# AI vs database: obesity drug landscape (run 2026-11-06)

Question: "Every obesity drug program in Phase 2 or later, with company, phase and next catalyst."

## Contestants
- **A. AI from memory** (Claude, no search), written before looking at any data: `A_memory_only.md` — 40 programs.
- **B. AI with web search** (Claude, 12 searches/fetches, read 4 pipeline-tracker sites): `B_web_search.md` — 66 rows.
- **Ground truth**: ClinicalTrials.gov API, active industry Phase 2/3 drug trials for obesity, cleaned by hand: `C_ground_truth.md`.
  231 trials -> 181 raw sponsor/drug names -> 94 distinct programs after merging aliases
  (97 in the file, minus 3 duplicates that B's sources revealed: zenagamtide = amycretin, PF-08653944 = MET-097i, LY3841136 = eloralintide).

## Score against the 94 programs in active trials
| | Found | Recall | Missed |
|---|---|---|---|
| A. Memory only | 28 | 30% | 66 |
| B. Web search | 41 (46 counting 5 Phase 1/2 programs it saw but excluded) | 44% (49%) | 48–53 |

## What B missed (examples)
- China / Asia: zovaglutide, ZT006 (QL Bio), BGM1812, SYH2069 (CSPC), DA-302168S, HRS-4729, HDM1005, HDM1702, MWN109, SRSD384, THDBH120, VCT220, HM15275 (Hanmi), IN-B00009 (inno.N), MDR-001, KDS2010, GL0034 (Sun Pharma), MKP10241 (Mankind)
- Small biotechs: AT673, ALV-100, NA-931, BIO101, KBP-336, CSTI-500, ERX1000, ECC4703, GUB-UCN2, RN3161, TE-8105, S1B-509, LUM-201, BMB-101, EMP16
- Big pharma early programs: BI 3034701, Novo NNC0519-0130 and NNC0662-0419, Lilly LY3457263 / LY3532226 / LY3549492, Roche RO7204239, Regeneron mibavademab and ALN-CIDEB, Alnylam ALN-4324

## Quality problems in B's answer
- 13 of 66 rows came from the model's background knowledge, unconfirmed (flagged by the model itself).
- About 28 of 66 rows had a concrete next catalyst; most of the rest say "—" or "not found".
- Sources conflicted on phase for several drugs (petrelintide, lenbelintide, orforglipron T2D).

## What B found that the single database query missed
Survodutide, petrelintide, enicepatide (CT-388), CT-996, emugrobart, pemvidutide, VDB-101, nimacimab, apitegromab, ARD-101,
TG103, supaglutide, PB-119, HEC88473, plus approved older drugs. Their recent trials are finished or registered under other
conditions, so one ClinicalTrials.gov filter drops them. Lesson for the product: union completed trials, news and filings, not just one query.

## Name matching is the hidden work
Mazdutide appeared under 3 names and 2 sponsors; bofanglutide under 3 names and 2 sponsors; amycretin as a code and an INN;
aleniglipron under a subsidiary ("Gasherbrum Bio"). Neither AI nor the raw database does this merging for you.

## Caveats
One disease area, one run each. ChatGPT not tested (Claude stands in). Ground-truth merges and the obesity/non-obesity cut are
hand judgements; 4 merges are marked "confirm". Phase 1/2 trials are counted as Phase 2.