# Figures: AlphaCitation Research Series No. 1 (EASD 2026)

Generated by `scripts/easd-2026-study/figures.ts` from the frozen record of the analysis of 11 Oct 2026 (D20, with the errata of D21), every input matched to the freeze record `docs/evidence/computed/easd-2026-freeze.json` (frozen 2026-10-12T09:50:57.567Z at repository commit `6d1b45d3`). Every value drawn is in the figure's CSV. Nothing is read from the database, the network or a temporary file.

Measured medicines are named, by the founder's decision (D22): UAE, semaglutide (Wegovy); Singapore, empagliflozin (Jardiance); Saudi Arabia, teplizumab (Tzield). Clinics, pharmacies and local businesses are not named on any figure (§10, D22); the script stops if any cited hostname other than the source shown in Figure 8 appears in an output.

Cycles are identified by neutral cycle codes (for example UAE-C01); cycle slugs are not published.

Visual language: one neutral ink and one accent (blue: registered data, bridged bands); ochre marks everything added after registration (the Saudi arm, post-registration readouts, robustness intervals); hatching marks material that is unverified or quarantined; dashed outlines mark what was not measured or not admitted. Colour never carries meaning alone: each figure also labels it, or uses marker shape or dash pattern.

## Figure 1. Study timeline

Files: `fig01-study-timeline.svg`, `fig01-study-timeline.csv`

**Caption.** Panel A: the four observation windows of protocol §3 (before on instrument v4; before on v5, from the first v5 cycles of 27 Sep to 28 Sep 06:00 UTC; during, 28 Sep 06:00 – 2 Oct 22:00 UTC; after, 2 Oct 22:00 – 11 Oct 23:59 UTC); the instrument change v4 → v5 (27 Sep); the congress (28 Sep – 2 Oct); the registration of the protocol (2 Oct, 09:50 UTC, commit time per D21) and the dated deviations that followed (ochre); and every cycle of the three series in the frozen pull, admitted (dot) or excluded (cross, numbered, with the reason). Panel B: presentation times of the seven registered readouts and of the readouts added after registration (OCTANE, LBA 59, D3, presented 30 Sep 09:45 UTC; the Saudi type 1 diabetes programme, D10; other EASD 2026 data, the latency anchor for E1 classes c and d). Units: UTC dates and times. N: 67 cycles in the pull, 64 admitted (UAE 22, Singapore 23, Saudi Arabia 19). Admitted cycles are marked at their recorded scan-completion time (perCycleNaming.scannedAt): UAE about 14:05 UTC throughout (first cycle 17:44); Singapore about 14:03 UTC throughout (first cycle 17:42); Saudi Arabia about 14:05 UTC before the congress and 16:34 UTC from 4 Oct (first cycle 19:40). Cycles were started by the scheduler (14:00 UTC, and a second scheduled job at 16:30 UTC, D20) and completed a few minutes later; from 5 Oct every series ran daily (D9). Excluded cycles have no committed scan time and are marked at mid-day of their UTC date.

**Observation period.** 22 Aug – 11 Oct 2026 (panel B: 28 Sep – 2 Oct 2026)  
**Units.** UTC date and time  
**N.** 67 cycles (64 admitted, 3 excluded); 10 readouts (7 registered, 3 post-registration)

**Source files and fields.**

- docs/evidence/computed/easd-2026-final-coding.json → admissions[] (series, slug, date, window, engineVersion, admitted, reasons, promptsCaptured, promptCount)
- docs/evidence/computed/easd-2026-saudi-final-coding.json → admissions[]
- docs/evidence/computed/easd-2026-e1.json → readouts[] (id, series, presentedUtc, postRegistration)
- docs/design/easd-2026-evidence-to-answer-prereg.md → header (registration time), §3 windows, §9 D9, D10, D15, D20
- docs/evidence/computed/easd-2026-bridge.json → series.*.ranOn

**Key reading.** Every after-window observation sits on instrument v5. One admitted v5 cycle precedes the congress in the UAE series; none does in Singapore (its 27 Sep cycle failed the capture rule) or in Saudi Arabia (its 27 Sep cycle failed the capture rule). That is why the v4 bands must be bridged rather than reused.

## Figure 2. Measurement architecture (proposed framework)

Files: `fig02-measurement-architecture.svg` (conceptual; no data, so no CSV)

**Caption.** A proposed conceptual chain from the availability of new evidence to downstream behaviour: evidence available; discovered or retrieved by the engine; cited as a source; represented accurately in the answer; integrated into the answer (naming, positioning, status statements); preference and behaviour. For each stage the figure states which endpoint of this study observes it and with what registration status, and marks the two stages the study cannot observe (retrieval without citation; user preference, prescribing and demand). E4 is shown as quarantined (D17). The bracket marks the registered attribution rule (I3), which links citation (stage 3) to integration (stage 5). Conceptual figure; no data are plotted.

**Observation period.** Not applicable (conceptual)  
**Units.** None  
**N.** None

**Source files and fields.**

- docs/design/easd-2026-evidence-to-answer-prereg.md → §3 readout set, §4 endpoints E1–E5, §9 D3, D10, D16 (I3), D17, D19

**Key reading.** The study observes citation, stated content and answer integration, but neither retrieval that leaves no citation nor any downstream behaviour, so no result in the paper speaks to treatment preference or prescribing.

## Figure 3. Observation flow

Files: `fig03-observation-flow.svg`, `fig03-observation-flow.csv`

**Caption.** Flow of observation units in the frozen pull (22 Aug – 11 Oct 2026): cycles run, excluded (with the failed admission rule and its deviation number) and admitted per series; admitted cycles and captured answers per window (E2's scored population is a different set and appears only in Figures 4, 4b and 5); the after-window denominators of each series' endpoints; and, for the Saudi exploratory endpoint E3-i, the chain from molecule namings to the seeded sample (at most 60 per molecule per window) to namings coded as a type 1 diabetes use and those with no statement that the use is unapproved. Units: cycles, answers, or namings (answer × molecule) as labelled. N: 67 cycles, 64 admitted.

**Observation period.** 22 Aug – 11 Oct 2026  
**Units.** cycles; answers; namings (answer × molecule)  
**N.** 67 cycles; 64 admitted

**Source files and fields.**

- docs/evidence/computed/easd-2026-final-coding.json → admissions[]
- docs/evidence/computed/easd-2026-saudi-final-coding.json → admissions[], headline.e3i.byWindowAgent, e3i.rows[].code
- docs/evidence/computed/easd-2026-analysis.json → inputs.pull, e2[].beforeRef/during/after, e3.namingsByWindow, e3.clinicallyMaterialByWindow, e5, saudi.opportunity (answers, opportunityAnswers), saudi.e3NamingsByWindow, saudi.e3iExploratory

**Key reading.** 3 cycles were excluded under rules fixed before results (two pre-congress capture failures and the cached UAE cycle of 5 Oct), and each endpoint rests on its own denominator, from 1,680 UAE after-window answers for E3 to 39 sampled Saudi type 1 use namings for E3-i.

## Figure 4. UAE naming trajectories

Files: `fig04-uae-naming-trajectories.svg`, `fig04-uae-naming-trajectories.csv`

**Caption.** Share of scored UAE answers that name the brand or its molecule (semaglutide or Wegovy; the registered E2 measure, registered primary endpoint) for each engine and for the four engines combined, in every admitted cycle (13 v4 cycles; 2 v5 cycles before the after window; 7 after-window cycles), from the governed per-cycle counts. Error bars are Wilson 95% intervals computed for this figure (not part of the registered analysis). Grey band: the registered v4 single-cycle range (§4). Tinted band: the bridged band against which the after window is judged (§6 offset and its 95% interval, I1). Right column: the after-window pooled rate (K/N shown) with its question-clustered 95% interval (post-registration robustness, D16/D19; never a replacement for the registered rule); filled marker = movement declared, open = inside the band. Units: per cent of scored answers. N: 49–54 scored answers per engine per cycle; after window Perplexity 378, Gemini 378, ChatGPT 378, Claude 378, 1,512 combined. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** 22 Aug – 11 Oct 2026  
**Units.** per cent of scored answers naming the medicine  
**N.** 22 admitted cycles; after window 1,512 scored answers

**Source files and fields.**

- docs/evidence/computed/easd-2026-analysis.json → perCycleNaming[series=uae-glp1] (slug, scannedAt, window, engineVersion, byEngine k/n); e2[series=uae-glp1]: band, bridgedBand, beforeRef, during, after (k, n, clusterCI95), movementDeclared, precision (afterRateExact, cycleBand, distanceAsDecidedPp, clusterCI95ContainsNearerEdge)
- docs/evidence/computed/easd-2026-final-coding.json → admissions[] (excluded cycle)

**Key reading.** Only Claude's pooled after-window rate (256/378, 67.7%) lies beyond its bridged band, by 0.395 pp; Perplexity stepped from 17/54 on 6 Oct to 40/54 on 7 Oct, yet its pooled 64.3% sits 0.074 pp inside; both clustered intervals span the band edge, so both registered verdicts stand but are borderline.

## Figure 4b. Saudi Arabia naming trajectories (post-registration)

Files: `fig04b-saudi-naming-trajectories.svg`, `fig04b-saudi-naming-trajectories.csv`

**Caption.** Share of scored Saudi Arabia answers that name the brand or its molecule (teplizumab or Tzield; the registered E2 measure, post-registration arm, D10; exploratory) for each engine and for the four engines combined, in every admitted cycle (11 v4 cycles; 8 after-window cycles), from the governed per-cycle counts. Error bars are Wilson 95% intervals computed for this figure (not part of the registered analysis). Grey band: the arm's own v4 single-cycle range (I5). Tinted band: the bridged band against which the after window is judged (§6 offset and its 95% interval, I1). Right column: the after-window pooled rate (K/N shown) with its question-clustered 95% interval (post-registration robustness, D16/D19; never a replacement for the registered rule); filled marker = movement declared, open = inside the band. Units: per cent of scored answers. N: 50–55 scored answers per engine per cycle; after window Perplexity 440, Gemini 440, ChatGPT 440, Claude 437, 1,757 combined. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** 24 Aug – 11 Oct 2026  
**Units.** per cent of scored answers naming the medicine  
**N.** 19 admitted cycles; after window 1,757 scored answers

**Source files and fields.**

- docs/evidence/computed/easd-2026-analysis.json → perCycleNaming[series=saudi-t1d] (slug, scannedAt, window, engineVersion, byEngine k/n); e2[series=saudi-t1d]: band, bridgedBand, beforeRef, during, after (k, n, clusterCI95), movementDeclared, precision (afterRateExact, cycleBand, distanceAsDecidedPp, clusterCI95ContainsNearerEdge)
- docs/evidence/computed/easd-2026-saudi-final-coding.json → admissions[] (excluded cycle)

**Key reading.** No engine's after-window rate lies outside its bridged band; the four-engine rate went from 247/2,402 (10.3%) on v4 to 175/1,757 (10.0%) after the congress.

## Figure 5. E2 band distances at full precision

Files: `fig05-e2-band-distance.svg`, `fig05-e2-band-distance.csv`

**Caption.** Panel A: for each engine and for the four engines combined, the after-window pooled naming rate (dot; filled = movement declared under the registered rule, open = inside the band), the bridged band (shaded), and the question-clustered 95% interval of the after-window rate (dashed; post-registration robustness, D16/D19). Right: the signed distance from the pooled rate to the nearer band edge at full precision, against the band as decided and against the band rebuilt from exact single-cycle rates (D19), the registered verdict, and whether the clustered interval contains that edge (borderline). UAE rows are the registered endpoint; Saudi rows are the post-registration arm (D10), exploratory. Panel B: the three borderline UAE comparisons on a ±1 pp scale around the upper edge. Units: per cent of scored answers; distances in percentage points (pp). N: 378 scored after-window answers per UAE engine (1,512 combined); Saudi 437–440 per engine (1,757 combined); 54 (UAE) or 55 (Saudi) question clusters. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** After window, 4–11 Oct 2026 (7 UAE cycles; 8 Saudi cycles)  
**Units.** per cent; percentage points  
**N.** UAE 378 per engine, 1,512 combined; Saudi 437–440 per engine, 1,757 combined

**Source files and fields.**

- docs/evidence/computed/easd-2026-analysis.json → e2[]: after (k, n, clusterCI95, questions), bridgedBand, movementDeclared, precision (afterRateExact, bridgedAsDecided, distanceAsDecidedPp, distanceExactRegisteredBandPp, bridgedExactCycleBand, distanceExactCycleBandPp, nearerEdge, clusterCI95ContainsNearerEdge, movementUnderExactCycleBand)

**Key reading.** Claude's movement (+0.395 pp beyond the edge; +0.441 pp on the exact-cycle band) and Perplexity's no-movement (−0.074 pp inside; −0.093 pp on the exact-cycle band) are each decided by less than half a percentage point, while their question-clustered intervals extend 10.58 pp below and 10.32 pp above the rate (Claude) and 7.68 pp below and 7.67 pp above (Perplexity), and 6.74 pp below and 7.01 pp above for the four-engine total; both registered verdicts stand, but the data cannot distinguish either from its opposite.

## Figure 6. E1 inspection coverage

Files: `fig06-e1-inspection-coverage.svg`, `fig06-e1-inspection-coverage.csv`

**Caption.** Coverage of the E1 source route (registered primary endpoint; Saudi rows post-registration, D10) by series, engine and window. Each bar is the set of new-URL citations (URLs not cited by the same series before the congress) in that engine's answers: solid = page read and classified; hatched = page unreadable (fetch failure or HTTP ≥ 400), which is unverified and never counted as negative. Text gives the share read, the counts, and the answers citing at least one new or unreadable page. Saudi Arabia has no during-window cycle. Units: citations (URL occurrences in answers) and answers. N: 5,268 new (series, URL) pairs; per-row N as labelled. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** During and after windows, 29 Sep – 11 Oct 2026  
**Units.** citations; answers  
**N.** 5,268 new (series, URL) pairs

**Source files and fields.**

- docs/evidence/computed/easd-2026-e1.json → inspectionCoverage.bySeriesEngineWindow (answers, citingNewUrl, citingUnreadable, newUrlCitations, unreadableCitations, newUrlCitationsRead); screen

**Key reading.** In the after window, between 68.5% (Saudi Arabia Gemini, 3,532 of 5,158) and 90.1% (Saudi Arabia Perplexity, 701 of 778) of new-URL citations could be read in every series and engine; in the congress window the share ranged from 0.0% (Singapore ChatGPT, 0 of 1) to 100.0% (Singapore Claude, 5 of 5); the unread remainder is unverified, not negative.

## Figure 7. E1 penetration and latency

Files: `fig07-e1-latency.svg`, `fig07-e1-latency.csv`

**Caption.** Days from each readout's presentation to the first admitted cycle in which an engine's answer cited a congress-derived source carrying that readout or reported its result (E1, registered primary endpoint). Markers identify engines and sit at the first qualifying cycle; the bar behind each marker is the latency interval (previous admitted cycle, first qualifying cycle], because latency resolution is set by the cycle schedule (post-registration correction, D24 item 11a). Dashed arrows mark right-censored cells, reported as "not observed by day N", where N runs from presentation to the completion of the series' last admitted cycle (UAE 11 Oct, 14:05 UTC; Singapore 11 Oct, 14:02 UTC; Saudi Arabia 11 Oct, 16:34 UTC). Groups: the seven registered readouts in their own series; readouts added after registration (OCTANE, D3; the Saudi type 1 diabetes programme, D10), exploratory; other EASD 2026 data outside the readout set (E1 classes c and d), per series. The bottom row is the cross-arm entry: TRIUMPH-2, registered to the UAE series, carried by the diaTribe round-up that Claude cited in a Saudi answer at day 8.4, unverified at citation (D18/D19), reported apart and never counted. Units: days (one decimal). N: 8 observed and 40 censored engine × readout cells. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** Presentation (28 Sep – 1 Oct 2026) to the after-window cutoff (11 Oct 23:59 UTC)  
**Units.** days from presentation  
**N.** 48 engine × readout cells (8 observed, 40 censored) plus 1 cross-arm entry

**Source files and fields.**

- docs/evidence/computed/easd-2026-e1.json → latency[] (series, engine, readout, presentedUtc, firstCycle, days, censoredAtDay); readouts[] (postRegistration); crossArm[] (answerSeries, registeredSeries, engine, readout, firstCycle, days, contentAtCitation)
- docs/evidence/easd-2026-e1-ledger.json → sources[diatribe].readoutsUnverifiedAtCitation, adjudication
- docs/evidence/computed/easd-2026-e1-manifest.json → entries[prnewswire OCTANE copy] (firstObservedInAnswer, readable)
- docs/evidence/computed/easd-2026-robustness.json → item11E1Corrections.a_latencyIntervals.rows (interval.lowerExclusiveDay, interval.upperInclusiveDay; checked against e1 latency days)

**Key reading.** No registered readout was verified in the inspected citations or text of its own series (not observed by day 10–12.1); congress material arrived only through post-registration readouts and other EASD 2026 data, in answers from Perplexity, Gemini and Claude and never from ChatGPT, each latency known only to within its cycle interval (for example Perplexity, OCTANE: (0, 4.2] days).

## Figure 8. Source-to-claim illustration: the diaTribe case

Files: `fig08-diatribe-source-to-claim.svg`, `fig08-diatribe-source-to-claim.csv`

**Caption.** One qualifying source traced from page to answer. Left: the diaTribe round-up "Top Diabetes News: EASD 2026" (class d media, dated 5 Oct) and its items as read on 11 Oct. Right: the Saudi-arm answer by Claude that cited the exact URL (8 Oct 16:31 UTC; the question is described, its text withheld under protocol §10). The teplizumab (Tzield) real-world item is verified at citation because the answer repeats its wording (quoted verbatim). The retatrutide item, which reports TRIUMPH-2 data without naming the trial, is unverified at citation: the answer does not use it and no archived copy exists; it is reported and never counted (D19), and it belongs to the UAE series, where no answer cited the page. Bottom: dates, and what the case counts toward under E1. Saudi arm: post-registration (D10). Units: dates (UTC); days from presentation. N: one source, one answer.

**Observation period.** 29 Sep – 11 Oct 2026  
**Units.** dates (UTC); days  
**N.** 1 cited page; 1 answer

**Source files and fields.**

- docs/evidence/easd-2026-e1-ledger.json → sources[url=…top-diabetes-news-easd-2026]: url, class, pageDate, fetchedAt, readVia, note, revisions[0].reason, readoutsUnverifiedAtCitation, adjudication, contentAtCitation
- docs/evidence/computed/easd-2026-e1.json → latency[saudi-t1d, anthropic], latency[uae-glp1, anthropic, TRIUMPH-2], crossArm[0], readouts[].presentedUtc

**Key reading.** A single citation carried two congress items with different evidential status: one demonstrably reached the answer's wording, the other cannot be shown to have been on the page when it was cited, so citing a page is not evidence that its content was represented.

## Figure 9. Saudi denominators

Files: `fig09-saudi-denominators.svg`, `fig09-saudi-denominators.csv`

**Caption.** Saudi Arabia, teplizumab (Tzield) (post-registration arm, D10; exploratory throughout). Panel A (unit: answer): every captured type 1 diabetes answer of the admitted cycles, the set E3 searched (before 2,446, after 1,789; the E2 naming population, 2,402 and 1,757, is given for reference only, D21); within that set, the answers to questions inviting new or emerging treatments, identified by one fixed rule on the question text (opportunity denominator, D16: 44 and 32); and answers naming any of the ten EASD 2026 type 1 agents (E3: 0 and 0). Panel B (unit: naming of semaglutide, tirzepatide or sotagliflozin within an answer): namings, the seeded sample (at most 60 per molecule per window), namings coded as a type 1 diabetes use, and those with no statement that the use is unapproved (E3-i). Panel C: the share with no status statement, with its question-clustered 95% interval (post-registration robustness). Before = 11 admitted v4 cycles (24 Aug – 26 Sep); after = 8 cycles (4–11 Oct).

**Observation period.** Before: 24 Aug – 26 Sep 2026 (v4); after: 4–11 Oct 2026  
**Units.** answers (A); namings (B); per cent of type 1 use namings (C)  
**N.** A: 2,446 and 1,789 answers; B: 112 and 206 namings, 56 and 39 type 1 use

**Source files and fields.**

- docs/evidence/computed/easd-2026-analysis.json → saudi.opportunity (answers, opportunityAnswers, governedAnswers), saudi.e3NamingsByWindow (all.namings, all.answers), saudi.e3iExploratory (namings, sampled, distinctAnswersNaming, t1dUseNamings, t1dUseDistinctAnswers, noStatusStatement, share, clusterCI95, questions)

**Key reading.** No answer named any of the ten congress type 1 agents, even among the 32 after-window answers that invited new treatments, and the share of type 1 use namings without an unapproved-use statement moved from 67.9% to 64.1% with heavily overlapping intervals.

## Figure 10. Singapore positioning (descriptive)

Files: `fig10-singapore-positioning.svg`, `fig10-singapore-positioning.csv`

**Caption.** Average rank of empagliflozin (Jardiance) among the medicines named in Singapore answers that name it (by brand or molecule, empagliflozin or Jardiance; 1 = named first; axis inverted), for each admitted cycle (23: 15 v4 cycles, 22 Aug – 26 Sep; 1 during, 2 Oct; 7 after, 5–11 Oct), with the During window shaded, the instrument change marked, and the excluded 27 Sep cycle. Dotted lines give the range of the v4 cycles for reference only. Descriptive (I4): naming is at its ceiling, the §6 bridge measures naming, and no offset exists for rank, so no movement is tested across the instrument change. Units: average rank position. N per cycle: up to four engines × 19 scored answers (the governed population; 75–76 per cycle), and rank is defined over the answers that name the medicine; the per-answer ranks are not in a committed file, so no interval is shown.

**Observation period.** 22 Aug – 11 Oct 2026  
**Units.** average rank position (1 = first)  
**N.** 23 admitted cycles (1 excluded)

**Source files and fields.**

- docs/evidence/computed/easd-2026-analysis.json → singaporePositioning[] (slug, date, window, avgRank); e5 (protocolFirstLine, candidates)
- docs/evidence/computed/easd-2026-final-coding.json → admissions[series=singapore-sglt2] (engineVersion, admitted, reasons)

**Key reading.** Average rank sat between 1.46 and 1.67 in the after window against 1.59–1.80 across v4 cycles, but the after-window cycles ran on instrument v5 and no rank offset exists, so the shift is described, not attributed.

## Figure 11. Claude's rise and the answers citing one local page

Files: `fig11-claude-cited-page.svg`, `fig11-claude-cited-page.csv`

**Caption.** Registered §5 descriptive analysis, reported late (D24); descriptive association, not a causal estimate. Panel A: for Claude (UAE), the share of scored answers naming semaglutide or Wegovy (the registered E2 measure) in each cycle from 27 Sep, split into answers that cite one new page (P61, a local commercial blog post (UAE), August 2026, before the congress) and answers that do not, with K/N and Wilson 95% intervals; the v4 cycles are pooled (336/702, Claude's registered v4 reference), drawn as the dashed reference level; the congress window and the page's first citation (29 Sep) are marked. Panel B: questions that drew the page in the after window against questions that never did, before (v4 and 27 Sep) and after, with the after-window split by citation and the excess over each question's own pre-congress rate. Panel C: the echo test of the page's content at citation. The page is never named. Units: per cent of scored answers; namings. N: 378 after-window scored answers (152 citing the page). Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** 22 Aug – 11 Oct 2026 (per cycle from 27 Sep)  
**Units.** per cent of scored answers; namings; answers  
**N.** 164 citing answers from 29 Sep; 378 after-window scored answers

**Source files and fields.**

- docs/evidence/computed/easd-2026-robustness.json → item12PageP61: page (type, pageDate, firstCitedInSeries), a_citationsByEngine, b_contentAtCitationByEcho (sentences, claudeByCycle, preCongressEchoes, byEngineAndWindow), c_whereClaudesRiseSits (byWindow, byCycle, groups, decomposition)
- docs/evidence/computed/easd-2026-analysis.json → perCycleNaming (reconciliation), e2[uae-glp1, anthropic].beforeRef

**Key reading.** After the congress, Claude named semaglutide or Wegovy in 143/152 (94.1%) of answers citing one pre-congress local page and in 113/226 (50.0%) of the rest, close to its v4 level of 47.9%; 152 of 164 citing answers repeat the page's sentences and 0 non-citing answers do. The rise is concentrated in the answers that cite the page: a descriptive association, not a causal estimate.

## Figure 12. E2 under three readings of the band rule

Files: `fig12-e2-band-readings.svg`, `fig12-e2-band-readings.csv`

**Caption.** Post-registration robustness (D24). Panel A: for each engine, the after-window pooled rate of answers naming the brand or its molecule (registered E2 measure) against the bridged band under three readings of "widened by the offset's 95% interval": (i) the registered reading I1 (D16); (ii) shift by the point offset and widen each end by half the interval width; (iii) shift and widen by the full width. Signed distances in pp; verdicts that differ from the registered one are marked. UAE is the registered endpoint; Saudi Arabia is the post-registration arm (D10). Panel B: the cycle-level change estimate (mean after-window cycle rate minus mean v4 reference cycle rate minus the bridge point offset), with its 95% interval and z; a different estimand from the band rule. Units: per cent; percentage points. N: UAE 378 scored answers per engine after; Saudi 437–440; cycles as listed in the CSV. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** After window, 4–11 Oct 2026, against the v4 reference cycles  
**Units.** per cent; percentage points; z  
**N.** UAE 378 per engine (1,512 combined); Saudi 437–440 per engine (1,757 combined)

**Source files and fields.**

- docs/evidence/computed/easd-2026-robustness.json → item2BandReadings (readings, rows[].after, readings.*.band/movement/signedDistancePp, flipsUnderReadings), item3CycleLevelChange (estimand, rows[].changeNetOfOffsetPp, primary.ci95Pp, primary.z)
- docs/evidence/computed/easd-2026-analysis.json → e2 (reconciliation of reading i)

**Key reading.** Two registered verdicts depend on how the band is read (UAE Perplexity under (ii) half-width (+0.546 pp, movement); UAE Claude under (iii) full width (−8.865 pp, no movement)), while the cycle-level estimate puts Claude's change at +16.8 pp (95% CI +7.2 to +26.4) and Perplexity's at +12.2 pp (−3.0 to +27.4); the registered verdicts stand.

## Figure 13. Source-class mix

Files: `fig13-source-class-mix.svg`, `fig13-source-class-mix.csv`

**Caption.** Registered §5 descriptive analysis, reported late (D24). Panel A: share of each engine's cited links by source class (regulator or government health body; journal, literature or trial registry; congress; sponsor or manufacturer; news or medical media; clinic, hospital or care provider; consumer or retail, including pharmacy; other classified; unclassified), per window, in each series (Saudi Arabia: post-registration arm, D10). Panel B: the class mix in the two cycle pairs where a registered-series engine's naming stepped: Claude 27 Sep against 29 Sep, and Perplexity 6 Oct against 7 Oct, with governed naming for each cycle. Classes are assigned by host; no local business is named. Units: per cent of links; links. Engines, as configured: ChatGPT, gpt-4o-mini, OpenAI Responses API with the web_search tool; the model decides per request whether to search; Claude, claude-haiku-4-5, Messages API with the web_search_20250305 server tool, max_uses 5; the model decides whether to search; Gemini, gemini-2.5-flash with Google Search grounding; grounding is model-decided; Perplexity, sonar, chat completions; retrieval on every request.

**Observation period.** 22 Aug – 11 Oct 2026 (Panel B: 27 and 29 Sep; 6 and 7 Oct)  
**Units.** per cent of cited links; links  
**N.** 151,614 classified links

**Source files and fields.**

- docs/evidence/computed/easd-2026-robustness.json → item1RegisteredSection5ReportedLate.c_sourceClassMix: rows[] (series, window, engine, answers, links, counts), comparisons[] (from/to: slug, date, governedNaming, links, counts), classifierCoverage
- docs/evidence/computed/easd-2026-analysis.json → perCycleNaming (reconciliation)

**Key reading.** Claude's mix barely moved between 27 and 29 Sep (largest class change News / medical media −1.9 pp) although its naming rose, while Perplexity's whole mix changed between 6 and 7 Oct (Other (classified) +40.3 pp) as its naming stepped; class shares cannot separate one source from another, which is why Figure 11 follows the answers citing a single page.

## Inputs and their sha256

| File | sha256 | Status |
|---|---|---|
| `docs/evidence/computed/easd-2026-analysis.json` | `18790c6af0808dddf520d13a9f9aef9ebb193be359eb53e703346bb0ead39432` | matches the freeze record |
| `docs/evidence/computed/easd-2026-e1.json` | `9384d569c49ca5908c8ef442663786d8db74518d35e4af44884d87a049f1de05` | matches the freeze record |
| `docs/evidence/easd-2026-e1-ledger.json` | `07d8f49d5bfab27561d584575a82b1c7a4d8114f6cce1e9b4f0df93ad9f3ac8c` | matches the freeze record |
| `docs/evidence/computed/easd-2026-e1-manifest.json` | `e6403e881f771f48d231aa8035efb9b33bb9dd2406deee7b6682580d699fad5e` | matches the freeze record |
| `docs/evidence/computed/easd-2026-robustness.json` | `5c6806bb5ca0897970381b5946e63c5ba2401776f9ca6a1dea2da99eb2cac124` | matches the freeze record |
| `docs/evidence/computed/easd-2026-final-coding.json` | `4f773dbcf8b337864c13175107727ec3b976dc433238eacb4829292518508bd3` | matches the freeze record |
| `docs/evidence/computed/easd-2026-saudi-final-coding.json` | `8cbaed40819bdb98cae8059124b2402f37af38684081e1cd4bbf438273c5b1d8` | matches the freeze record |
| `docs/evidence/computed/easd-2026-bridge.json` | `8286987d7d0581520acdcd8fc91252fe1cf2261c068a662be7d394a5e89ed146` | matches the freeze record |
| `docs/design/easd-2026-evidence-to-answer-prereg.md` | `204153f68887dc60102fa49f8d0bfb201c7cb0de1d6c1b27f19f3c366c087756` | matches the freeze record |