Signals · checked 22 September 2026
Several lanes move. None is the same verdict.
The watch labels the signal first, then asks what would make it causal.
Claude availability
Resolved
Fable, Mythos, and Opus errors affected Claude surfaces for a short window today. Recovery narrows the action to retry or wait; it does not measure quality.
Surface identity
Chat ≠ Work
OpenAI says the recent Sol update applied to Chat, not the Work/Codex version. Product surface is a first-class comparison field.
Community quality
Narrow lead
Fast or shallow replies, file/tool failures, and Chat-versus-Work reports need a fixed replay, not a universal label.
Measurement process
Judge + age
Judge versions can move scores, and the visible drift board’s latest snapshot is 13 September. Expose judge identity and snapshot age.
Signal table
| Signal | Current read | Next test |
|---|---|---|
| Claude multi-model errors | Declared availability incident, resolved 02:35 UTC. | Check post-recovery fixed tasks by model, surface, account, and region. |
| Chat versus Work reports | Detailed but self-selected system-quality lead. | Replay one task with product surface, served model, route, context, and artifact receipts. |
| Account access limits | Officially acknowledged access lane; mechanism and response route remain opaque. | Record access state, fallback notice, served identity, and task outcome. |
| Drift monitor and judge | Movement and evaluator instability are both plausible signals. | Pin judge/version, use fixed anchors, publish coverage and snapshot age. |