The journal shows published findings. Each entry is a change the detector flagged. Pass-rate findings use a two-proportion test with Benjamini-Hochberg FDR control (movement survives at α=0.01, watch is nominally significant, unconfirmed). Serving-metric findings (thinking tokens, latency) use relative-shift thresholds; latency is operational context, never a standalone drift claim. The advisory feed contains an append-only log of non-latency Movements alongside current Watches, latency Context, and coverage status, refreshed each run.

Same model, same battery. Daily (canary) and weekly (deep) readings are different batteries and stay separate. Traces run against each model's own past, never against other models.

Exploratory changes live in the dossier. The seismogram and dossier show every day-to-day difference, previous reading subtracted. Those are for browsing, not findings; comparisons are withheld across missing readings and roster revisions.

Cause is a separate question. Each entry records an observation under its stated rule; Watches remain unconfirmed and latency is operational context. The charter separates measurement from causal attribution; product-harness behaviour above the API is out of scope.

Coverage sits beside pass rates. A run with no reported active dimension is a missing reading, and a 100% pass rate can coexist with failed calls, so each entry shows the call errors on that reading. Quota or billing errors do not establish a provider-wide outage.

Every finding is traceable. Each entry links to the underlying witnessed reading and its Rekor witness bundle when a matching bundle is available, with an explicitly labelled directory fallback. Readings carry the battery hash, roster, runner and grader versions, the code commit, and per-dimension sample and probe counts. Full history: series.json.