Before you read on, three things this is not. It is not an accusation against Anthropic. It does not say anything about every Claude deployment — only about the route we watched. And it does not replace the incident page itself, which is the record we cannot edit.
Timeline
Timeline and published measurements
- Frozen null floor
- 0.700
- α
- 0.050
- Alarm at
- 20
Incident cd6308af · opened 2026-08-05 17:39:04 UTC · published values
| Field | Published value |
|---|---|
| Model / route id | Anthropic Claude Haiku 4.5 anthropic/claude-haiku-4.5 |
| Task family | Code reasoning |
| Status at open | Open |
| Opened (UTC) | 2026-08-05 17:39:04 |
| Frozen null floor | 0.700 (α = 0.050) |
| Estimated effect size | 0.700 |
| 95% anytime-valid uncertainty interval | [-0.161, 0.700] |
| E-value at open | 20.829 |
| log10 E at open | 1.319 |
| Onset estimate | 5 (diagnostic only) |
| Serving observation | Via aggregator (not direct adapter) |
| Pool / checker | code_reasoning-v4 / code_reasoning-v2 |
| Incident epoch | 5dad03b2-09b2-49a3-bdf7-ccfaf3543703 |
| Epoch transition | checker_refresh from a4d45cef-6c8c-4239-8d1f-35fc2033c630 |
Validity attaches to the e-process alarm. Default α = 0.05 → threshold 20; opened E = 20.829 sits just over that latch.
Route caveat
Aggregator and route caveat
Watchtower records a route: model · upstream · quantisation. Served path via openrouter; upstream/quantisation recorded but not yet pinned. Alarm describes the route as measured.
Observed via aggregator: through an aggregator serving layer, not a direct adapter.
Reference rows never open public incidents by design.
Sibling
Sibling context on the same UTC day
A separate math incident on the same model opened earlier that day, 2026-08-05 09:24:26 UTC (effect 0.500, E = 24.823) — that entry. Do not collapse the two into one “model is broken” story.
As of the 2026-08-11 snapshot: 2 of 5 families in alarm, 770 complete panels — re-check the live model page before publishing status.
Code reasoning
What “code reasoning” means on this instrument
Deterministic checkers on frozen five-member panels; samples ask for exact stdout of fixed Rust programs. Terminal failure scores 0.
Measured
What was measured; what is not claimed
- Effect size + interval as published
- E-value trajectory
- Quiet elsewhere matters (other families)
- Independence/FAR — methodology; 56.38% block-5 dependence stress
- No customer logos — Actuary’s own public monitor
Why it matters
Why this matters if you do not serve Haiku
CI can stay green while a served aggregator route shifts on one family. Frozen floor + α yields a citable incident. Publishing quiet families avoids selective reporting.
Next step
Next step
Open incident: cd6308af-bd8f-4ecb-b554-e148a43f7020.