Explore the evidence

AlphaCitation Research Series · Paper No. 01 · Does New Medical Evidence Change AI Treatment Answers?

A Longitudinal Study of Evidence Citation, Medicine Naming and Selected Treatment Framing Around EASD 2026

Package version v1.0.2. Frozen analysis: freeze record freeze/easd-2026-freeze.json, repository commit 6d1b45d360c8d74cdd026cf7d2d7eb256e2d3e07, frozen at 2026-10-12T09:50:57.567Z. First released with the paper on 12 October 2026 (v1.0), after a disclosure audit and the author's approval; this version, v1.0.2, released 16 October 2026.

What this package is

It is the study's evidence, published so that a reader can check every material finding of the paper against the files that produced it. It holds the registered protocol with every dated deviation (D0–D24), the endpoint rules, the frozen registered analysis and E1 outputs, the instrument bridge, the baseline and final coding with their ledgers and hand codes, the blind second coding pass, the E4 audit and reference values, the regulator records, the post-registration robustness analyses with a per-answer outcome table, the protocol and results of the optional human validation, the data behind every figure, the verifier's output, the freeze record and the verification scripts the founder approves for release.

Most files are byte-identical to the frozen files and can be matched against the sha256 in the freeze record. The rest were changed only where a published restriction requires it, by fixed rules that are logged in REDACTION-LOG.md with a count for every file. MANIFEST.json lists, for every file, its source, both hashes and the rules applied.

Two things are never in it: full AI answers and the question battery. Short answer spans quoted for a coding decision stay, as in the ledgers the coders worked from, except Gemini's: no Gemini answer text and no link Gemini supplied is in the package (see "Gemini" below).

Cycles are named by neutral codes (UAE-C01–UAE-C23, SG-C01–SG-C24, SA-C01–SA-C20), because the study's own cycle identifiers open live public pages (the source-identification policy). Every file uses the same code for the same cycle; the map from code to internal identifier is held by AlphaCitation.

Registered and post-registration. The registered analysis is in analysis/ and e1/. Everything in robustness/ is POST-REGISTRATION ROBUSTNESS (deviation D24): computed after the results were seen, by a separate script (not released), into separate files, and it never replaces a registered verdict. The three registered §5 descriptive source analyses that were never run before are reported there too, late, under D24. The paper refers to the measured medicines generic-first: semaglutide (Wegovy), empagliflozin (Jardiance), teplizumab (Tzield). The files keep the names the study used.

How the package maps to the paper and the supplement

How to verify a finding: three worked paths

Path 1. A headline E2 number back to the per-cycle counts (registered)

The paper reports that Claude's UAE answers named semaglutide or Wegovy, the registered E2 measure, in 256 of 378 answers in the after window (67.725%), a registered movement beyond the bridged band.

1. Open analysis/easd-2026-analysis.json and find the e2 entry with series = uae-glp1 and engine = anthropic.
2. after.k = 256 and after.n = 378: the pooled after-window count of answers naming the medicine, over every admitted after-window cycle (4 to 11 Oct; the 5 Oct cycle is excluded under D14). precision.afterRateExact = 0.6772486772486772.
3. The same file's perCycleNaming holds the governed count per cycle. The 7 after-window UAE cycles give, for Claude (byEngine.anthropic):

4. The band. band is the registered v4 single-cycle range, 40.7%–55.6% (protocol §4). offset is the bridge's v5−v4 offset, 0.0309 with 95% interval [-0.0617, 0.1173] (bridge/easd-2026-bridge.json, series.uae-glp1.offsetsV5minusV4.anthropic). Reading I1 adds the interval bounds to the band ends: bridgedBand = [0.3453, 0.6733].
5. The decision. 0.6772486772486772 lies above 0.6733, so movementDeclared = true. The distance beyond the edge is precision.distanceAsDecidedPp = 0.3949 percentage points, or 0.4409 on the band rebuilt from exact single-cycle rates (D19). The magnitude against the bridged pre-congress reference is magnitudePp = +16.8 (MAJOR on the house scale, reading I2).
6. Uncertainty. The question-clustered 95% interval, a post-registration robustness analysis (D16), is after.clusterCI95 = [0.5714, 0.7804]; it contains the band edge (precision.clusterCI95ContainsNearerEdge = true), so the verdict is described as borderline. It is never reversed. Under another reading of the band rule (post-registration robustness, D24 item 2, robustness/easd-2026-robustness.json item2BandReadings) the verdict changes; the registered reading I1 stands.
7. Attribution. attribution.e1Answers = 0: no Claude UAE after-window answer qualified under E1, so the registered wording is movement without congress sources.
8. The figure. figures/fig05-e2-band-distance.csv carries the same row: uae-glp1,anthropic,true,256,378,0.6772486772486772,0.3453,0.6733,0.345679012345679,0.6728395061728395,0.6733,0.3949,0.3965,0.4409,0.5714,0.7804,true,true,true.

The build script recomputes every step above from the public files (pooled rate, per-cycle sums, bridged band, decision and distance) and stops if any number differs.

Path 2. The diaTribe E1 entry back to the ledger (registered E1; cross-arm row post-registration, D18)

The paper reports that one E1 source carried a registered readout into an answer of another series, and that part of what it carries cannot be shown to have been on the page when the engine cited it.

1. e1/easd-2026-e1.json crossArm: readout TRIUMPH-2 (retatrutide), registered for uae-glp1, reached an answer of series saudi-t1d from engine anthropic in cycle SA-C17, day 8.4 after its presentation, with contentAtCitation = unverified. The registered latency table counts a readout only in its own series, so this row never enters it (D18).
2. e1/easd-2026-e1-ledger.json sources: the entry with url = https://diatribe.org/diabetes-research/top-diabetes-news-easd-2026 has class = d-media (news or medical media dated 28 Sep or later reporting congress data), pageDate = 2026-10-05, readouts = "EASD 2026 type 1 diabetes programme (agent set + teplizumab abstracts)", "TRIUMPH-2 (retatrutide)", "Other EASD 2026 data (not in the readout set)", and readoutsUnverifiedAtCitation = "TRIUMPH-2 (retatrutide)". The revisions list keeps the earlier classification (previousClass = d-media, revised 2026-10-11).
3. The same entry's contentAtCitation and adjudication fields explain why: there is no archived copy of the page, the answer repeats the wording of the real-world teplizumab (Tzield) item (verified at the time of citation), and it does not use the TRIUMPH-2 item (unverified, reported and never counted; D18, D19).
4. e1/easd-2026-e1-manifest.json: the entry with the same URL was first observed in an answer at 2026-10-08T16:31:57.389+00:00, cited by Claude in 1 answer, and fetched at 2026-10-10T17:31:36.316Z.
5. e1/easd-2026-e1.json perEngineCycle.saudi-t1d.SA-C17.anthropic: viaSource = 1 of n = 56: the one qualifying Claude answer in that cycle.
6. figures/fig08-diatribe-source-to-claim.csv lays out the same chain element by element, with the question text withheld.

Path 3. Figure 11 back to the per-answer table (post-registration robustness, D24)

Figure 11 and §6.1 of the paper report a post-registration, descriptive association, not a causal estimate: after the congress, Claude's UAE answers that cite one pre-congress local page (page P61, a local business, withheld) name semaglutide or Wegovy far more often than the answers that do not cite it.

1. figures/fig11-claude-cited-page.csv, rows with window = after: the answers citing the page sum to 143/152 named (94.1%) over 7 cycles, and the answers not citing it to 113/226 (50.0%).
2. robustness/easd-2026-robustness.json item12PageP61.c_whereClaudesRiseSits.byWindow, window after: citingP61 = 143/152 and notCiting = 113/226, the same numbers.
3. robustness/easd-2026-per-answer-outcomes.csv (skip the # comment line): keep the rows with series = uae-glp1, engine = anthropic, window = after and inGovernedPopulation = true. There are 378 rows, and 256 have namedRegistered = true: the two groups of Figure 11 together (143 + 113 named, 152 + 226 answers), and the registered pooled count of Path 1.
4. Per cycle, the per-answer rows of each after-window date equal the two Figure 11 rows of that date added together, and the perCycleNaming count of Path 1:

5. The split itself, answer by answer. In the same rows, citesLocalPageP61 marks the answers that cite the page: those with citesLocalPageP61 = true give 143/152 named and those with false give 113/226, exactly Figure 11's two groups. The column names no page: it is a true/false flag.
6. The echo, answer by answer. Keep every Claude UAE row with window = during or after (all captured answers from 29 Sep, not only the scored ones). Of the answers citing the page, 152/164 have echoesLocalPageSentence = true; of the answers not citing it, 0/316 do. Figure 11's two echo rows carry the same counts. An answer echoes when it shares a run of at least 40 normalised characters with one of the page's two sentences, which the robustness output names only by neutral labels (S1, a class statement; S2, an efficacy-figure sentence); the sentences themselves are withheld.

The build script checks every number in this path against the public files, including each cycle's reconciliation, and stops if any differs.

Verification and replication: two different things

(a) Computational verification of the published results is what this package supports. Every number in the paper can be traced to the files here and recomputed from them: Paths 1–3 above need no script, and code/verify.ts (with its reduced lib.ts), run on the public coding files, reproduces the frozen verifier output. The build itself recomputes every step of the worked paths and stops if any differs.

(b) Independent replication of the full measurement process is not possible from this package. It needs the active question battery, the capture pipeline and the production scorer, which are proprietary and not released (the release policy), and the held pull of the answers. What a replicator can do is run a comparable study: ask the engines comparable questions over comparable windows, apply the rules the protocol and endpoints/ENDPOINT-SPECIFICATIONS.md publish, and compare the results with the published ones.

Local-business identifiers are withheld from the publication and public evidence files under the study's disclosure policy. This restriction does not prevent independent researchers or AI systems from identifying or citing the same publicly accessible sources when conducting comparable measurements.

What can be reproduced from the public files, and what needs restricted data

"Checked at build" means the build script recomputes it from the public files and stops if it differs.

Human validation

The study had blinded human validation of the interpretive coding (50 sampled E3, E5 and E3-i judgements), by the author as a single reader: 47/50 agreed; the representative sample 17/17. It is not independent validation. Naming (Part A) was not hand-coded by the author's decision. Its re-scoring is a deterministic consistency check (the production scorer reproduces every governed per-cycle count); naming accuracy was not independently assessed against human-coded answers. validation/human-validation-protocol.md is the protocol and validation/human-validation-results.md the results, with every disagreement. The blind second coding pass (a separate instance of the same model family) is a consistency check, not independent validation either.

Restrictions and why

Question-instrument availability (the release policy, verbatim)

Question-instrument availability. The exact question battery is withheld to preserve the integrity
of AlphaCitation's longitudinal measurement instrument. Public disclosure of the active questions could
enable content to be optimized specifically against the instrument, potentially influencing future
observations and compromising comparability over time.

To support scientific scrutiny, the study reports its question taxonomy, sampling design, endpoint
definitions, coding procedures, analytical methods and reproducibility controls. The complete instrument
is preserved in the frozen study record.

This restriction represents a departure from CHART Item 5b and limits independent replication of the
exact data-collection procedure.

Consequences in this package: no question text appears anywhere; the semantic question identifiers in the coding files and the per-answer table are replaced by opaque stable identifiers (uae-q01, sg-f03, …; f marks a focused question, q a brand-neutral one, the classes of the paper's question taxonomy), because the original identifiers name each question's topic and so disclose battery composition (the robustness output's own question numbers, Q01…, are a different, text-keyed numbering); the fixed instruction that every question in all three batteries carries is withheld wherever it appeared (the suffix-stripping regex in second-pass/easd-2026-second-pass-builders.py (2); the quotation in deviation D13 of protocol/easd-2026-evidence-to-answer-prereg.md; the framing wording in a comment of ); and the which-medicine rule in is withheld because its phrases are tuned to the battery's wording. The rule is described in coding-scheme decision S4, and its result for every coded row is public.

Source identification (protocol §10, D22; the source-identification policy)

The three measured medicines are named (D22). The registered boundary protects local clinics, pharmacies and businesses: those located in the measured markets. One source-identification policy applies to every engine, window and series, and no source is named or withheld according to the direction of a finding. Its criterion is the major non-local organisation: an established organisation operating nationally or internationally under its own name, outside the measured markets, is named; anonymity is not extended to it without a specific reason. Every cited domain is classified in redaction-domain-classification.csv:

• Named (834 entities). Regulators, governments and intergovernmental bodies; journals and scientific publishers; medical and drug databases; academic institutions; academic medical centres and non-profit health systems outside the measured markets, and public university and government hospitals; professional societies and non-profits; news media, press-release wires and financial media; sponsors and manufacturers; reference works; market-research publishers and other established research organisations; international retailers; major non-local commercial companies, including consumer health companies and apps; and major non-local healthcare companies: pharmacy chains, insurers (at group level), telehealth companies, laboratory companies and hospital groups. A named source keeps its URL and title, except where the page was cited only by Gemini (see "Gemini" below).
• Withheld (1276 entities, each replaced by one stable typed token such as [local-pharmacy-07], with a page index for each of its URLs). Everything operated or headquartered in a measured market (UAE, Singapore, Saudi Arabia): local clinics, hospitals, pharmacies, laboratories, insurers and care providers, local businesses, and regional platforms and regional entities of international groups headquartered or operating there; franchise outlet pages of international brands; and non-local small businesses, clinics, pharmacies, practitioners, distributors, affiliates and personal sites, which are not major organisations. Entities whose operator could not be identified from the domain and page titles are not shown to be major and are withheld too; each was reviewed and none is left unresolved (0 flagged).
• Platform pages (49 platforms). Social media, video, maps, review, marketplace and file-hosting pages carry a platform token ([platform-page:<platform>]/pNNN): the platform is named, the account, seller or item is not.
• A public-entity page whose title or path names a withheld entity keeps its host but not its title, path or exact date (rules R1f and R1q, 28 pages). The names are matched as domain labels, written whole or split, as reviewed trading names (120 Latin and 43 Arabic-script aliases), and an Arabic title that names a business is withheld whenever the check cannot tell.
• Publishers are named wherever they are. News media, journals, medical databases and research publishers based in a measured market are named like any other publisher; the location rule concerns businesses and care providers, not publishers.

Every classification was reviewed entity by entity for this version; the full mapping is held by AlphaCitation in the restricted archive. The review notes are counted in REDACTION-LOG.md. redaction-token-map.json lists every token of this build with its type, citations and series (tokens only, no names); documents that quote tokens read them from there, because tokens are stable within a version, not across versions.

Cited is not supported, and supported is not caused. The package records that an answer cited a source: the engine listed it. Whether a source supports a claim is a separate judgement, made only where the E1 ledger reads the page and records its qualifying sentence, and it is shown to hold at the time of citation only where contentAtCitation says so. No file here shows that a source caused an answer to change; the association reported for page P61 is descriptive (D24).

Local-business identifiers are withheld from the publication and public evidence files under the study's disclosure policy. This restriction does not prevent independent researchers or AI systems from identifying or citing the same publicly accessible sources when conducting comparable measurements.

Discovering the same source independently is not an exact replication: the question battery is withheld (§2.8).

Page P61. The local page behind Claude's UAE rise belongs to a local business. Its company, domain and URL never appear, and neither do the three of its sentences that some answers repeat: every run of 40 or more normalised characters of any of them is replaced by a marker (rule R1i), including inside the frozen E4 spans that quote one of them, and every figure beside such a marker is withheld (rule R1k). The robustness output refers to two of them by neutral labels (S1, S2); the third (S3, a second efficacy-comparison sentence) was found by the disclosure audit and is checked the same way. The frozen robustness output counts an answer as echoing the page when it repeats S1 or S2 (it says "two sentences"); answers that repeat S3 are outside that frozen count. The build stops if any 40 characters of these sentences, its company name, domain, URL or page title appear anywhere in the package; the sentences are checked by window hashes and are never written here. In the public copy of the robustness script, the window-hash tables, the figure-text hashes and the hostname hashes are withheld (rules R1i, R1j and R1k), because overlapping window hashes would let a reader rebuild a sentence one character at a time and a hash of a short string confirms a guess.

Gemini (the Gemini reporting policy)

Gemini results are reported in aggregate; verbatim outputs and grounding links are not redistributed.

The analysis includes Gemini as registered; its results appear here as codes, counts and aggregates. In every public file, each Gemini answer span and figure string is replaced by a marker with the sha256 of the withheld string beside it (spanSha256, textSha256; rule R1m, 2907 spans and 1971 figure strings), and every quoted fragment of a Gemini answer in a coder's note is replaced by a marker (R1n, 333 fragments). 8 answers whose engine no public file records are treated the same way, under a neutral marker ([answer text withheld under the Gemini constraints; engine not recorded]). In the robustness output, the two Gemini answers of the text-route check show their figure as [Gemini figure text withheld], with the count and the reference value kept. No grounding-redirect link appears (R1l): an as-cited redirect reads [Gemini grounding redirect]. A page that only Gemini cited is given by a neutral token only; its URL, publisher, title and any quotation are withheld (R1o, 1575 pages, token [gemini-only]/pNNNN); a page another engine also cited keeps its URL under the source policy above. Domain classes and counts are unchanged. The build stops if any Gemini span, figure string, quoted fragment, run of 40 or more characters of Gemini answer text, redirect link or resolved Gemini-only URL survives (G15). Search Suggestions were never captured, and none are published.

Full AI answers

Never published. Coded spans of up to about 240 characters, quoted for a coding decision, stay in the ledgers; the few longer spans are listed in the redaction log for review. The per-answer table carries outcomes only, no text.

Operational and confidential details (protocol §10)

Withheld: the parameter changed at v5 and its values, version history beyond the fact and effect of the change, calibration figures, regex diagnostics and the production scorer's internals; credential environment-variable names; internal key namespaces, database row identifiers and local machine paths; internal deployment and job names; the production LLM budget the bridge shared (its daily cap, default ceiling and hard stop; rule R4h; the bridge's own metered costs stay); and one commercial-relationship detail in deviation D10. The local Gemini redirect cache is held and identified by its sha256.

Code

code/ holds 2 of the 15 scripts the freeze record hashes (14 study scripts and this package's builder, evidence-package.ts, which is withheld with the other 12): the verification utilities the founder approves for release (the release policy). They were changed only as listed here; every change is counted per file in REDACTION-LOG.md. Nothing from the application source (src/) is included.

Not released (implementation proprietary; the computed outputs and specifications are published): analyze.ts, bridge.ts, code.ts, cycles.ts, e1.ts, figures.ts, freeze.ts, pull.ts, regulatory-status-t1d.ts, regulatory-status.ts, release-scan.ts, robustness.ts. These scripts produced the frozen outputs in this package from the held pull and the production scorer. A reader can check every published number against the published outputs (Paths 1–3 below need no script), but cannot re-run these steps from the package: the analysis, coding, E1 screening, bridge, regulator snapshots, figures and robustness computations. Their rules are specified in the protocol, endpoints/ENDPOINT-SPECIFICATIONS.md and the coding-scheme decisions, and their hashes are in the freeze record.

The public lib.ts which-medicine regex matches nothing, so isChoiceQuestion in the public copy always returns false; the per-row choiceQuestion values in the coding outputs are the ones the study used.

scripts/easd-2026-study/evidence-package.ts, which built this package, is not included: it carries the classification lists the review depends on.

Version identifiers

• Package: v1.0.2, released 16 October 2026; v1.0.1 released 15 October 2026; v1.0 released 12 October 2026.
• Documentation revision, 13 October 2026: the README's release status updated; the naming check described as a deterministic consistency check, with naming accuracy not independently assessed (as in the paper and supplement); the corrected E1 attribution calculation added to the reproduction table, with the correction record it uses. No data file changed; every data file's sha256 is as released.
• Post-publication note PN1, 15 October 2026: a change in the structure of Perplexity's responses between 6 and 7 October 2026 (paper §6.2). Adds addenda/PN1-response-structure.csv and addenda/PN1-naming.csv, computed after publication from AlphaCitation's stored answer captures (structure only, no answer text) and from the stored naming counts. They lie outside the freeze record; their hashes are in MANIFEST.json. Every data file released in an earlier version keeps its sha256.
• Post-publication note PN1, corrected, 16 October 2026: the note's wording corrected (paper §6.2, Abstract, §11, Figure 13 caption and version history; supplement §4.1; executive brief; this package's DATA-DICTIONARY definition of the monitoring panel): version 1.0.1 described the monitoring panel as about 12 product categories unrelated to diabetes, although one of them concerns weight-management medication and tracks diabetes medicines, among them semaglutide, the medicine of the UAE series; called the study's post-registration Saudi series a question set unrelated to the congress; said that the instrument sent the same parameters throughout, although they changed at v5 on 27 September (paper §4.3); in the Figure 13 caption, widened the note's scope to every question set the instrument ran; and, in supplement §4.1 and the executive brief, said that every stored answer carried exactly 10 search results after the change, where at least 98% did on each day. Adds addenda/PN1-naming-panel-11.csv, the panel's daily naming counts in its 11 categories that track no diabetes or weight-management medicine, computed after publication from the stored panel readings by the measure of the monitoring-panel rows of addenda/PN1-naming.csv. It lies outside the freeze record; its hash is in MANIFEST.json. Every data file released in an earlier version keeps its sha256.
• Freeze record: repository commit 6d1b45d360c8d74cdd026cf7d2d7eb256e2d3e07, frozenAt 2026-10-12T09:50:57.567Z, 33 files and 15 scripts, problems: none.
• Not published: the held pull (question and answer text), pull-final.json, pulled 2026-10-11T16:56:27.066Z, 67 cycles; 110898974 bytes, sha256 de98a60469b0e7b100110f16170b16760cc32fe333696eb4292bf48219eed265. Any copy can be matched by this hash.
• Not published: the held local Gemini redirect cache (cited URLs, including local businesses); 9714719 bytes, sha256 0aa4612beaac5d2f2113ff2bf41129c3b7fb60e89159a65efdb11fbec3487427. Any copy can be matched by this hash.
• The package builder (scripts/easd-2026-study/evidence-package.ts) was revised after the analysis freeze, so its hash in the freeze record does not match the version that built this package. It touches no analysis: it only copies, redacts and checks. Its own sha256 is in MANIFEST.json (builder).
• The release scan (scripts/easd-2026-study/release-scan.ts), a release tool (no analysis), was revised with the package builder after the freeze; it is not released. The freeze record's hash is 8aeba6af218ea34bdaff0273de84f9c7400d239cee9d569ec8ee7c027f867153; the version used for this package is ae55ed219c75254a5c4d0e9557a9f8a805d527f0a072c331e7ab80824b84e3c5.
• Not published: the internal cycle-code map (docs/evidence/computed/easd-2026-cycle-codes.json), which links the neutral cycle codes to the study's cycle identifiers; frozen, 8848 bytes, sha256 a8f53c8b96ab06b78df4d74340db939bb49ca55d9f0f819ad5a2ad0cfbdfc619.
• robustness/easd-2026-robustness.json records the freeze in force when it was computed (84a2214c, 2026-10-11T20:23:25.294Z); all 9 of its frozen inputs are byte-identical to those in the current freeze record (6d1b45d3), which also hashes the Gemini redirect cache.
• Every public file's sha256 and its source's sha256 are in MANIFEST.json. Files marked verbatim there are byte-identical to the frozen file, so their hash equals the one in freeze/easd-2026-freeze.json.

Licences

Published by AlphaCitation Research, the research programme of AlphaCitation, a digital product of Akunudo LLC-FZ. © 2026 AlphaCitation (Akunudo LLC-FZ). AlphaCitation is a product of Akunudo LLC-FZ, a free-zone company registered in the United Arab Emirates. Licensed under the release policy; LICENCES.md lists every file and folder with its licence. In short: the study scripts the founder approves are in code/ under the MIT licence, and every other study script is proprietary and not in the package; the redacted evidence datasets are under CC BY 4.0 where Akunudo LLC-FZ holds the redistribution rights, and any file or field derived from third-party material (regulator records, cited-page titles and publishers, AI-answer excerpts) keeps its own rights, with the Gemini constraints of Gemini reporting policy; this documentation and the figures are under CC BY 4.0. The active question battery, AlphaCitation's measurement instruments, production software, confidential methodology implementations and unreleased datasets are proprietary and are not licensed. Nothing outside this package is licensed by it.

Licence and third-party material

Third-party material in this package is reproduced as follows. It implies no endorsement by any source.

• European Medicines Agency. Source: European Medicines Agency (medicines data; EPAR product information), ema.europa.eu, accessed October 2026.
• Health Sciences Authority, Singapore. Contains information from Listing of Registered Therapeutic Products accessed on 2 and 11 October 2026 from data.gov.sg, which is made available under the terms of the Singapore Open Data Licence version 1.0 (https://data.gov.sg/open-data-licence).
• US Food and Drug Administration. FDA data via openFDA (CC0).
• diaTribe and sponsor releases. The E1 ledger quotes one qualifying sentence from each page it classifies, with the page's URL: among them The diaTribe Foundation, "Top Diabetes News: EASD 2026" (diaTribe, 5 October 2026, https://diatribe.org/diabetes-research/top-diabetes-news-easd-2026), and the press releases of the sponsors and manufacturers named in the ledger. Each quotation is attributed to its source.
• Emirates Drug Establishment, Saudi Food and Drug Authority and EASD. Register entries and programme details are factual excerpts, each with its source named; they imply no endorsement by these bodies.
• AI answers. The answer excerpts are the study's own API outputs, each labelled by engine (ChatGPT, Claude, Perplexity). Gemini results are reported in aggregate; verbatim outputs and grounding links are not redistributed.

Correspondence and corrections

AlphaCitation Research is committed to open scientific scrutiny. We welcome methodological critiques, requests for clarification and substantive commentary on our published research. Material corrections and updates will be documented in any subsequent versions.

Scholarly correspondence, methodological critiques and correction requests may be directed to analyst@alphacitation.com. Substantive corrections will be evaluated and documented in the publication's version history.

Name the file and field (or ledger row) concerned. Corrections are issued as new versions of the paper, the supplement and this package.