← All incidents

Drift incident · Raven Alpha (behaviour fingerprint)

gpt-5.4

Model page
✓ ResolvedConfidence: Low
Responses to fixed prompts no longer match this model’s own history (672 recent vs 448 baseline responses). Biggest shifts: first token on “List three uses for a brick.”; token probabilities on “Pick a number between 1 and 1000. Reply with only the number.”; token probabilities on “Suggest a name for a pet cat. Reply with only the name.”
Oct 5change began ≈ 8:22 PM UTC
Oct 7incident opened 2:04 AM UTC
Oct 24resolved 11:30 AM UTC
2.4e-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 7, 2026, 2:04 AM UTC UTC. Resolved Oct 24, 2026, 11:30 AM UTC UTC: the series returned to its baseline. Methodology.

Share

Permanent link: https://martingale-bench.vercel.app/incidents/inc-gpt-5.4-fingerprint-2026-10-07T02

Incident id inc-gpt-5.4-fingerprint-2026-10-07T02. Every public incident is also in the Atom feed and the JSON feed, and in the raw data. Methodology v0.2.