MartingaleBench
← All incidents

Drift incident · Raven Alpha (behaviour fingerprint)

gpt-6-sol

Model page
✓ ResolvedConfidence: Low
Responses to fixed prompts no longer match this model’s own history (672 recent vs 6076 baseline responses). Biggest shifts: reasoning tokens on “Write a haiku about autumn.”; output length on “Write a haiku about autumn.”; whether it reasoned at all on “Write a haiku about autumn.”
Oct 23change began ≈ 1:00 AM UTC
Oct 24incident opened 7:13 PM UTC
Oct 26resolved 6:13 AM UTC
3.3e-3adjusted p

What this means

Raven Alpha compares how the model answers a fixed set of open-ended prompts with its own history. A persistent, significant shift means the model, its configuration or its serving path changed, even when every scored test still passes.

Opened Oct 24, 2026, 7:13 PM UTC UTC. Resolved Oct 26, 2026, 6:13 AM UTC UTC: the series returned to its baseline. Methodology.

Share

Permanent link: https://martingale-bench.vercel.app/incidents/inc-gpt-6-sol-fingerprint-2026-10-24T19

Incident id inc-gpt-6-sol-fingerprint-2026-10-24T19. Every public incident is also in the Atom feed and the JSON feed, and in the raw data. Methodology v0.2.