← All incidents

Drift incident · Raven Alpha (behaviour fingerprint)

gpt-4.1

Model page
✓ ResolvedConfidence: Low
Responses to fixed prompts no longer match this model’s own history (672 recent vs 2744 baseline responses). Biggest shifts: first token on “Pick a random integer between 1 and 50. Reply with only the number.”; token probabilities on “Pick a random integer between 1 and 50. Reply with only the number.”; token probabilities on “Flip a coin. Reply with only heads or tails.”
Oct 10change began ≈ 1:42 AM UTC
Oct 11incident opened 9:44 AM UTC
Oct 11resolved 8:00 PM 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 11, 2026, 9:44 AM UTC UTC. Resolved Oct 11, 2026, 8:00 PM UTC UTC: the series returned to its baseline. Methodology.

Share

Permanent link: https://martingale-bench.vercel.app/incidents/inc-gpt-4.1-fingerprint-2026-10-11T09

Incident id inc-gpt-4.1-fingerprint-2026-10-11T09. Every public incident is also in the Atom feed and the JSON feed, and in the raw data. Methodology v0.2.