MartingaleBench
← All incidents

Drift incident · Raven Alpha (behaviour fingerprint)

gpt-6-luna

Model page
✓ ResolvedConfidence: Low
Responses to fixed prompts no longer match this model’s own history (1344 recent vs 13278 baseline responses). Biggest shifts: whether it reasoned at all on “Write one sentence about the sea.”; visible output length on “Name one programming language. Reply with only its name.”; whether it reasoned at all on “Name one programming language. Reply with only its name.”
Oct 19change began ≈ 7:56 PM UTC
Oct 20incident opened 10:50 PM UTC
Oct 21resolved 8:47 AM UTC
8.0e-4adjusted 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 20, 2026, 10:50 PM UTC UTC. Resolved Oct 21, 2026, 8:47 AM UTC UTC: the series returned to its baseline. Methodology.

Share

Permanent link: https://martingale-bench.vercel.app/incidents/inc-gpt-6-luna-fingerprint-2026-10-20T22

Incident id inc-gpt-6-luna-fingerprint-2026-10-20T22. Every public incident is also in the Atom feed and the JSON feed, and in the raw data. Methodology v0.2.