Which Evidence Settles Which Claim

Version 3.0 · 24 October 2026Version history

1.0, 23 October 2026: first publication (doi:10.5281/zenodo.23577317). 2.0, 23 October 2026, after a critique of version 1.0 received that day (it is not peer review): two listings of one pharmacy that bracket the claim's date within three days each side now support and no longer settle (P5); a medicine is investigational only while a study of the use is ongoing on the claim's date, a new sub-proposition (P3d); P8 is divided into the regulatory safety statement (P8a, per market) and a safety finding (P8b, international); "settles" is defined as a verdict on authority, distinct from interpretation and adjudication (section 2); a listing that does not show the product for sale, and a supply record of unstated kind, settle nothing; the machine-readable files, the test vectors, the reference implementation and the conformance rule are named (section 7); AlphaCitation's own software is described in its own section (section 9). Verdicts change for P3, P5 and P8 claims, so this is a major version (doi:10.5281/zenodo.23584875). 2.1, 23 October 2026: Annex B, on dividing an answer into claims and assigning each to a proposition, with a classification test set that measures a classifier and is not part of conformance (section 7); two clarifications: that a status claim qualified only by a phrase such as "for use", "for sale" or "for distribution" is a registration claim (P1); and that label content other than a condition, population, use, safety statement or clinical effect maps to no proposition (P2); plainer wording of the P5 rule in the brief and the summary table, which a reader had taken to mean that pharmacy listings only support; the annex of evidence classes renamed Annex A; section 9 naming version 2.0 where it describes software; the versioning of classification rules and the addresses that name no version (section 7); and section 10, open questions for the next major version. No verdict that a source receives for a proposition, market and date changes, so this is a minor version (doi:10.5281/zenodo.23588555). 3.0, 24 October 2026, after an external review of version 2.1 (it is not peer review): the claim's market is fixed by the market step, which takes the market the answer states, else the question's, else the market the deploying system declares, records which step governed and never infers a market (section 2; Annex B, M1 to M7); P6 is divided into P6a (reimbursed or covered, settled by a payer's decision within its reach), P6b (health technology assessment recommendation) and P6c (formulary listing), with two new evidence classes; a formulary the sentence does not name makes a formulary-listing claim that no source can be matched to (Annex B, C9 and E2); P8c, a safety signal, is added, with a new evidence class, and never establishes causation, and its negation is that a review found no new signal; the answer change is settled per market, by observations of that market; an availability claim with no word of breadth takes the same claim form as its negation (P5); a hedged claim is recorded as a hedged unit, visible as its own category and never judged (Annex B, E5); a stated date governs only its clause, an undated earlier time leaves the claim's date not determinable, and the answer's date is read in the market's local time (Annex B, T3, T4 and T7); section 7 states the release history, the policy for the stable version and how changes are proposed and recorded; classification under Annex B was measured in a pre-registered study and the results are reported (section 9). Before release, Annex B was sharpened where two blind labellers of a development set of real AI answers, and three machine readers of the test set, read its wording two ways, and the questions that versions 2.1 and the draft of 3.0 had left open were decided: a date the answer gives for an event dates an event claim, judged against the record of the event (Annex B, T6; Rule 7), and for a claim about a state, a stated period that contains the answer's date is read as the answer's date (T2); one unit per use, adverse effect, endpoint or study (U1); a status named only in a noun phrase, and a reference back to a claim already made, make no unit (U5); a list's lead-in makes no unit, a label's section heading supplies its predicate, and an entry whose heading is not given makes no claim (U7); a bare "Yes" or "No" asserts or denies the question's proposition (U9); P2 covers the legal category of supply, which in the EU the Commission's decision states, and the dose, route, form or restricted programme the authorisation permits, and a statement of who may receive the medicine, while a statement that a group must not receive it is a label warning, P8a (C2, C4, C7); P5 has a named-channel form, settled by the regulator's decision or label that provides a regulatory access route for the medicine by name or by the named supplier's own record, and a stock form, which nothing settles (C6); P6b covers added-benefit and value ratings (C9); "investigational" with no use named is read through P1 and "being studied" with none is P3d, P3c covers observational studies, and P4 covers a guideline named by its issuer, while a recommendation attributed to unidentified guidelines is a P4 claim that is not determinable (C10 to C12); a finding of no new safety signal is P8c negative (C14); a claim about a class makes its proposition for the class (E11); which uses of "may" hedge a claim (E5); a market attached to a noun, a body whose remit is part of a country, which governs a claim about its own decision, and places below the country, recorded by ISO 3166-2 codes, for which an authority's record speaks and a supply record does not (M1, M6); answers that differ at one time are not the answer change (C18). Verdicts change for claims whose market only the question or the deploying system states, for coverage, safety-signal and answer-change claims, for unqualified availability claims, for event claims, for claims about a place below the country, a named channel, stock, a class, the terms of a label, an HTA rating and a guideline named by its issuer or by none, so this is a major version.

The standard in brief

An answer about a medicine bundles separate claims: that it is approved, for whom, available, paid for, effective, safe, better than a competitor, recommended. Each claim except the recommendation can be settled only by the classes of evidence this standard names for it, in force on the claim's date; a claim about when an event took place is settled by the record of that event (Rule 7). For approval, indication, availability, reimbursement, health technology assessment, the regulator's safety statements and claims that AI answers changed, that means evidence that speaks for the market the claim is about: the market the answer states, else the one the question states, else the one the deploying system declares, and never one inferred from the user. For availability, a licensed pharmacy's listing dated the claim's date settles only that the pharmacy listed the product for sale that day, which settles a claim that the medicine is not available anywhere in the market and only supports a claim that it is available across it; two of its listings that both show it for sale, one dated before and one after the claim's date, each within three days of it, support at most. For trial, efficacy, safety-finding, safety-signal and comparative claims, it means international evidence; a safety signal never establishes that the medicine causes the effect. For a guideline or formulary claim, it means the named guideline or formulary, within its own scope. A recommendation is never settled: it is judged through the claims it rests on.

A conclusion may rest only on a source that settles; without one, the claim is unknown, never false.

1. Purpose

AI answers about medicines mix claims that only different kinds of evidence can settle. An answer that rules out a medicine as unapproved and unavailable, and recommends a competitor, carries a registration claim (settled by the market's regulator), an availability claim (settled by launch statements and supply records; an authorisation does not establish supply) and a recommendation (which no record settles).

The standard fixes in advance, for every market, which evidence may settle which claim. It decides the first of three questions about a claim: whether a source has the authority to settle it (section 2). The worked examples re-read two earlier AlphaCitation publications against the standard. AlphaCitation offers it for others to apply, test and cite; section 9 records how AlphaCitation's own software applies it.

2. Terms

• Proposition. One checkable statement inside an answer. Annex B sets out how an answer is divided into propositions.
• Evidence class. The kind of record a source is: one of 27, listed in Annex A.
• Authority. The body that issues a record. A record of a class only an authority issues (Annex A, first list) speaks for the jurisdiction of the authority that issued it and nothing else. An authority cited for a class it does not issue settles nothing: an EMA "authorisation" does not exist.
• The claim's market. The country or bloc the claim is about, as ISO 3166-1 codes, or ISO 3166-2 codes for a place below the country (Annex B, M6), fixed by the market step. It is the market the answer states for the claim, in the sentence or in a heading or earlier clause the sentence continues, where naming a market's regulator, payer or health technology assessment body states that market; otherwise the market stated in the question; otherwise the market the deploying system declares for its user, as configuration; otherwise it is unresolved (Rule 5). The first step that states a market governs, even a market word that names no market ("worldwide"); a later step that states a different market is recorded as a conflict and never governs. A market is never inferred from the user's address, language, spelling, currency, units or time zone. Every verdict records the step that fixed the market. Annex B (M1 to M7) applies the step to the words of an answer. For centrally authorised medicines, EU bodies speak for the 27 member states and for Iceland, Liechtenstein and Norway, which adopt corresponding decisions under the EEA Agreement; a nationally authorised medicine is settled by the national record.
• The claim's date. The date of the answer, or the earlier date the answer itself states for the claim ("as of October 2023"). Annex B (T1 to T7) fixes it clause by clause.
• Settles. A source settles a claim when it has the authority to decide it: it is of a class this standard names for the proposition, issued by an authority competent to issue it, speaks for the claim's market where the proposition is settled per market, meets the time test of Rule 7 (in force on the claim's date, or, for an event claim, the record of the event issued by the answer's date), and meets every further condition section 4 and section 5 set for it. That is the first of three questions. The second, interpretation, is whether the record, read correctly, concerns the claim's product, population, endpoint and date and says what the claim says. The third, adjudication, is whether the claim is then supported, contradicted or unresolved. This standard answers the first only (section 8). A source that settles is competent to establish what it records; its class alone does not establish that a reported effect is reliable. Where this text says that a source settles that something is so, or that a claim is settled false, it means that the source settles the claim and, read as the cited study read it, records that it is so or contradicts it.

3. Summary of the taxonomy

Sources must be in force on the claim's date, or, for a claim about when an event took place, be the record of that event issued by the answer's date (Rule 7). Section 4 names the classes that only support; any other class is irrelevant. P3 has sub-propositions P3a (trial registered), P3b (design), P3c (result reported) and P3d (study ongoing); P6 has P6a (reimbursed or covered), P6b (HTA recommendation) and P6c (formulary listing); P8 has P8a (regulatory safety statement), P8b (safety finding) and P8c (safety signal); a further proposition covers the answer change, settled per market. "P6" and "P8" are no longer proposition identifiers: a claim mapped to either is not assessable.

4. The ten proposition types

Each proposition is defined for a medicine. A claim about a class of medicines makes the same proposition for the class: for registration (P1), indication (P2), availability (P5), reimbursement (P6a), formulary (P6c) and regulatory safety-statement (P8a) claims it is about the class's members, and for the others only a source that speaks of the class itself can settle it (Annex B, E11).

Worked examples use only facts published in Field Note 07 [1] (GPT-4o mini through OpenAI's API with web search; Wegovy, UAE), Paper No. 01 [2] (EASD 2026) and the regulators' documents cited with them [3, 4, 5], and, in P5, the day of each availability answer, from Field Note 07's derived per-answer codes file, which it offers on request.

P1. Registered or authorised

Definition. The medicine holds a marketing authorisation, or a current registration, in a market.

Evidence. Settled by the competent authority's authorisation decision, register entry or authorised label for that market. A label in force is part of the authorisation record: the EU summary of product characteristics is Annex I to the European Commission's decision, the conditions of the authorisation, including the legal category of supply, are Annex II, and the labelling and package leaflet are Annex III. Supports, never settles: a regulatory opinion, institutional guidance, a manufacturer's or launch statement, commentary.

Example. Field Note 07 tested answers that said Wegovy was not approved in the UAE. The records of the class that settles that claim are the Emirates Drug Establishment's. Its announcement of 1 June 2026 is its own record of an authorisation decision: it states that it approved the oral form, and gives no registration date and does not mention the injection. Its directory, as recorded on 2 and 5 October 2026, listed the injection as registered on 14 November 2022 and the tablet on 4 May 2026. Field Note 07 states that those values rest on an AI-assisted transcription of a 5 October read, on records of 2 October and on a search-indexed copy of the injection's directory record seen on 13 October, that no directory page from those days was preserved, and that none of the values could be re-read on 19 October. The announcement therefore settles that the oral form is approved in the UAE, which contradicts an answer that, in the present tense and naming no form, says Wegovy is not approved there. A Dubai Health Authority document of 3 September 2024 calls the medicine "Approved by FDA and registered for obesity pharmacotherapy in UAE". It supports a UAE registration claim and settles nothing: it is guidance, and the FDA is a foreign regulator.

P2. Authorised indication, population or terms of supply and use

Definition. The medicine is authorised in a market for a stated condition, population or use, or on stated terms of supply or use, stated as what the authorisation permits: its legal category of supply (such as prescription only), its dose, its route, its form, or a restricted programme it requires for every supply.

Evidence. Settled only by that market's authorised label, in the version in force, and, for the legal category of supply or a restricted programme, also by the authorisation decision where it states it: in the EU, Annex II to the Commission's decision ("Conditions or restrictions regarding supply and use") states the legal category, and the summary of product characteristics does not. A register's category is not an indication, and a guideline is not the licence. A phrase that names the scope of an authorisation, such as "for use", "for sale" or "for distribution", is not a use: a status claim qualified only by one is P1 (Annex B, C3). A risk, warning or contraindication in the label is P8a, and an effect from its clinical section is P7. Supports, never settles: the authorisation decision (except as to the legal category or restricted programme it states), a regulatory opinion, a regulator's assessment report, a manufacturer's statement.

Example. Paper No. 01's Saudi arm concerned teplizumab (Tzield). Its US label, in the June 2026 revision the paper cites [3], indicates it to delay the onset of Stage 3 type 1 diabetes in patients aged 1 year and older with Stage 2 disease, with a further accelerated-approval indication in patients aged 8 to 17 recently diagnosed with Stage 3 disease. The European Commission authorised it, as Teizeild, on 8 January 2026 to delay the onset of Stage 3 type 1 diabetes in patients aged 8 years and older with Stage 2 disease [4]. Neither settles a Saudi indication claim, and Paper No. 01 did not establish teplizumab's Saudi registration status. In these versions the two foreign labels differ on the youngest eligible age.

P3. Investigational

Definition. The medicine is being studied for a use for which it is not authorised in the market.

Evidence. No single source settles P3. It is composite: that an interventional study of that use is ongoing on the claim's date (P3d, international), and that the market's authorised label does not carry that use (P2, per market), each settled on its own terms. Where the claim names no use, its parts are that an interventional study of the medicine is ongoing (P3d) and that the medicine holds no marketing authorisation in the market (P1, per market). Only a status word ("investigational", "experimental") makes such a claim (Annex B, C10). A registered study that is not ongoing on the claim's date does not make a medicine investigational. A registry record, a publication of record or a manufacturer's statement supports P3.

Example. In Paper No. 01, retatrutide was named in 10 of 1,680 UAE after-window answers, and 8 were coded as stating its status correctly (that it is investigational, in trials or not approved). The paper marks these codes provisional, because UAE status was established only to 2 October 2026. Orforglipron was excluded because the EDE directory recorded it as registered from 2 April 2026. Under this standard that registration settles P1 only. A UAE claim that orforglipron is investigational for a use is settled through P3d and P2 separately: a trial of that use, ongoing or not, does not make it investigational in the UAE if its UAE label carries that use.

P3a, P3b, P3c and P3d. Registration (P3a) is settled by the registry record; a publication of record or a manufacturer's statement supports it. Design or planned endpoints (P3b) are settled by the registry record version in force on the date asked, or the published protocol; a later amendment cannot rewrite an earlier registration, and a publication of record supports it. A reported result of a study, interventional or observational (P3c), is settled by results posted to a registry or the study's publication of record; a conference abstract, a manufacturer's release or commentary supports it. That the study is ongoing (P3d) is settled only by its registry record in the version in force on the claim's date, stating an ongoing overall status: not yet recruiting, recruiting, enrolling by invitation, or active and not recruiting (ClinicalTrials.gov's terms; another registry's status is read as its equivalent, and a status with no equivalent settles nothing). A version stating the study completed, terminated, suspended or withdrawn settles that this study was not ongoing on that date; it never settles that no study of the medicine exists. A version that states no status, states it as unknown, or states an expanded-access status settles nothing. A publication of record or a manufacturer's statement supports P3d.

P4. Guideline recommendation

Definition. A clinical guideline recommends the medicine, or its class, for a use: the guideline the claim identifies by its name or by its issuer, or guidelines the claim does not identify, which no source can be matched to (Annex B, C12).

Evidence. Settled by the guideline the claim identifies, at the version in force, within its own stated scope and population; the claim's market does not govern. A claim that names only the issuer is about that issuer's guideline for the use. A claim that attributes a recommendation to guidelines it does not identify is not determinable (section 6): no source can be matched to it. A systematic review and commentary support P4.

Example. Paper No. 01 coded Singapore answers that framed SGLT2 inhibition as first-line instead of or ahead of metformin against a reference frame it fixed before results, the 2022 ADA/EASD type 2 consensus. Under the protocol count, 1 of 202 candidate answers before the congress and 1 of 53 after did so. Both framed it for high-risk patients only, consistent with that consensus's cardiorenal recommendations. First-line use together with metformin was reported, not counted. The paper, not the answers, chose the guideline. Under this standard an answer that names the guideline or its issuer makes a P4 claim settled by that guideline; an answer that attributes the framing to guidelines it does not identify makes a P4 claim that no source settles, and a framing attributed to no guideline makes none. Either may be compared with a reference frame, which settles neither.

P5. Launched or available

Definition. The medicine is on sale or supplied in a market.

Evidence. Settled by a manufacturer's dated statement that the product has launched in that market, or by a supply record (a distributor's listing or a procurement record) within its own reach. A licensed pharmacy's listing settles only that the pharmacy listed the product for sale on the listing's own date; a listing that does not show the product for sale settles nothing, and nor does a supply record that does not say whether it is a distributor's listing, a procurement record or a licensed pharmacy's listing: each is not determinable (section 6). Two listings of the same pharmacy that both show it for sale, one dated before the claim's date and one after it, each within three days of it, support and do not settle that it was listed on the claim's date: the product may have left the listing between the two reads. A listing dated the claim's date is judged on its own. A listing is not proof of stock and does not establish market-wide supply. A listing dated the claim's date therefore settles a claim that the medicine is not available anywhere in the market, which it contradicts, and only supports a claim that the medicine is available across the market. A claim that the medicine is available in the market without a word of breadth ("available in the UAE") takes the same claim form as a claim that it is not available anywhere there: a claim and its negation are one proposition, and the evidence that settles one settles the other (Annex B, C6). A claim that the medicine is available across the market needs a word of breadth, and a claim that it is not available across the market is that claim's negation, settled by the same evidence. A claim that the medicine can or cannot be obtained through a named channel is settled, for a regulatory access route (a named-patient or special-access import route, compassionate use, a restricted programme), only by the market regulator's own decision that provides the route for that medicine by name: its authorisation decision or authorised label (a restricted programme the label requires), or its decision granting the route for the medicine (an import permit or a compassionate-use authorisation that names it), each an authorisation_decision or authorised_label. For a named supplier, it is settled by that supplier's own supply record, as above. The regulator's published rules for a route, which name no medicine, support the claim and never settle it: they show that the route exists, not that the medicine can be obtained through it. A record of another channel supports it. A claim about stock is settled by no class this version names: a listing is not proof of stock, and the claim is not determinable (section 6). The authorisation decision, a register entry, the authorised label and other manufacturer statements support P5 and never settle it, except that an authorisation decision or label that itself provides a regulatory access route for the medicine by name settles a named-channel claim about that route: registration does not prove supply.

Example. In Field Note 07, 14 answers said Wegovy was not available in the UAE. The regulator's records cannot test them. Under an availability standard the study added after the first day of collection, each was coded false against one licensed UAE pharmacy chain's listings, read on 13 and 16 October 2026, which showed the injection and the tablet for sale at that chain. By the study's per-answer codes, 9 of the 14 were given on 13 or 16 October, the days of the reads, and 5 on 14 and 15 October, between them. For the 9, a listing of the same day settles that the chain listed both forms, which contradicts "not available" in the UAE. For the 5, the listings either side support that contradiction and do not settle it. Eight of the 14 also said Wegovy was not approved. Those claims are judged under P1, or under P2 where the answer names a use, against the regulator's records and not the listings; Field Note 07 sets out the evidence for each approval answer, and the readings it lists as disputed, in its Exhibit 2. Of the six that claimed only that it was unavailable, the three given on 13 October are settled false; the three given on 14 October rest on no source that settles, so they are unknown, not false (Rule 8). Field Note 07 coded all 20 under its own stated rule and reports the listings as secondary evidence, not proof of stock on a shelf. Without the listings, only the approval answers count as false: 14 of 52.

P6. Reimbursed, recommended or listed

Definition. A claim about coverage takes one of three forms, judged separately. P6a, reimbursed or covered: a payer covers or reimburses the medicine, in the market or for the members of a payer or plan the claim names. P6b, HTA recommendation: a health technology assessment body recommends, or does not recommend, the medicine for a use, or rates its added benefit, therapeutic value or cost-effectiveness for a use. P6c, formulary listing: the medicine is on a formulary or list, such as a hospital's, an insurer's or a national formulary, or an essential medicines list; a formulary the claim does not name is recorded as unnamed, and no source can be matched to it (Annex B, C9). "P6" is no longer a proposition: a claim mapped to it is not assessable.

Evidence. P6a is settled only by a payer's coverage decision for its own market, within its reach. A decision that covers the market's population settles a claim of coverage in the market and a claim that names that payer; the decision of one payer or plan settles a claim that names that payer or plan and only supports a claim of coverage across the market; a decision that does not state its reach settles nothing. Whether a decision's indication, population and eligibility are the claim's is interpretation (section 8). An HTA appraisal, a formulary listing, a manufacturer's statement and commentary support P6a and never settle it: a recommendation that does not bind a payer does not establish reimbursement. P6b is settled only by the published appraisal of the body the claim is about, for the jurisdictions the appraisal speaks for (a NICE technology appraisal speaks for England; its application in Wales is the Welsh Government's direction, a different record); another jurisdiction's appraisal is a declared benchmark (Rule 2), and a manufacturer's statement and commentary support it. P6c is settled only by the named formulary's own listing, at the version in force, within its own stated scope; a payer's decision, a manufacturer's statement and commentary support it. An authorisation establishes none of them.

Example. Neither publication tested reimbursement. The pharmacy listings in Field Note 07 are supply records, and a supply record does not establish coverage.

P7. Efficacy result

Definition. The medicine produced a stated effect on a stated endpoint in a stated population.

Evidence. Settled by the trial's publication of record, a systematic review, a regulator's assessment report or an authorised label's clinical section, all international. Posted results, a conference abstract and a manufacturer's statement support P7. An unappraised posted result is a reported result (P3c), not an efficacy result.

Example. Two Gemini answers in Paper No. 01's UAE series, found by hand check, stated the headline figure of OCTANE, a retrospective real-world cohort added to the study after registration and presented as a late-breaking abstract at EASD 2026. The abstract supports the claims that the result was reported (P3c) and that it holds (P7), and settles neither.

P8. Safety

Definition. A claim about a medicine's risks takes one of two forms, judged separately. P8a, the regulatory safety statement: the medicine's authorised label in a market, or that market's regulator in a safety communication, states a risk, warning, contraindication or adverse effect. P8b, a safety finding: the medicine is associated with a stated adverse effect. P8c, a safety signal: a regulator or pharmacovigilance body has identified a signal that the medicine may be associated with a stated adverse effect, or is assessing one; its negation is that such a body has identified no new signal, or has closed its assessment of a signal without finding a causal association. A safety claim not assigned to P8a, P8b or P8c maps to no proposition (section 6).

Evidence. P8a is settled only by that market's authorised label, in its safety sections, and by that market's regulator's safety communications. A label or communication from another market supports it as a declared benchmark (Rule 2), and a systematic review is irrelevant to it: a review cannot say what a label carries. P8b is settled by a systematic review (international). A trial's publication of record, a regulator's assessment report, a label's safety section and a regulator's safety communication, from any market, support P8b and do not settle it. P8c is settled by a regulator's or pharmacovigilance body's safety signal record, or by its safety communication, from any market: the claim is that a body raised the signal, which the body's own record establishes wherever it sits. A P8c settlement establishes that the body identified, or is assessing, the signal, and never that the medicine causes the effect; a negative settlement establishes only that the body's review identified no new signal, or closed its assessment, and never that the medicine does not cause an effect; a signal record supports P8b and never settles it. A regulator's assessment report supports P8c. Case reports, forums, commentary and manufacturer statements support all three.

Example. On 8 October 2026 a Claude answer in Paper No. 01's Saudi arm, which was added after registration and is exploratory throughout, repeated a media round-up's wording on the first real-world teplizumab courses in Europe, including "temporary drops in white blood cell counts, rashes, and gastrointestinal issues". The paper identified no individual case safety information: the terms come from a published media summary as repeated by the answer. The answer makes a safety finding (P8b), which a media summary supports at most and a systematic review would settle, wherever the question was asked. Whether the Saudi label or the Saudi Food and Drug Authority states those risks is a separate claim (P8a), settled only by a label it authorised or by its safety communications. Paper No. 01 did not establish whether teplizumab is registered in Saudi Arabia.

P9. Comparative

Definition. One medicine is better, worse or no different from another on a stated endpoint.

Evidence. Settled by a head-to-head trial or an appraised indirect comparison with stated limits, both international. Commentary supports P9. Separate trials set side by side neither settle nor support.

Example. SMARTEST, one of Paper No. 01's registered readouts, compared dapagliflozin with metformin; no primary-results publication had been found by 11 October 2026. For the paper's Singapore series about empagliflozin it was class evidence, not trial evidence about the medicine asked about.

P10. Recommendation by the answer

Definition. The answer tells the reader to choose, use, continue, stop or avoid a medicine or a class of medicines.

Evidence. Nothing settles P10. A recommendation is never judged true or false; it is assessed through the propositions it rests on, P1 to P9.

Example. In 19 of the 20 answers Field Note 07 coded false (19 to 21 of 52 under the codes it lists as disputed), the medicine's non-approval or unavailability was the stated reason for recommending the competitor. The recommendation was not judged. Its grounds were: the registration claim against the regulator's records (P1) and the availability claim against the pharmacy listings (P5).

The answer change

A claim that AI answers changed is settled only by preserved observations under a longitudinal design: a fixed battery, the engine and parameters recorded, a statistical floor. A single answer supports and never settles. Paper No. 01 shows the design and where a study can fall short of it. It admitted 64 of 67 cycles and declared one registered movement, a verdict the paper says depends on the reading of its band rule. Its battery was a maintained set of question slots with text changes documented. The model field each engine returned was not recorded, and the paper's post-publication note records that, for Perplexity, the per-request parameters were not stored. Observations speak for the markets they observed: an answer-change claim is settled per market, under the market step (section 2), and observations of one market are a declared benchmark for another (Rule 2). Whether observations speak for engines, parameters or framings of a question they did not observe is interpretation (section 8). A claim that answers differ at one time, between engines or framings, states no change over time and is not the answer change (Annex B, C18).

5. Rules that hold across every proposition

Rule 1. Each proposition names the only evidence that can settle it.

Rule 2. A foreign regulator's record never settles a local claim. Where a proposition is settled per market (P1, P2, P5, P6a, P6b, P8a and the answer change), the source must speak for every place of the claim's market (each country of a group, or the subdivision a claim is about: Annex B, M6), so one country's authority never settles a claim about the Gulf states together, a subdivision's body never settles a claim about its country, and a supply record speaks for the places Annex B, M6, assigns it. A record for another market is a declared benchmark: it supports and never grounds a finding. Paper No. 01, written before this standard, likewise recorded US and EU status for context and never scored it.

Rule 3. A register settles registration only, and "not listed" settles nothing. A register entry settles P1 and nothing else. A read that found no matching record, found the product only under a status not read as current, or could not read the register settles nothing. A copy whose currency is not assured settles only up to the date its currency is evidenced. On 19 October 2026 the EDE directory held 1,179 products and returned no entry for either form of Wegovy [1]. That read decided nothing.

Rule 4. An EMA opinion is not the European Commission's decision. The EMA recommends; the Commission decides. An EMA opinion supports an EU authorisation (P1) and never settles it. A Commission decision is matched to its opinion by the EMA procedure number, never by date: a later decision, even one citing an opinion adopted on the same date, may concern another procedure, such as an annual renewal.

Rule 5. An unknown or unspecified jurisdiction is unresolved, never false. Where a proposition is settled per market, a claim whose market the market step (section 2) leaves unresolved, because no step states one or the one stated is unrecognised or "global", or a source that does not state its market, is not determinable.

Rule 6. Institutional guidance is not authorisation. Bodies that do not authorise medicines, such as the Dubai Health Authority, issue guidance, which supports a registration claim and settles none.

Rule 7. No retroactive judgement. An answer is judged only against evidence in force on the claim's date or, for an event claim, against the record of the event issued by the answer's date (below). For a claim about a state, a source that came into force later, or had ceased to be in force, does not settle it: a later decision is an evidence change, not a historical error. A pharmacy listing is in force on its own date only; two listings that bracket the claim's date support at most (P5). A source with no effective date, or a date that is not valid ISO 8601, decides nothing. Sotagliflozin was authorised in the EU on 26 April 2019 as an adjunct to insulin in adults with type 1 diabetes and a BMI of at least 27 kg/m² whose glycaemic control was inadequate on optimal insulin, and the authorisation was withdrawn on 22 March 2022 at the marketing-authorisation holder's request [2, 5]. An answer calling it "approved for type 1" must therefore be read against the record of its own date. Two Field Note 07 answers dated their claim "As of October 2023": the EDE's announcement of 1 June 2026 approving the oral form cannot speak to that date, and only the injection's directory registration date could, a value Field Note 07 could not re-read on 19 October 2026.

An event claim (Annex B, T6) asks when an event took place. It is judged against the record of that event, of a class that settles the proposition, issued on or before the answer's date. A record issued after the answer's date does not settle it. A record that later ceased to be in force, because the authorisation was withdrawn, the label revised, the appraisal replaced or the registry record updated, still settles when the event it records took place, unless before the answer's date it was annulled, retracted or corrected as to that event. A register copy settles the date it records for the event whatever its currency on the answer's date; Rule 3 bounds claims about a state. An answer of October 2026 saying that sotagliflozin "was approved in the EU in April 2019" is settled by the decision of 26 April 2019, although the authorisation was withdrawn in 2022. Whether the date the record gives for the event falls within the claim's date, or, for an undated event claim, on or before the answer's date (Annex B, T6), is adjudication (section 8).

Dates are never merged: claim, event or decision, entry into and exit from force, currency, retrieval and recording. A superseded record is closed, never rewritten.

Rule 8. Unknown is not false. A claim that no source settles is unknown, whether its sources only support it, are irrelevant to it or are not determinable (section 6): no finding may be raised on it, and it is never counted as an error. A hedged claim (Annex B, E5) is recorded and reported as hedged; it is not judged, and it is never counted as an error, whatever settles it.

6. Verdicts

Each source gets one verdict per proposition, market and date; two listings of one pharmacy that bracket the claim's date under P5 get one verdict together, which is at most Supports. Every verdict decides authority only (section 2).

Every verdict records the rule that decided it, a one-sentence reason, the step that fixed the claim's market and, for the checks it reached, its limitations by dimension: authority, jurisdiction, time and accessibility. Abstention is per claim: one claim that is not assessable leaves the rest of the answer assessable.

7. Citing the standard, licence and versions

Ezeife OC. Which Evidence Settles Which Claim: The AlphaCitation Proposition Taxonomy and Evidence-Authority Standard, version 3.0. Dubai: AlphaCitation Research; 24 October 2026. doi:10.5281/zenodo.23594909

Licence. The text is licensed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). Anyone may copy, adapt and apply it, including commercially, with attribution, a link to the licence (https://creativecommons.org/licenses/by/4.0/) and an indication of changes. Quoted third-party text is excluded from that grant, and no trademark, AlphaCitation's included, is licensed.

Versions. Versions are numbered semantically. A change that can alter an existing verdict raises the major number. Such changes include what settles a proposition, a jurisdiction or time rule, the meaning of a verdict, and any addition (a proposition, an evidence class, an authority, a market) that changes the verdict a source already judged would receive. An addition that changes no such verdict raises the minor number. A correction that changes no rule raises the patch number. Every change is recorded in the version history, every version stays available, and the DOI's concept record resolves to the newest. A finding cites the version it was judged under. The rules for classifying the claims in an answer (Annex B) are versioned in the same way: a rule that changes how a sentence an earlier version classified is classified, or that assigns a sentence to a proposition whose definition in section 4 does not cover it, raises the major number; a rule that classifies a sentence no earlier version classified, within the definitions of section 4 and without deciding what settles it, raises the minor number. The classification test set deposited with this version (classification-test-set-v3.0.json) measures how well a classifier applies Annex B to divide answers into propositions; that measure is distinct from conformance, which the test vectors alone decide. AlphaCitation's measurement of three machine readers, against the 123 items the set held before Annex B was sharpened and under that draft, is deposited beside it (classification-validation-v3.0.md), and the pre-registered study of the released text is deposited in two records of its own: its protocol (doi:10.5281/zenodo.23600058) and its results with the benchmark (doi:10.5281/zenodo.23602110; section 9).

Machine-readable files and conformance. Three machine-readable files are deposited with each version from 2.0. The file names keep the series name "v1"; each file states the version it belongs to. The newest version's files are also served at addresses that name no version, https://alphacitation.com/standards/evidence-authority/latest/ followed by standard.pdf, standard.json, schema.json, test-vectors.json, classification-test-set.json or reference-implementation.zip, and each version's record at https://alphacitation.com/standards/evidence-authority/ followed by its number (2.0, 2.1, 3.0). The machine-readable companion (evidence-authority-standard-v1.json) states, with identifiers, the proposition types and sub-propositions, the classes that settle and the classes that support each, their jurisdiction and time rules, the 27 evidence classes and the 9 issued only by an authority, the verdicts, the limitation dimensions, the rules of section 5 and the order in which the conditions of section 6 are applied. Its JSON Schema is evidence-authority-standard-v1.schema.json. The test vectors (evidence-authority-standard-v1.test-vectors.json) give one case for each worked example of section 4 and cases at the edges of each rule; each gives its inputs, the expected verdict and the rule that decides it, and marks every input that is not a published fact as synthetic. Where a file and this text differ, the text governs. An implementation conforms to a version when it reproduces every test vector of that version: for each, the claim's date, the verdict, the deciding rule and, where the vector states them, whether the registry record settles that a study was ongoing or that it was not (P3d) and the step that fixed the claim's market, with whether a later step conflicted. AlphaCitation's reference implementation, @alphacitation/evidence-authority, is licensed under the Apache License 2.0; its source is deposited with each version from 2.0 (version 3.0.0 with this version) and served at https://alphacitation.com/research/standard-v1/alphacitation-evidence-authority-3.0.0.zip. Version 3.0 changes three test vectors of version 2.1, each with a note saying why (EAS-V044, EAS-V050 and EAS-V057), and adds 68 (EAS-V094 to EAS-V161), 161 in all: an implementation that conforms to version 2.1 conforms to version 3.0 only once it applies the market step, the divided P6, P8c, the answer change's jurisdiction, the time test for event and undated event claims, the stated period that contains the answer's date, the P5 named-channel, stock and channel-kind forms, the P2 terms of supply and use, P3 with no use named, guidelines and formularies the claim does not identify, and places below the country by ISO 3166-2 codes. It decides from the source metadata it is given; it does not fetch, verify or interpret documents, and it holds no evidence. It takes the answer's date as supplied: a caller applies T7 by giving the date, or a date-time in the market's local offset, whose calendar day is the answer's date.

Stability, release history and proposals. Version 3.0 is the stable version. It changes only by corrections that change no rule (patch versions) and by fixes to security issues in the published implementation or API; any other change, including one that would raise the minor number, waits for the next major version. The release history is the version history above; every version stays available at its DOI. Proposals are recorded in a public issue register, https://alphacitation.com/standards/evidence-authority/issues, each with its status (open, adopted, declined or deferred), the version it targets and its resolution, and the process is set out at https://alphacitation.com/standards/evidence-authority/governance. Proposals received while 3.0 is the stable version are recorded and considered for the next major version. External reviewers are named only with their consent, and an invitation to review implies no endorsement.

8. What the standard is not

It is neither clinical judgement nor a regulatory determination. It does not say whether a medicine is right for a patient, grades no clinical evidence and decides no legal status; the authority's record remains the determination.

It decides authority, not truth. Interpretation (whether a record, read correctly, concerns the claim's product, population, endpoint and date and says what the claim says) and adjudication (whether the claim is then supported, contradicted or unresolved) belong to other steps, as does whether a source may be used at all (its licence and terms). A source of a settling class is competent to establish what it records; whether a reported effect is reliable is a question of the evidence's quality, which frameworks for grading evidence address and this standard does not.

9. Implementation status

This section records what AlphaCitation's own software does, and how closely AI models applying Annex B agree with reference labels, as of this version's date. It is not part of the standard's rules. The software it describes changes without a new version of the standard; this section is not updated between versions.

AlphaCitation adopted the standard on 22 October 2026 for its own findings about AI answers. Its software checks a finding decided by a regulator's label or register before AlphaCitation publishes it, whether a person verifies the finding or the accuracy gate publishes it, and from the release of version 3.0 it applies version 3.0. A finding records no market of its own: the check reads the market for which the scan that raised it was configured. The software does not yet apply the market step to its own findings: it does not take a market the answer states, and it does not record which step fixed the market. By 24 October 2026 the accuracy gate had cleared no class of finding to publish without a person's verification. AlphaCitation publishes the reference implementation and an open API that returns this standard's verdict for a claim and the metadata of a source, at https://alphacitation.com/api/v1/evidence-authority/judge (described at https://alphacitation.com/api/v1/evidence-authority/openapi.json). The API applies version 3.0 from its release and returns the market step's result with every verdict. It decides from what it is sent: it does not fetch or verify sources, holds no evidence, and does not read claims out of free text.

How closely AI models apply Annex B. A pre-registered study (protocol doi:10.5281/zenodo.23600058; results and data doi:10.5281/zenodo.23602110) measured how closely six AI models, applying Annex B blind, agree with reference labels on 294 sentences of the answers that three AI engines (gpt-4o-mini, claude-haiku-4-5 and sonar, through their APIs with web search) gave to 96 neutral questions about medicines written by AlphaCitation. With Annex B in full, the three flagship models' F1 over the full record of each claim (its proposition, polarity, hedge, market, the step that fixed the market, its date and the date's kind) was 0.68 to 0.75, against 0.21 to 0.25 with a minimal guide of the definitions of section 4 and the record of B1 without Annex B's rules: 2.7 to 3.4 times as high. With the market step set aside, since the study's answer format, unlike B1, let a reader return M7 as a step, the figures are 0.72 to 0.79 against 0.30 to 0.37, 2.0 to 2.6 times as high. Their micro F1 over propositions was 0.853 to 0.905 (0.849 to 0.918 against the reference as corrected after the runs; 0.847 to 0.916 on the same 284 sentences under the frozen labels), and one of the three (0.905) was shown to exceed the pre-registered floor of 0.85 over all 294 sentences by the protocol's decision rule (Holm-adjusted p 0.005), although its interval, [0.873, 0.934], is wider than the precision the protocol planned (about plus or minus 0.026), and the protocol reports such a result as inconclusive, not as support; on the 207 public sentences alone none was (Holm-adjusted p 0.0504 for that model, above 0.05). Small models trailed their flagships in micro F1 over propositions by 0.08 to 0.19, and none was shown to be within the pre-registered 0.05 of its flagship. A model run twice on the same requests changed its propositions on 4% to 23% of the sentences (5% to 23% with calls the provider refused counted as abstentions, as the protocol reads). Only OpenAI's flagship reached the pre-registered 0.95 (0.959), and only with those calls asked again (0.946 with them counted as abstentions), so the hypothesis that every reader reaches it does not hold. Where the three flagships gave the same propositions, on 78% of the sentences, those propositions matched the reference on 97% of them (92% where only OpenAI's and Google's flagships, whose models did not write the labels, agree: 231 of 250 sentences). On the 207 public sentences alone, which anyone can rescore with the released data, the full-record figures are 0.65 to 0.74 against 0.22 to 0.26 (2.6 to 3.3 times as high; 2.0 to 2.5 with the market step set aside), the three flagships give the same propositions on 76% of the sentences, and a model run twice changes its propositions on 4% to 25% of them (6% to 25% with the refused calls counted as abstentions). Most disagreement is in whether a sentence makes a claim, the market step and the date's kind; the sentences the text did not decide are section 10, questions 3 to 10. The reference labels were written and checked by blind instances of one model (claude-opus-5-5), which is also one of the readers, and labelling of 100 of the sentences by people outside AlphaCitation is invited. A pilot under a draft of Annex B, on 123 constructed and published sentences, is reported in classification-validation-v3.0.md, deposited with this version.

10. Open questions for the next major version

This section is not part of the standard's rules. It lists questions whose answer would change a verdict, and so can be settled only in a major version (section 7). Each is recorded in the issue register (section 7) and open for comment.

Before its release, version 3.0 decided the sixteen questions that version 2.1 and the draft of 3.0 listed here: event dates and records of earlier events (Annex B, T6; Rule 7; Annex B, B6, the launch statement); places and payers below the country (M1, M6); P5 claims about a named channel, about stock and about the market as a whole (C6; section 4, P5); P3 with no use named (C10; section 4, P3); label content outside P2 (C2; section 4, P2); claims about a class (E11); reported results of studies that are not trials (C11; section 4, P3c); guidelines and formularies named only by their issuer, or not named (C9, C12; section 4, P4); answers that differ at one time (C18); a class member named only in a ground (E11); products entered under a heading (U7); lists of adverse effects and coordinated uses (U1); a market attached to a noun (M1); and a regulator's warning reported by the answer (C14). Two questions remain from the design, and the classification study added eight: each arises from items a blind adjudicator found that Annex B, as written, does not decide (issue register, EA-042 to EA-049).

1. Shortage registers. A claim about stock (P5, the stock form) is settled by no class this version names, so it is not determinable. Whether a regulator's register of medicine shortages settles it, for the market and the dates the register covers, is not decided.
2. Claims that a medicine is no longer studied. A claim that development of a medicine for a use has stopped is P3 negative (Annex B, C10). A registry record speaks for its own study only and never settles that no study of the medicine exists (P3d), so such a claim is in practice never settled. Whether a sponsor's dated statement that it has stopped developing the medicine for the use settles it is not decided.
3. Suitability. Whether a statement that a medicine is "unsuitable" or "not suitable" for a group is a contraindication or warning (P8a, Annex B, C2 and C14), a statement of who may receive it (P2, C4), or, where its subject is the medicine's use in practice ("their use in"), no proposition (X1).
4. What a medicine does, stated without a result. Whether an effect stated as what the medicine does or is used to do, with no study, magnitude or population ("is used to improve blood flow and lower pressure"), is an effect on an endpoint (P7, C13) or a description of its mechanism or use (X1).
5. A regulator's clinical hold. Whether a statement that a regulator placed a medicine or its study on clinical hold states that the study is not ongoing (P3d negative, C11) or only entails it (U5); whether the hold is an event (T6); and how the reason given for it is classified (C14, E8).
6. Tense in other languages. Whether T6 reads tense by its form or by its function in the answer's language: the French passé composé and the German Perfekt are present perfects in form and the ordinary narrative past in function.
7. The dates of an answer change. Which date an answer-change claim takes when the sentence states only the earlier of the two times it compares (C18; T1 to T7).
8. Drug interactions. Whether a label's statement of an interaction is a warning (P8a, C14), a term of dose (P2, C2) or, where it states neither a risk nor a dose, no proposition (X1).
9. A "Yes" that is not bare. When a "Yes" or "No" is bare (U9), and whether a statement of uncertainty about the question's proposition in the same sentence hedges it (E5).
10. The tense of a reported result. Where a sentence states a study's result with no word that reports it ("in the trial, it reduced the risk"), whether the tense of the result's own predicate is the tense of the report, which T6 makes an undated event in the past tense, or the result is a state on the answer's date.

How to propose a change. Write to analyst@alphacitation.com, with the case the standard decides wrongly or leaves undecided and the verdict you would expect. Each proposal is entered in the issue register (section 7). While a version is the stable version, a change to a rule waits for the next major version, with test vectors that show the change; a correction that changes no rule may be made as a patch version. Every change is recorded in the version history.

Disclosures

The author founded Akunudo LLC-FZ, which owns and operates AlphaCitation. AlphaCitation sells AI information intelligence services to life-sciences organisations and applies this standard to its own findings. The worked examples name medicines of Novo Nordisk (Wegovy), Sanofi (Tzield), Eli Lilly (retatrutide, orforglipron) and Boehringer Ingelheim (the empagliflozin series of Paper No. 01), and name dapagliflozin and sotagliflozin by their non-proprietary names. Paper No. 01 states that Novo Nordisk, Boehringer Ingelheim and Sanofi are commercial prospects of AlphaCitation and that none is a client. Paper No. 01 also states that AlphaCitation may present its findings to them when offering its services, and that its Saudi teplizumab series, used in the P2 and P8 examples, was added after registration to measure the medicine of a commercial prospect. Both publications state that AlphaCitation had sent analyses of the medicines they measured to the manufacturers' staff in commercial conversations before publication. The author was employed by Novo Nordisk from 2013 to March 2025, last as Pricing and Access Director for its obesity and diabetes portfolio in Asia-Pacific. He holds no shares, share options, pension entitlement or deferred compensation from Novo Nordisk, and has no employment or advisory role with it. No pharmaceutical manufacturer commissioned, funded or reviewed this standard.

Field Note 07 states that AlphaCitation has no commercial relationship with Eli Lilly, and that AlphaCitation may present that note's finding to Novo Nordisk or Eli Lilly when offering its services. The author holds no shares or share options in Sanofi, Boehringer Ingelheim or Eli Lilly, and has no employment or advisory role with any of them.

POST /api/v1/evidence-authority/judge
{
  "claim": {
    "proposition": "P1",
    "market": [
      "AE"
    ],
    "answerDate": "2026-10-14"
  },
  "evidence": {
    "judged": "single",
    "sources": [
      {
        "evidenceClass": "authorisation_decision",
        "authority": "ae_ede",
        "effectiveFrom": "2026-06-01",
        "currency": "assured"
      }
    ]
  }
}

POST /api/v1/evidence-authority/judge
{
  "claim": {
    "proposition": "P1",
    "market": [
      "AE"
    ],
    "answerDate": "2026-10-14"
  },
  "evidence": {
    "judged": "single",
    "sources": [
      {
        "evidenceClass": "authorisation_decision",
        "authority": "us_fda",
        "effectiveFrom": "2021-06-04",
        "currency": "assured"
      }
    ]
  }
}

POST /api/v1/evidence-authority/judge
{
  "claim": {
    "proposition": "P1",
    "market": null,
    "answerDate": "2026-10-14"
  },
  "evidence": {
    "judged": "single",
    "sources": [
      {
        "evidenceClass": "authorisation_decision",
        "authority": "ae_ede",
        "effectiveFrom": "2026-06-01",
        "currency": "assured"
      }
    ]
  }
}