30 of 300 differed over 3 sets (fails from 16) · 26 of 299 differed over 3 sets (fails from 16) · the real model differs from itself on 1% · too close to call on the first set, so the auditors asked more questions
Where it missed
Medical (drug half-lives)7/19
Programming (release years)5/22
Internet culture (dates and numbers)3/21
Biology (chromosome numbers)3/21
For example
Vutrisiran median terminal elimination half-life (hours)real 5.2 · provider 120
NGC 5189 central binary orbital period reported by Manick et al. (2015) (days)real 4.04 · provider 0.1683
龙厦铁路象山隧道长度(米)real 15,898 · provider 8,090
✓Reasoning switched off with reasoning effort none, as the real model was asked.
audited 18 Nov, 18:12 UTC · epoch 31
test power 98% · median answer 9.7 s · 2 audits
Audit rounds
Round
Passed
Failed
Not tested
Multiplier
17 Novepoch 31
3
1
1
×1.04
Audited by approved auditors · evidence on IPFS, hash-checked on-chain.