# EASD 2026 study: analysis

Generated by `scripts/easd-2026-study/analyze.ts` from a read-only pull (2026-08-22 → 2026-10-11, pulled 2026-10-11T16:56:27.066Z). Registered rules as written; readings fixed in D16 (I1–I5). Rows marked POST-REGISTRATION are robustness or the Saudi arm (D10) and never replace a registered figure.

Scorer validation: 64 admitted cycles rescored; every per-engine count equals the governed count.

## E2: naming movement against the bridged band

| Series | Engine | Before (ref) | During | After | After, question-clustered 95% (post-reg.) | Bridged band | Movement | Magnitude | Beyond band | E1 answers (named) | Others (named) | Fisher p | Verdict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| uae-glp1 | Perplexity | 193/371 (52.0%) | 22/54 (40.7%) | **243/378 (64.3%)** | 56.6%–72.0% | 42.0%–64.4% | no | +12.3pp DECISION-RELEVANT | 0pp | 1 (1) | 377 (242) | 1 | congress sources without movement |
| uae-glp1 | Gemini | 416/701 (59.3%) | 35/54 (64.8%) | **206/378 (54.5%)** | 45.5%–63.8% | 35.1%–78.4% | no | -0.5pp SMALL | 0pp | 1 (1) | 377 (205) | 1 | congress sources without movement |
| uae-glp1 | ChatGPT | 212/702 (30.2%) | 15/54 (27.8%) | **110/378 (29.1%)** | 18.8%–39.1% | 14.2%–46.9% | no | +0.8pp SMALL | 0pp | 0 (0) | 378 (110) | — | neither |
| uae-glp1 | Claude | 336/702 (47.9%) | 31/54 (57.4%) | **256/378 (67.7%)** | 57.1%–78.0% | 34.5%–67.3% | **yes** | +16.8pp MAJOR | 0.4pp | 0 (0) | 378 (256) | — | movement without congress sources |
| uae-glp1 | Four engines | 1273/2794 (45.6%) | 103/216 (47.7%) | **815/1512 (53.9%)** | 47.2%–60.9% | 36.9%–54.0% | no | +9.1pp POTENTIALLY RELEVANT | 0pp | — | — | — | — |
| saudi-t1d (post-reg.) | Perplexity | 58/589 (9.8%) | 0/0 (n/a) | **26/440 (5.9%)** | 2.9%–9.1% | 3.8%–23.6% | no | -4.5pp SMALL | 0pp | 3 (1) | 437 (25) | 0.1674 | congress sources without movement |
| saudi-t1d (post-reg.) | Gemini | 54/603 (9.0%) | 0/0 (n/a) | **40/440 (9.1%)** | 4.3%–14.5% | 3.0%–30.3% | no | -4.7pp SMALL | 0pp | 1 (0) | 439 (40) | 1 | congress sources without movement |
| saudi-t1d (post-reg.) | ChatGPT | 105/605 (17.4%) | 0/0 (n/a) | **81/440 (18.4%)** | 9.8%–27.7% | 6.7%–35.1% | no | +1.1pp SMALL | 0pp | 0 (0) | 440 (81) | — | neither |
| saudi-t1d (post-reg.) | Claude | 30/605 (5.0%) | 0/0 (n/a) | **28/437 (6.4%)** | 3.2%–10.3% | -1.2%–13.3% | no | +0.8pp SMALL | 0pp | 0 (0) | 437 (28) | — | neither |
| saudi-t1d (post-reg.) | Four engines | 247/2402 (10.3%) | 0/0 (n/a) | **175/1757 (10.0%)** | 6.4%–13.9% | 6.5%–23.3% | no | -1.8pp SMALL | 0pp | — | — | — | — |

### E2 at full precision (POST-REGISTRATION check, D19; the registered verdict above stands)

The decision uses the registered band as written (§4 quotes the UAE limits to 0.1%) and the offset bounds as stored. Here the after-window rate is unrounded, the offset bounds are restored to their exact values (multiples of 1/(paired questions × runs)), and the band is also rebuilt from the exact single-cycle rates. Distance is signed: positive means beyond the nearer edge, negative means inside it by that much. The last column says whether the question-clustered 95% interval of the after-window rate contains that edge, which is when a verdict should be read as borderline.

| Series | Engine | After (K/N, exact) | Distance, registered band (pp) | Distance, exact cycle band (pp) | Registered verdict | Same verdict on exact cycle band | Clustered 95% contains the nearer edge |
|---|---|---|---|---|---|---|---|
| uae-glp1 | Perplexity | 243/378 = 64.286% | -0.074 | -0.093 | no movement | yes | **yes (borderline)** |
| uae-glp1 | Gemini | 206/378 = 54.497% | -19.357 | -19.312 | no movement | yes | no |
| uae-glp1 | ChatGPT | 110/378 = 29.101% | -14.870 | -14.903 | no movement | yes | no |
| uae-glp1 | Claude | 256/378 = 67.725% | +0.395 | +0.441 | movement | yes | **yes (borderline)** |
| uae-glp1 | Four engines | 815/1512 = 53.902% | -0.078 | -0.049 | no movement | yes | **yes (borderline)** |
| saudi-t1d (post-reg.) | Perplexity | 26/440 = 5.909% | -2.135 | -2.135 | no movement | yes | **yes (borderline)** |
| saudi-t1d (post-reg.) | Gemini | 40/440 = 9.091% | -6.064 | -6.061 | no movement | yes | no |
| saudi-t1d (post-reg.) | ChatGPT | 81/440 = 18.409% | -11.740 | -11.742 | no movement | yes | no |
| saudi-t1d (post-reg.) | Claude | 28/437 = 6.407% | -6.922 | -6.926 | no movement | yes | no |
| saudi-t1d (post-reg.) | Four engines | 175/1757 = 9.960% | -3.466 | -3.470 | no movement | yes | **yes (borderline)** |

### Bridge self-check (registered, §6)

§6 registered that the v4-setting run-to-run spread "should reproduce the v4 band's width, a check on the bridge itself", and that the v5 spread is reported beside it. No threshold was registered, so the comparison is shown, not scored. The bridge ran three same-day runs per setting; the v4 band is the single-cycle range over the pre-congress weeks. A same-day spread much narrower than the band means most of the variation the band captures arises between days, not between repeated runs on one day.

| Series | Engine | v4 runs (K/N) | v4 same-day spread | v5 runs (K/N) | v5 same-day spread | Reference band (UAE: registered §4; Saudi: I5; Singapore: v4 range, no registered band) | Band width | Spread / width | Reproduces width |
|---|---|---|---|---|---|---|---|---|---|
| uae-glp1 | Perplexity | 19/54, 16/54, 16/54 | 29.6%–35.2% (5.56pp) | 18/54, 15/54, 18/54 | 27.8%–33.3% (5.56pp) | 46.3%–58.8% | 12.5pp | 0.445 | no |
| uae-glp1 | Gemini | 31/54, 32/54, 30/54 | 55.6%–59.3% (3.7pp) | 31/54, 26/54, 29/54 | 48.1%–57.4% (9.26pp) | 48.1%–74.1% | 26pp | 0.142 | no |
| uae-glp1 | ChatGPT | 17/54, 16/54, 15/54 | 27.8%–31.5% (3.7pp) | 16/54, 15/54, 14/54 | 25.9%–29.6% (3.7pp) | 20.4%–44.4% | 24pp | 0.154 | no |
| uae-glp1 | Claude | 37/54, 36/54, 33/54 | 61.1%–68.5% (7.41pp) | 36/54, 38/54, 37/54 | 66.7%–70.4% (3.7pp) | 40.7%–55.6% | 14.9pp | 0.497 | no |
| singapore-sglt2 | Perplexity | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 100.0%–100.0% | 0pp | n/a | n/a (ceiling) |
| singapore-sglt2 | Gemini | 19/19, 18/19, 19/19 | 94.7%–100.0% (5.26pp) | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 94.7%–100.0% | 5.26pp | 1 | n/a (ceiling) |
| singapore-sglt2 | ChatGPT | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 100.0%–100.0% | 0pp | n/a | n/a (ceiling) |
| singapore-sglt2 | Claude | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 19/19, 19/19, 19/19 | 100.0%–100.0% (0pp) | 94.7%–100.0% | 5.26pp | 0 | n/a (ceiling) |
| saudi-t1d | Perplexity | 0/55, 1/55, 0/55 | 0.0%–1.8% (1.82pp) | 0/55, 1/55, 1/55 | 0.0%–1.8% (1.82pp) | 3.8%–21.8% | 18.04pp | 0.101 | no |
| saudi-t1d | Gemini | 5/55, 3/55, 3/55 | 5.5%–9.1% (3.64pp) | 9/55, 2/54, 8/55 | 3.7%–16.4% (12.66pp) | 3.6%–20.0% | 16.36pp | 0.222 | no |
| saudi-t1d | ChatGPT | 11/55, 9/55, 8/55 | 14.5%–20.0% (5.45pp) | 10/55, 7/55, 11/55 | 12.7%–20.0% (7.27pp) | 10.9%–30.9% | 20pp | 0.273 | no |
| saudi-t1d | Claude | 0/55, 1/55, 2/55 | 0.0%–3.6% (3.64pp) | 1/55, 2/55, 1/55 | 1.8%–3.6% (1.82pp) | 0.0%–10.9% | 10.91pp | 0.334 | no |

## Singapore E2: positioning (descriptive, I4)

| Cycle | Date | Window | Average rank | Top competitor mention rates |
|---|---|---|---|---|
| `SG-C01` | 2026-08-22 | before-v4 | 1.72 | Invokana 98.7%, Farxiga 98.7%, Forxiga 90.7% |
| `SG-C02` | 2026-08-24 | before-v4 | 1.74 | Farxiga 100.0%, Invokana 98.7%, Forxiga 90.8% |
| `SG-C03` | 2026-08-26 | before-v4 | 1.7 | Invokana 100.0%, Farxiga 100.0%, Forxiga 92.1% |
| `SG-C04` | 2026-08-28 | before-v4 | 1.79 | Farxiga 100.0%, Invokana 97.4%, Forxiga 90.8% |
| `SG-C05` | 2026-08-30 | before-v4 | 1.72 | Farxiga 98.7%, Invokana 97.4%, Forxiga 85.5% |
| `SG-C06` | 2026-09-04 | before-v4 | 1.72 | Invokana 100.0%, Farxiga 100.0%, Forxiga 80.3% |
| `SG-C07` | 2026-09-06 | before-v4 | 1.64 | Invokana 100.0%, Farxiga 100.0%, Forxiga 89.5% |
| `SG-C08` | 2026-09-08 | before-v4 | 1.63 | Invokana 97.4%, Farxiga 97.4%, Forxiga 90.8% |
| `SG-C09` | 2026-09-10 | before-v4 | 1.64 | Invokana 100.0%, Farxiga 100.0%, Forxiga 88.2% |
| `SG-C10` | 2026-09-13 | before-v4 | 1.69 | Farxiga 98.7%, Invokana 97.4%, Forxiga 85.5% |
| `SG-C11` | 2026-09-15 | before-v4 | 1.8 | Invokana 100.0%, Farxiga 100.0%, Forxiga 85.5% |
| `SG-C12` | 2026-09-17 | before-v4 | 1.61 | Invokana 100.0%, Farxiga 100.0%, Forxiga 80.3% |
| `SG-C13` | 2026-09-19 | before-v4 | 1.79 | Invokana 100.0%, Farxiga 100.0%, Forxiga 89.3% |
| `SG-C14` | 2026-09-21 | before-v4 | 1.59 | Farxiga 98.7%, Invokana 97.3%, Forxiga 82.7% |
| `SG-C15` | 2026-09-26 | before-v4 | 1.64 | Invokana 100.0%, Farxiga 100.0%, Forxiga 86.8% |
| `SG-C17` | 2026-10-02 | during | 1.68 | Invokana 100.0%, Farxiga 100.0%, Forxiga 88.2% |
| `SG-C18` | 2026-10-05 | after | 1.54 | Invokana 100.0%, Farxiga 100.0%, Forxiga 90.8% |
| `SG-C19` | 2026-10-06 | after | 1.46 | Invokana 100.0%, Farxiga 100.0%, Forxiga 81.6% |
| `SG-C20` | 2026-10-07 | after | 1.67 | Invokana 100.0%, Farxiga 100.0%, Forxiga 86.8% |
| `SG-C21` | 2026-10-08 | after | 1.67 | Invokana 100.0%, Farxiga 100.0%, Forxiga 84.2% |
| `SG-C22` | 2026-10-09 | after | 1.55 | Invokana 100.0%, Farxiga 100.0%, Forxiga 82.9% |
| `SG-C23` | 2026-10-10 | after | 1.58 | Invokana 100.0%, Farxiga 100.0%, Forxiga 88.2% |
| `SG-C24` | 2026-10-11 | after | 1.57 | Invokana 100.0%, Farxiga 100.0%, Forxiga 80.3% |

## E3 (UAE): investigational agents named, by window

| Window | Perplexity | Gemini | ChatGPT | Claude | All | Status correct / absent / wrong | Clinically material | Material, strict (post-reg.) |
|---|---|---|---|---|---|---|---|---|
| before-v4 | 1/765 | 3/779 | 4/780 | 2/780 | 10/3104 | 9 / 1 / 0 | 1 | 0 |
| before-v5 | 0/58 | 0/60 | 1/60 | 0/60 | 1/238 | 1 / 0 / 0 | 0 | 0 |
| during | 0/60 | 0/60 | 0/60 | 0/60 | 0/240 | 0 / 0 / 0 | 0 | 0 |
| after | 5/420 | 1/420 | 0/420 | 4/420 | 10/1680 | 8 / 2 / 0 | 2 | 0 |

## E5 (Singapore): first-line framing, by window

| Window | Answers | Candidates | General | High-risk | With metformin (reported, not counted) | Not first-line | Protocol first-line | Strict (post-reg.) |
|---|---|---|---|---|---|---|---|---|
| before-v4 | 1554 | 202 | 0 | 1 | 48 | 153 | **1** | 0 |
| before-v5 | 0 | 0 | 0 | 0 | 0 | 0 | **0** | 0 |
| during | 104 | 9 | 0 | 0 | 4 | 5 | **0** | 0 |
| after | 728 | 53 | 0 | 1 | 11 | 41 | **1** | 0 |

## E4 (UAE): efficacy figures by source category, by window — QUARANTINED (D17)

> **Not for publication as coded.** The "no identifiable source" and "congress-window" columns are hand-checked (D4). The "label" and "published" columns are the script's attribution, which failed its D17 audit (8.0% of a random sample wrongly attributed, Wilson 95% 4.6–13.5%; `docs/evidence/easd-2026-e4-label-audit.json`). Shown for the record only.

| Window | Answers | With a coded figure | Label | Published | No identifiable source | Congress-window |
|---|---|---|---|---|---|---|
| before-v4 | 3104 | 387 | 325 | 0 | 62 | 0 |
| before-v5 | 238 | 30 | 28 | 0 | 2 | 0 |
| during | 240 | 29 | 26 | 0 | 3 | 0 |
| after | 1680 | 266 | 244 | 0 | 20 | 2 |

## Saudi Arabia (POST-REGISTRATION, D10)

E3: answers naming any of the ten EASD 2026 type 1 diabetes agents, by window, against all answers and against answers to questions that invite new or emerging treatments (opportunity rule `\b(new|newest|latest|emerging|upcoming|future|pipeline|research|trials?|investigational|experimental|cure|breakthrough|innovative|advances?|development)\b` on the question text).

Denominators: the namings were searched over every captured answer of the admitted cycles (focused questions included), so "Answers searched" and "Opportunity answers" are counted over that same set (D21). The governed naming population used by E2 is shown for reference only.

| Window | Namings | Answers searched | Opportunity answers (within those searched) | Governed E2 population (reference) |
|---|---|---|---|---|
| before-v4 | 0 | 2446 | 44 | 2402 |
| before-v5 | 0 | 0 | 0 | 0 |
| during | 0 | 0 | 0 | 0 |
| after | 0 | 1789 | 32 | 1757 |

E3-i (EXPLORATORY): approved molecules offered for type 1 diabetes. Denominator: sampled namings coded as a type 1 diabetes use (stated, absent or wrong); namings that only rule the molecule out are excluded.

| Window | Namings (sampled) | T1D-use namings (distinct answers) | No status statement | Share | Question-clustered 95% | Strict share | In which-medicine answers |
|---|---|---|---|---|---|---|---|
| before-v4 | 112 (91) | 56 (45) | 38 | 67.9% | 44.4%–87.3% | 67.9% | 8 |
| before-v5 | 0 (0) | 0 (0) | 0 | n/a | n/a | n/a | 0 |
| during | 0 (0) | 0 (0) | 0 | n/a | n/a | n/a | 0 |
| after | 206 (120) | 39 (31) | 25 | 64.1% | 40.6%–82.0% | 64.1% | 8 |