gpt-6-luna: thinking-token use fell on the structured output probe set
Movement on the daily battery. This model measured against its own witnessed baseline.
| model | gpt-6-luna |
|---|---|
| identity | alias · a floating name the provider may re-point; a change here means what is served under this name changed |
| dimension | Structured output |
| metric | thinking_tokens_mean |
| decision rule | Relative shift past the serving-tripwire threshold |
| first flagged | 2026-10-02 · still flagged 2026-10-04 (2 readings) |
| coverage | 0 call errors on this reading |
Pattern: thinking-token use moved while the pass rate on the same probes stayed flat (1.00 before, 1.00 at onset). That is the signature of a serving-side change - an effort default or a routing change - rather than a different model behind the name, which usually moves accuracy too. Stated as a pattern, not a cause.
Dated records alongside this reading
An annotation attaches a dated external record (a provider changelog entry, a documented deprecation, a known configuration artefact) to a reading the instrument already produced. It changes no reading and establishes no cause; the instrument detects, attribution stays a human judgment.
Predates the thinking-token floor · 2026-10-02
This entry predates the detector's absolute floor for the thinking-token tripwire (PR #67, merged 2026-10-05): a baseline under 20 mean thinking tokens no longer fires the relative-shift rule, because a model running its minimum-thinking channel emits single-digit thinking tokens and a swing of a few tokens reads as a large multiple. This entry's baseline is below that floor and would not publish under the current rule. The log is append-only, so it stays; read it as a rule-calibration artefact, not a serving change. The pattern note above was computed on the same near-zero numbers and should be read the same way.
Trace the evidence
The witnessed reading for this date and battery, and the Rekor witness bundle. The full series is in series.json.
This entry records an observation under the stated decision rule. Watches remain unconfirmed; latency is operational context. The instrument does not establish the cause of a change.