← 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 6414 baseline responses). Biggest shifts: output length on “Give me a one-line motivational quote.”; visible output length on “Give me a one-line motivational quote.”; wording on “Write one sentence about rain.”
Oct 18change began ≈ 2:11 PM UTC
Oct 19incident opened 10:14 PM UTC
Oct 20resolved 9:23 AM UTC
2.0e-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 19, 2026, 10:14 PM UTC UTC. Resolved Oct 20, 2026, 9:23 AM UTC UTC: the series returned to its baseline. Methodology.

Share

Permanent link: https://martingale-bench.vercel.app/incidents/inc-gpt-6-sol-fingerprint-2026-10-19T22

Incident id inc-gpt-6-sol-fingerprint-2026-10-19T22. Every public incident is also in the Atom feed and the JSON feed, and in the raw data. Methodology v0.2.