stated = no.

Rule 3: dosed in humans by 20 Oct 2026, by any sponsor, including dosing by an originator before the current developer acquired the molecule. Any one of: - a trial registry record (ClinicalTrials.gov, China CTR, ChiCTR, CTIS/EU CTR, ANZCTR, jRCT, CRIS, PACTR) with actual enrolment above zero, or an actual (not estimated) start date together with a recruiting, active or completed status; - a dated developer, partner or regulator statement that a first participant was dosed, or that a trial started dosing or completed; - a developer-authored paper reporting human dosing results.

When dosing rests only on a developer paper or release, the coder must also try the trial registry record and say what it shows. Fails: IND or CTA clearance alone; "expected to start"; a registry record with an estimated start only; withdrawn with zero enrolment.

Rule 4: red team. A second reviewer argues for exclusion and searches for evidence that: - the molecule predates the AI step claimed (a legacy compound or an in-licensed asset, with the AI claim retrofitted later); - the developer's own detailed account (paper, patent, filing) describes the design or target choice without ML, contradicting the headline claim; - the claim is about a sibling, back-up or the programme rather than this molecule; - the AI step was minor (one filter among many), or AI was used only after the molecule was chosen; - the claim was withdrawn, or the developer now describes the molecule without AI; - dosing is not actually confirmed.

The row survives if the red team finds no evidence for these. An argument without evidence does not defeat a row. Evidence that defeats rule 1 or 2 moves the row to borderline with the matching code; evidence that defeats rule 3, or shows AI played no part, moves it to excluded.

4. Borderline: dosed, an AI connection, not strict

Code Use when
physics_only The developer's and partners' statements about this molecule name no ML: its design is described only as computational (physics-based, structure-based, FEP, molecular dynamics, docking). ML in the platform's general description does not lift it out. If any developer or partner statement about the molecule names ML, the row is tested under rule 1 instead (v1.2).
platform_only The developer or partner connects this molecule to a platform, but the statement does not credit AI itself (rule 1, "What counts as crediting AI"): the platform is named without being called AI there and is described elsewhere as multi-method, or AI appears only in company boilerplate. (In v1.0 this code also covered rows that credit AI but fail the step test; from v1.1 those are strict with ai_step_stated = no.)
third_party_only Only third parties make a molecule-level AI claim; the developer is silent.
repurposing AI or ML picked a new indication or combination for an existing drug. Borderline even if the claim passes rules 1 and 2.
delivery AI worked only on an excipient, a delivery system or a formulation that is not part of the active substance (a lipid nanoparticle, a proteo-lipid or polymer vehicle, a new formulation of an existing active), and the developer or a partner credits AI for that work on this product as rule 1 requires (v1.8, § 17.2).
vaccine_antigen Retired in v1.6 (§ 15). Until then: AI or ML selected vaccine antigens or neoantigens (including personalised vaccines), borderline even if the claim passed rules 1 and 2.
other Anything else; the note says what (for example a scope question in § 6 decided as borderline).

Where two codes fit, use the first that applies in this order: repurposing, delivery, physics_only, platform_only, third_party_only, other. third_party_only applies only when neither the developer nor any partner connects this molecule to AI at all.

The difference between platform_only (borderline) and no_ai_link (excluded): borderline needs a developer or partner statement that connects this molecule to the AI platform (the molecule named alongside the AI platform). AI language about the company only, never linked to the molecule, is excluded.

5. Excluded rows

Each exclusion reason is re-tested against facts current on 20 Oct 2026, molecule by molecule, and given one code:

Code Meaning
not_dosed No dosing evidence by the as-of date (rule 3 fails).
no_ai_link No developer, partner or third-party claim connects AI to this molecule; or AI language is company-level only.
ai_disclaimed The developer says AI played no part, or its own account shows the molecule came from elsewhere.
in_licensed Discovered elsewhere; AI used only after acquisition (development, patient selection).
legacy_retrofit A legacy molecule whose AI claim was added later, with no contemporaneous ML evidence.
out_of_scope Not a drug candidate the registry covers, or AI not used on the product: observational study, nutraceutical, AI used for analytics, patient selection, dosing, trial operations or manufacturing, or a scope question in § 6 decided as out. (Formulation and delivery-vehicle AI were here until v1.8; they are now borderline delivery, § 17.2.)

Where two codes fit, use the first that applies in this order: not_dosed, ai_disclaimed, in_licensed, legacy_retrofit, out_of_scope, no_ai_link.

If a re-test finds the reason no longer holds (for example a not_dosed molecule is now dosed), the molecule is coded under § 3 as if new.

6. Scope questions (proposals for John; defaults apply in the review until he decides)

# Question Options Default and why Downside if wrong
S1 Cell therapies. GB-5267 is a Generate/Roswell Park CAR-T, dosed in NCT07489287 (actual start 18 Aug 2026; first patient dosed, announced 1 Oct 2026), with a developer claim that AI designed its CAR. (a) out of scope, excluded; (b) borderline other; (c) in scope, coded under § 3 like any biologic Decided (c), John E. Milad, 27 Oct 2026 (§ 13.1). The test is whether AI designed the therapeutic protein that gives the drug its specificity or activity, however it is delivered. GB-5267 is coded like any biologic. Overcount by 1 if readers expect small molecules and soluble biologics only; modality lets them filter cell therapies out.
S2 Delivery-vehicle AI. METiS says AI designed MTS105's lipid nanoparticle, not the encoded therapeutic. (a) excluded out_of_scope; (b) borderline other (a) excluded; confirmed by § 13.1. Rule 2 asks whether AI chose the target or designed the therapeutic molecule; a carrier is neither, the same as formulation (MTS-004). Superseded 15 Nov 2026 (§ 17): delivery-vehicle AI is borderline delivery. None for the strict count; a reader interested in AI formulation loses the entry.
S3 ISM3312. Insilico's own Nat Commun 2025 paper says a healthy-volunteer Phase 1 was completed in China (CTR20230768); its IND release says the molecule was "generated and designed" with Chemistry42. (a) code under § 3 like any row; (b) keep excluded until a registry record is read (a). The developer-authored paper is a dated developer statement of dosing (rule 3), and "generated and designed" names the design step (rule 2). Likely strict if the quotes check out. Overcount by 1 if the Phase 1 report cannot be verified.
S4 Phage cocktails. BX004 (BiomX) is a cocktail of three natural lytic phages; BiomX's deck credits its ML-based BOLT platform, its own methods paper does not. (a) excluded out_of_scope; (b) in scope as a biologic, coded under § 3 Decided (b), John E. Milad, 27 Oct 2026. Choosing phages from a library is selection, which v1.3 counts when a trained model did it; a cocktail qualifies when a trained model chose, designed or engineered its phages. BX004 is borderline other: the developer's methods account selects by host-range coverage scoring, not a trained model (§ 12.5). None for the strict count; an engineered phage stays eligible.

7. Process

  1. Coders. Two independent coders per row, blind to each other and to the current tier, notes and ai_role. They see the molecule name, developer, partner, modality, target, indication and the current source URLs as leads only. Each records the § 2 fields, a pass/fail for rules 1 to 3 with the evidence, the tier and code, and a confidence (high, medium, low).
  2. Red team (rule 4) on every row either coder passes as strict.
  3. Adjudication. Rows where both coders and the red team agree take that tier. Every disagreement on tier, and every proposed move into or out of strict, goes to John with the evidence (the sign-off record (not published)).
  4. Verification. Before anything is written to the CSVs, the maintainer's session re-fetches every ai_claim_url and checks that the quote is on the page. A quote not found is not used, and the row is re-coded without it.
  5. Agreement is reported as percent agreement and Cohen's kappa on the tier (strict, borderline, excluded) and on the borderline code.

8. What this does not change

Target precedence, status, phase and attribution tier rules (METHODOLOGY § 3) are unchanged. Attribution tier A still needs documentation independent of the developer.

9. Changes from v0.9 (pilot) to v1.0

The pilot (24 rows, two blind coders) met its pre-set criteria: tier agreement 23 of 24 (κ 0.93), and 18 and 19 of 19 keyed rows reproduced. Four changes, each from a pilot disagreement:

Change Pilot case that prompted it
Rule 2: the AI must be the instrument of the step in the quote; "designed using" a multi-method platform is platform_only. MRT-2359: "discovered and rationally designed using our QuEEN" split the coders (strict vs borderline). QuEEN combines AI-guided chemistry with libraries, structural biology and proteomics.
Borderline order: platform_only before third_party_only; third_party_only only when no developer or partner ties the molecule to AI. AO-252: the licensee ties its only programme to A2A's "AI/ML-driven de novo design platform"; coders split between the two codes.
Exclusion order added. OPL-0401: ai_disclaimed vs in_licensed, both true.
Rule 3: dosing by any sponsor counts; a registry check is required when dosing rests on a paper. OPL-0301 was dosed by Sanofi (as SAR247799) before Valo in-licensed it, so "never dosed" was wrong; ISM3312's paper says both "completed" and "being evaluated".

v1.0 was used unchanged for the full re-review (158 rows, 20 Oct 2026).

10. Deviation: v1.0 → v1.1 (John E. Milad's decision, 20 Oct 2026)

Change. The strict line is the acid test (rule 1) plus dosing (rule 3) and the red team (rule 4). The step test no longer decides inclusion; it becomes the ai_step_stated flag. Physics-only, third-party-only, repurposing and vaccine-antigen rules are unchanged. Published counts are placed on a cumulative ladder (METHODOLOGY § 5).

Why. Under v1.0 the step test cut strict from 49 to 32. It mostly measured how detailed a developer's press release is, not whether the developer credits AI for the molecule. The lirafugratinib error that started this review was an acid-test failure (no one claimed it), which rule 1 still catches.

Status as a deviation. This is a definition change made after the pilot and the full review, by the maintainer, with the review results in view; it is not pre-registered. The pilot's inter-coder agreement (κ = 0.93) and the full review's (κ = 0.87) still cover the tests that remain: attribution, dosing, and the borderline codes. Rows that v1.0 coded platform_only were re-coded under v1.1 by two fresh blind coders, with a red team on every new strict pass (the sign-off record (not published) § E).

11. Deviation: v1.1 → v1.2 (John E. Milad's decision, 25 Oct 2026)

Change. physics_only applies only when the developer's and partners' statements about the molecule name no ML, even if the platform description does. A molecule whose developer names ML for it is tested under rule 1 like any other.

Why. SGR-1505 and SGR-3515 were coded physics_only because Schrödinger's design is physics-led, yet Schrödinger names machine learning for each molecule in its own words: - SGR-1505: "combining free energy calculations with machine learning methods" (J Med Chem 2025, doi:10.1021/acs.jmedchem.5c01494). - SGR-3515: "machine learning (Active Learning FEP+)" in the case study "en route to development candidate SGR-3515".

Both pass the acid test, and John decided both are strict. Relay's lirafugratinib and zovegalisib stay physics_only: their own discovery papers (PNAS 2024, doi:10.1073/pnas.2317756121; Cancer Discovery 2024, doi:10.1158/2159-8290.CD-23-0704) credit molecular dynamics and name no ML. SGR-2921 stays physics_only: Schrödinger names only physics and the platform.

Status as a deviation. This was decided by the maintainer after the review, with its results in view. Every other row coded physics_only was re-tested under v1.2 by two blind coders, with a red team on any pass (the sign-off record (not published) § E). The coders agreed on 15 of 17. Of the 7 both passed, BBO-8520 and SNX-281 survived the red team and became strict. Zasocitinib, iscartrelvir, omilancor, GT-02287 and ATH-063 stay borderline (other), because their developers' own detailed accounts describe the design without ML. Both SGR rows carry ai_step_stated = no, because the ML is credited only in combination with FEP (step test, instrument rule).

12. Deviation: v1.2 → v1.3. Judge the method, not the vocabulary (John E. Milad's decisions of 26 Oct 2026)

12.1 What counts as AI. A model whose parameters were trained on data: - machine learning and deep learning; - generative models; - active learning and Bayesian optimisation; - knowledge-graph or language-model inference; - ML-guided physics: active-learning FEP, ML scoring functions, ML force fields.

Not counted, but tagged: - rule-based expert systems; - automation and orchestration (robotics, workflow software); - pure physics: FEP, MD, QM or docking with no trained model; - classical QSAR used only as a property filter.

12.2 The strict line (replaces the acid-test wording of § 3 rule 1; rules 3 and 4 unchanged). A molecule is strict when developer or partner evidence shows that a data-trained model chose its target, or produced or selected the molecule itself or its chemical starting point. ML-guided optimisation counts as producing the molecule when the developer's methods account shows a trained model identified the modifications that define the clinical molecule (v1.4, § 13.3). The "or" is inclusive: either one is enough, and both together also qualify. Record ai_step as target, design or both (rentosertib is both), or antigen for a product whose antigens a model selected (§ 15). John's words: "if X or Y; or X & Y". - Eligible evidence: papers; patents (methods and examples sections); conference posters and abstracts; SEC and other regulatory filings; investor decks; releases; talks or interviews quoted verbatim. Any language counts; record the original wording plus an English gloss. - The word "AI" is not required. "A Gaussian-process model ranked the enumerated designs" counts. "AI-designed" with no method also counts, as before (step flag no). - Third parties alone still do not count. A news article counts only where it quotes a named developer or partner speaker verbatim.

Property filters (John's decision, Q2). A trained model used only as a property filter (ML ADMET or hERG prediction, for example) does not count on its own: tag it ml in ai_methods but do not count it. It counts only if the evidence says the model selected the clinical compound or its series.

Conflicting developer evidence (John's decision, Q3). A specific methods account (patent, paper) of a step outweighs a headline AI claim for that step. Silence is not contrary: a paper that does not mention ML does not contradict a developer's claim that ML did the step. Nor is a construct descriptor: "humanized", "fully human", "bispecific" or "Fc-engineered" says what the molecule is, not how it was found or designed (John's decision on HXN-1001, Decision 12, 27 Oct 2026).

12.3 Platforms. A molecule is strict with ai_step_stated = no when two things hold: - the developer describes the platform as built around a trained model (for example a generative chemistry engine, a foundation model, an ML design platform); and - the developer ties that platform to the molecule by name.

A platform needs evidence that its trained model did the work on this molecule only when it combines the model with another way of finding molecules: physics or docking, structure-based design, or display or screening as the source of hits (for example QuEEN, Dynamo or SEE-Tx). Wet-lab testing that generates the model's training data, or tests what it proposes (lab-in-the-loop), never triggers this. Without the needed evidence the molecule is platform_only. (Reworded in v1.4, § 13.2; the v1.3 text also listed "experiments", which caught every lab-in-the-loop platform.)

12.4 Boilerplate. AI wording in company boilerplate next to the molecule still does not count. The coder must look for molecule-level evidence instead (§ 12.2 sources), and record what was searched.

12.5 Demotion. Contrary evidence can move a strict molecule out: - the developer's or partner's own detailed account of the same step credits non-ML methods (the specific methods account beats a headline claim; silence is not contrary); - the molecule or its series predates the platform or model; - the molecule was in-licensed, with the model used only afterwards.

Codes: physics_only (method evidence shows only physics or docking), other (contradicted claim), legacy_retrofit or in_licensed.

12.6 ai_methods. A new field in molecules.csv, recorded for every row reviewed, whether or not the methods count. It holds a semicolon list from a fixed set: ml, deep_learning, generative, active_learning, knowledge_graph, llm, ml_physics, physics, docking, expert_system, automation, qsar_filter, unknown. ai_step_stated keeps its meaning: yes when the evidence says the model chose the target or produced or selected the molecule.

12.7 Scope of the re-run. - In both directions: all 55 strict rows (looking for contrary as well as supporting evidence); the wording-sensitive borderlines (18 platform_only, 8 physics_only, the contested other rows ATH-063, zasocitinib, GT-02287, iscartrelvir, VRG50635, GS-0321 and PHI-101, and INDV-2000); the four proposed no_ai_link exclusions still in molecules.csv (AC699, NDI-219216, NEU-723, RLY-8161); and the 17 excluded rows coded no_ai_link, molecule by molecule. - Later: repurposing and vaccine-antigen rows, as a separate scope question. Repurposing was settled in v1.5 (§ 14) and vaccine-antigen rows in v1.6 (§ 15).

12.8 Standing procedure (SOP). Every borderline or new candidate goes through this procedure: 1. a deep evidence hunt over the § 12.2 sources; 2. two blind coders; 3. a red team on every pass and every demotion; 4. every quote re-fetched by the maintainer session; 5. agreement (κ) reported; 6. John's sign-off before any tier change reaches main.

12.9 Proposals follow the rule (26 Oct 2026). Whenever new evidence changes an entry's outcome under these rules, the proposal is recomputed before it goes to John: the tier, the codes, every count and the sign-off pack. That includes a red-team finding, a re-fetched source, or an identity or dosing fact settled later. A decision put to John is a genuine judgement call: a reading of a rule the text does not settle, a scope question, or release timing. It is never a default the evidence no longer supports, left for him to overrule. Where the rule decides an entry, the pack proposes the rule's outcome and records the losing argument as the dissent.

Status as a deviation. Decided by the maintainer after the review, with its results in view. The pilot's agreement covers the dosing and code rules but not this evidence standard; agreement under v1.3 is measured afresh on the re-run.

Approval. John E. Milad approved § 12 on 26 Oct 2026 ("yes to all"), including Q2 and Q3 and the inclusive "or" of § 12.2. It applies from the v1.3 re-run onward.

v1.3 re-run (26 Oct 2026). 112 cases were coded in both directions by two fresh blind coders. Tier agreement was 91.1% (κ = 0.85); strict versus not, 94.6% (κ = 0.89). A red team then challenged 64 cases: 7 of the 59 passes were defeated, and all 5 demotions were upheld. Strict went from 55 to 53. Two follow-up red teams (26 Oct, § 12.9) each argued against a demotion. Iscartrelvir returned to borderline: its identity was settled, and the developer's own methods credit docking. DSP-2342's not_dosed demotion was defeated by new evidence. Strict is 52. Results are in the sign-off record (not published).

13. Deviation: v1.3 → v1.4. Delivery, lab-in-the-loop and ML-guided optimisation (John E. Milad's decisions of 27 Oct 2026)

13.1 What is designed, not how it is delivered (S1, Decision 8). The test is whether AI designed the therapeutic protein or molecule that gives the drug its specificity or activity. The vehicle that carries it (a cell, a viral vector, a lipid nanoparticle, a conjugate's linker or payload) need not be AI-designed, and an AI-designed vehicle does not qualify on its own. GB-5267 (AI-designed CAR in a T cell) is coded like any biologic; MTS105 (AI-designed nanoparticle, conventional payload) stays excluded out_of_scope. modality names the delivery so readers can filter. Superseded in part by § 17.1 (15 Nov 2026): conjugate linkers and payloads and viral vectors' capsids and genomes are part of the active substance, and an AI-designed delivery system alone is borderline delivery, not excluded.

13.2 Lab-in-the-loop is not multi-method (Decision 10). § 12.3 reworded: only another way of finding molecules (physics or docking, structure-based design, display or screening as the hit source) calls for evidence that the model did the work. Wet-lab data feeding a model, and experiments testing its proposals, are part of the AI loop. The same wording is applied to § 3.

13.3 ML-guided optimisation (Decision 10). Added to § 12.2: ML-guided optimisation counts as producing the molecule when the developer's methods account shows a trained model identified the modifications that define the clinical molecule. ABBV-295 is strict on this basis (Gubra's 2022 TIDES poster: a model trained on ~3,000 screened analogues "identified 5 key amino acid substitutions that accurately control amylin receptor selectivity").

13.4 Re-run. The rows whose platform_only coding rested on the multi-method clause and could move under § 13.2 (JUV-161, BDTX-189, silevertinib, BDTX-4933, ADG125) are re-coded under § 12.8. Result (27 Oct): the two blind coders agreed on 1 of 5 (BDTX-4933, platform_only); the red team settled the four splits. JUV-161 becomes strict (step not stated; its legacy challenge found no evidence). BDTX-189 and silevertinib stay platform_only (Black Diamond's own filings credit cell and tumor screens, and group mutation families in a non-ML step). ADG125 stays platform_only (display panning is the hit source; the proposed exclusion fell to Adagene's 2022 deck naming it as from its AI platform). Rows tied to physics, docking or structure-based platforms are unaffected; the changes only widen the strict line, so no strict row needs re-testing.

Approval. John E. Milad, 27 Oct 2026, in the maintainer session: "I approve the revised rule" (13.1) and "yes, approve the rewording" (13.2–13.3), with the instruction to change the rules and apply them as the review proceeds.

14. Deviation: v1.4 → v1.5. The method test for repurposing (John E. Milad's decisions of 8–9 Nov 2026)

The § 12.7 scope question for repurposing rows. Until v1.5 a repurposing row needed only an AI connection to the new use. From v1.5 it needs the same kind of evidence as a strict step. Sign-off pack and evidence: docs/reviews/2026-11-09-repurposing-signoff.md.

14.1 What repurposing needs. From v1.5, the § 4 code repurposing needs the same kind of evidence as a strict molecule's AI step, applied to the repurposing step. That is developer or partner evidence (§ 12.2 sources) that a data-trained model (§ 12.1) did one of two things: - picked the new indication or combination for an existing drug; or - selected the drug for that use from a set the model scored.

How the rest of the protocol applies: - § 12.2's evidence rules apply as written. - The word "AI" is not required, and "AI-identified" with no method still counts. - A specific methods account of the repurposing step outweighs a headline AI claim. Silence is not contrary. - A stated reason (pharmacology, earlier papers, mouse data) is silence. It counts against the claim only when the developer's account says the use came from a source that was not a model. - Third parties alone still do not count. - § 12.3 applies to platforms. - A platform built around a trained model, and tied to the drug by name, passes. - A platform that combines a model with another way of finding the use (docking, shape or similarity screening, a screen as the source of hits) needs evidence that the model did this step. - For a screen, that means the developer's account of this drug says the model's readout picked it. - § 12.4 applies: company-level AI wording does not count. - Rules 1 and 3 are unchanged. - Language models and agents (John E. Milad's decision, 8 Nov 2026). - A language model or agent counts when its output proposed the new use, or when its judgement scored or ranked the candidates and the drug was taken from that ranking. - People or lab tests checking the top picks afterwards is validation. - A model that only searched, extracted or summarised evidence, with a person or a fixed rule making the pick, is assisted review and does not count. The same is true of people ranking a model's output and choosing the drug. - Record llm in ai_methods either way. - There is no novelty test. Earlier testing in people still excludes. - The same line applies to § 12.2's target and molecule steps.

14.2 When the test fails. Test the exclusion codes in § 5's order, then the borderline codes in § 4's order.

Exclusion codes: - ai_disclaimed: the developer says AI played no part, or its own account shows the new use came from elsewhere. - in_licensed: the drug was acquired for a use already chosen elsewhere, with AI used only afterwards. Judge each use separately. - legacy_retrofit: either limb is enough. - The developer's accounts from the time describe finding the use without a trained model, and the AI wording came later. This includes the same team relabelling its own work. - Or the drug had already been tested in people for that use (the same disease; for a combination, the same pair), by any sponsor, before the AI step. A model that only rediscovers a tested use adds no new insight (John E. Milad, 9 Nov 2026). - The second limb gives way only when the developer's evidence shows a model did another step that counts in the register for this programme, such as choosing a new combination partner (a new pair is a new use). A model that produced a new molecule makes a new drug, judged under § 12.2. AI used for patient selection, biomarkers, dosing or trial operations counts nowhere in the register (§ 3 rule 1), so it does not lift the exclusion. - no_ai_link: no developer, partner or third-party statement ties AI to this drug, or the AI wording is company-level only.

Borderline codes: - platform_only: the platform combines a trained model with another way of finding the use, and nothing shows which method did this step. - third_party_only: only third parties tie AI to the repurposing step. - other: the developer's own account shows a method other than a model did the step, or the step was done by a rule-based system or by assisted review.

14.3 What an existing drug is. A molecule already dosed in humans, for any use, before the AI step. A trained model choosing a compound never dosed in humans from a library is selection under § 12.2, not repurposing.

14.4 Fields and scope. - Record ai_methods for every repurposing row. - ai_step stays blank and ai_step_stated is n/a. - When a drug has several new uses, test the entry's lead indication and note the others. - Split two molecules sharing an entry.

14.5 The maintainer's own platforms. A row whose AI step was done by a platform the maintainer built or runs is coded and signed off by reviewers other than the maintainer, and the row says so.

14.6 Re-run (9 Nov 2026). All 19 repurposing rows were reviewed under § 12.8. The 14 older rows went to two blind coders each (15 items once LAM-001 was coded on its own; tier agreement 14 of 15, κ = 0.86; strict-or-not 15 of 15); the 5 rows added since 3 Nov were checked against the existing evidence; three red teams challenged every split and outcome, and every relied-on quote was re-fetched. Result (v1.0.16): SOM3355, PXT3003 and REC-2282 excluded legacy_retrofit; azelaprag excluded ai_disclaimed; LAM-001 split from LAM-002A and excluded no_ai_link; OC514, baricitinib, INV-001 and INVA-8001 recoded other; OC212e and REC-3599 platform_only; BEN-2001 third_party_only; eight rows stay repurposing. Strict is unchanged.

14.7 Readings settled with the release. A same-team relabel of a non-model finding is legacy_retrofit, as tolcapone was (no same-team exception for a repurposing step). A statistical model fitted to data, such as a Cox regression, counts under § 12.1 when the developer presents it as machine learning; the notes record the worked method (BGE-117, BGE-175). Whether classical statistics should leave § 12.1 is a separate review across every tier.

Approval. John E. Milad: the language-model rule, 8 Nov 2026 ("Yes, adopt it"); the rediscovery rule, 9 Nov 2026 ("if a study has already been done in an indication, AI hasn't provided any novel insights"); § 14 and the pack's defaults, 9 Nov 2026 ("Your defaults are accepted.").

15. Deviation: v1.5 → v1.6. Vaccines whose antigens a model selects (John E. Milad's decisions of 13 Nov 2026)

The § 12.7 scope question for vaccine_antigen rows. Until v1.6 a vaccine whose antigens AI selected was borderline "even if the claim passes rules 1 and 2"; no reason was recorded for that. From v1.6 these products are judged like any other. Sign-off pack and evidence: docs/reviews/2026-11-13-vaccine-signoff.md.

15.1 One standard. The vaccine_antigen code is retired. A vaccine or immunotherapy whose antigens were selected by a data-trained model is judged by the ordinary rules: rule 1 and § 12.2 (the developer's or a partner's evidence that a trained model did it), rule 3 (dosed), rule 4 and § 12.5 (the red team and demotion). What counts is what AI designed (METHODOLOGY § 1), whether that part is common to every patient or made for each one.

15.2 The antigen step. A passing row is strict with ai_step = antigen: a model selected the antigens or epitopes the product carries. It is a value of its own beside target, design and both, so figures by AI step keep their meaning. ai_step_stated is yes when the evidence says the model selected the antigens, no when the developer ties an AI platform to the product without saying so. Target precedence is not classified for these rows: there is no single drug target.

15.3 Unit. One product candidate: a named product with patient-specific batches is one row. Two versions under one product code and one trial stay one row, with both described (SLATE and SLATE-KRAS).

15.4 How the ordinary rules read for these products. - A ranking that blends a model's prediction with expression or clonality is still the model's selection. - Laboratory confirmation of what the model proposed is lab-in-the-loop (§ 13.2). - Review of the selection by a board is not contrary evidence, and a paper that does not mention ML is silent, not contrary (§ 12.2). - A developer's detailed account that describes a rule-based or deterministic selector with no trained model is contrary evidence (§ 12.5): the row is other. - (v1.7) A named public tool that happens to be a trained network (for example NetMHCpan) does not supply attribution unless the developer or a partner describes the selection as machine learning or AI, or describes the tool as a trained model. - (v1.7) Silence and contradiction are different. A paper that gives no account of the selection method is silent (§ 12.2). A developer's detailed account of a selector with no trained model in it is contrary, and the product is other. A weighted ranking counts as the model's selection only where the developer says the weights or the predictor were fitted to data. - (v1.7) A partner's platform supplies attribution for an antigen product when the partner says the platform selected this product's antigens and describes the platform as AI. Where no statement ties the platform to the product's selection, § 12.3 applies and the code is platform_only.

These readings match rows already strict: abbv-295 (people chose after the model), sgr-1505, sgr-3515 and snx-281 (a trained model inside a combined method), juv-161 (assays confirm the model's picks) and gv20-0251 (a claim made after first dosing).

15.5 Repurposing is different. It stays borderline because the molecule was dosed in people before the AI step (§ 14.3). A model-selected antigen makes a new product. An antigen product already dosed for another disease is treated as § 14 treats a drug.

15.6 Completeness. These rows are not a census until a class-wide search is run and logged; the site says so. A first screen (docs/reviews/2026-11-13-vaccine-coding/CLASS-SEARCH.md) found three likely candidates and seven to look at more closely; each goes through § 12.8. (Reviewed in v1.7, § 16.)

15.7 Re-run (13 Nov 2026). All ten vaccine_antigen rows were reviewed under § 12.8: evidence gathered again from primary sources; two blind coders (tier agreement 10 of 10, κ = 1.00; tier and code 9 of 10, κ = 0.82); a red team on every pass and every non-pass; every relied-on quote re-fetched, and one coder quote struck. The coders worked to a stricter three-part test written for the coding and passed five rows. The red team showed that test rejects evidence patterns that six strict rows had passed with, so the proposal was recomputed under the ordinary rules (§ 12.9). Result (v1.1.0): GRANITE, SLATE, GRT-R910, EVX-01, EVX-02, TG4050, NECVAX-NEO1, intismeran autogene and autogene cevumeran strict with ai_step = antigen; GRT-R910 and autogene cevumeran at low confidence, with the dissent in each inclusion_note; Tamavaq borderline other (its own detailed account describes a deterministic selector, and no participant is confirmed vaccinated). Strict 55 → 64, borderline 88 → 79.

Approval. John E. Milad, 13 Nov 2026: pooled into strict, not a separate class ("We're just trying to count drugs that have significant AI-design to them"); one standard for vaccines, the two low-confidence promotions and SLATE as one row ("defaults approved").

16. Deviation: v1.6 → v1.7. Three readings for antigen products; eleven candidates reviewed (John E. Milad's decision of 13 Nov 2026)

The class-wide screen of § 15.6 named ten vaccine and antigen-product candidates, and the weekly watch had wrongly closed HMBD-501 as not dosed. All eleven went through § 12.8. Sign-off pack and record: docs/reviews/2026-11-13-candidates-signoff.md.

16.1 Three readings added to § 15.4. The red teams showed these were being applied but were not written down: a public trained tool supplies no attribution on its own; silence and contradiction are different, and a weighted ranking is a model's selection only where the developer says it was fitted to data; a partner's platform supplies attribution only when the partner ties it to this product's selection. No existing row changes. The nine antigen rows of v1.1.0 and Tamavaq were already coded this way: intismeran autogene passes on Moderna's statement that the algorithm is machine learning, although its patent's selector is a weighted rule; Tamavaq's own account describes a selector with no fitted model.

16.2 Antigen products that are not vaccines. § 15.1 covers "a vaccine or immunotherapy". A cell therapy raised against model-selected antigens is judged on the antigen set, with ai_step = antigen; the cell is the delivery (§ 13.1).

16.3 Result (v1.2.0). Two blind coders agreed on tier and code for all ten vaccine candidates (κ = 1.00) and split on HMBD-501. A red team attacked every pass and every non-pass: no pass was defeated (two unweakened, four weakened), three of four non-passes were upheld and one, pEVAC-PS, was weakened by a co-developer's release both coders had missed. The maintainer session re-fetched the quotes each call rests on: 12 of 13 matched their sources, and the one that could not be opened is not relied on. Before release an independent reader checked the public text of the eleven rows against the record and the sources; its corrections are in the internal changelog. - Strict, ai_step = antigen: EVM16 and NEO-PV-01 (high confidence), VB10.NEO (medium), and iNeo-Vac-P01, cNeT / ATL001 and NEOVACC (low, with the dissent in each inclusion_note). The three low-confidence rows have the defects autogene cevumeran and GRT-R910 were admitted with in v1.1.0. For NEOVACC the AI description is the partner's and is indirect: myNEO ties its ImmunoEngine platform to this product's antigens, describes the platform's presentation model as deep learning trained on mass-spectrometry data, and calls itself an AI-driven company; its page for the platform itself says only "computational". - Borderline other: pEVAC-PS. The University of Cambridge, a co-developer, says machine learning designed the antigen; the developers' methods paper describes a design with no trained model, and a specific methods account outweighs a headline claim (§ 12.2). Strict with ai_step = design is the recorded dissent. - Borderline platform_only: HMBD-501. Hummingbird's platform pairs its machine-learning tool with immunization as the source of antibodies (§ 12.3), no source says what a model did for this product, and it carries the antibody of HMBD-001, which is already platform_only. - Excluded no_ai_link: PGV-001, NOUS-PEV, NOUS-209. For PGV-001 the developers' paper says two constants of the ranking's binding score were set by a logistic fit to published immunogenicity data. They never describe the ranking as machine learning or as a trained model, so § 14.7 and the first reading above keep it out: a fitted formula counts only when the developer presents it as a model. The fitted-weights reading is a necessary condition, not a sufficient one.

Strict 64 → 70, borderline 79 → 81, excluded 57 → 60. Both coders and the red teams are Claude models: the agreement figures measure consistency, not accuracy.

16.4 Completeness. Still not a census (§ 15.6). Chinese-language registries and company sites, Japanese and Korean developers and academic trials were not searched. Not read: the methods of Ott et al. (Cell 2020, NEO-PV-01), Bhardwaj et al. (Cancer Discovery 2025, PGV-001) and Leoni et al. (Cancer Res 2020, NOUS-209).

Approval. John E. Milad, 13 Nov 2026: the pack's three questions approved as recommended ("Approve as recommended").

17. Deviation: v1.7 → v1.8. The active substance, delivery and manufacturing (John E. Milad's decision of 15 Nov 2026)

Decision note and evidence: docs/reviews/2026-11-15-delivery-scope-decision.md and its folder; sign-off pack: docs/reviews/2026-11-15-delivery-scope-signoff.md.

17.1 The active substance. Replaces the vehicle list in § 13.1. The test is whether AI designed part of the active substance, as regulators define it, where that part gives the drug its specificity or activity. That covers the binder, payload and linker of a conjugate; the capsid and genome of a viral vector; the full sequence of an mRNA or oligonucleotide, coding and non-coding regions alike; and the engineered receptor of a cell therapy, however the product is delivered. These are design under the ordinary tests: a trained model (§ 12.1; a deterministic algorithm such as a dynamic-programming codon search is not one), the developer's or a partner's attribution (rule 1) and, for optimisation, a methods account (§ 13.3).

17.2 Delivery and formulation: borderline delivery. Replaces S2 and the formulation and vehicle items in § 5's out_of_scope. AI that worked only on an excipient, a delivery system or a formulation that is not part of the active substance is borderline delivery, when the developer or a partner credits AI for that work on this product as rule 1 requires. A weaker connection takes the usual code. § 12.3 applies as it does to strict: a platform that combines a model with screening or another non-model source needs evidence that the model did the work on this product's delivery system or formulation; without it, the code is platform_only. Computational work on a formulation or solid form with no AI credited is excluded no_ai_link, or out_of_scope when it is manufacturing (physics_only is for the design of the active substance). Where AI designed part of the active substance and also its delivery, the active substance is coded.

17.3 Manufacturing. AI used only to make the product (cell lines, expression systems, process control), with the product unchanged, stays excluded out_of_scope.

17.4 Why. The vehicle list contradicted § 13.1's own test, since a capsid gives a vector its specificity and a payload gives a conjugate its activity. Regulators' definitions of the active substance include these parts: - EMA's draft guideline on mRNA medicines, and WHO's naming of mRNA substances, cover the full sequence; - FDA's ADC guidance names the linker and payload as constituent parts; - FDA's draft comparability guidance treats a tropism-changing capsid as a new product.

The four published lists count delivery-only products: all four count METiS's AI-formulated product, and Wellcome and Haque count Entos's vaccine. Greenblatt's scores place formulation below "AI-discovered", where this registry's borderline sits. No list counts manufacturing.

17.5 Re-run (15 Nov 2026). Eight products went through § 12.8: the seven the old rule had excluded (MTS-004, MTS105, MTS109, Covigenix VAX-001 and VAX-002, nirmatrelvir, STX-001) and ProQR's AX-0811, never assessed before. Two blind coders agreed on tier for 8 of 8 (κ = 1.00) and on tier and code for 7 of 8 (κ = 0.83). Red teams upheld both strict entries, with the step flag changed to "not stated" on both; upheld two borderline codings (MTS-004, STX-001); defeated one coder's code (MTS109, delivery); weakened two (the Entos vaccines); and moved nirmatrelvir from physics_only to other on new evidence. The maintainer session re-fetched every quote relied on, 14 from web pages and 3 from METiS's Hong Kong prospectus, and found each one. Result (v1.3.0): - Strict, ai_step = design, step not stated: AX-0811 (high confidence) and MTS105 (medium). For MTS105 the developer's paper says its MrNatty algorithm designed the whole mRNA sequence; the developer's patents filed before dosing and its prospectus describe its sequence tools as trained deep-learning and generative models, so the link to a trained model is platform-level, and borderline delivery is the recorded dissent. - Borderline delivery: MTS-004 (first dosed 2022, not 2025 as recorded before), MTS109 and Covigenix VAX-002. - Borderline platform_only: Covigenix VAX-001 and STX-001. Their AI wording appeared years after first dosing; DF-006 is the precedent. - Borderline other: nirmatrelvir. Pfizer's 2022 statement credits AI and machine learning with its design; its detailed accounts describe structure-based chemistry with no machine learning, as for iscartrelvir and GTA182. - The four exclusion rows these products sat in are removed. Not admitted: Quotient Sciences' unnamed generic in an ML-chosen formulation (Sep 2026), a test of the algorithm, not a candidate.

The same release admits the backfill review's products (docs/reviews/2026-11-15-backfill-signoff.md), coded under the existing rules. Together: strict 70 → 75, borderline 81 → 95, excluded 60 → 68. Both coders and the red teams are Claude models: the agreement figures measure consistency, not accuracy.

17.6 Completeness. A class-wide search (docs/reviews/2026-11-15-delivery-scope/CLASS-SEARCH.md, 18 web searches) found no other dosed product with a developer's or partner's AI claim for an mRNA sequence, a capsid, a conjugate's linker or payload, or an oligonucleotide sequence. Voyager's VY1706 is dosed but fails: its capsid came from in-vivo screening, with no machine learning named. Not covered: Chinese-language registries and company sites, Japanese and Korean developers, conference abstracts, METiS's own pages (blocked) and engineered-receptor cell therapies. Two leads outside these classes, UKK-0018 (Ukko) and mRNA-4200 (Moderna), wait for the standing review. Still not a census.

Approval. John E. Milad, 15 Nov 2026: "Approved" (the decision note's three questions, as recommended) and "Approve both" (this rule's sign-off pack and the backfill review's).