Home › Evidence › Records › clm-0049

clm-0049

measured-herehigh ●●●
citable URL: https://halobench.com/records/clm-0049/ — this address never moves; the anchor /records/#clm-0049 keeps resolving

The 122B has a reproducible rule-precedence bug in policy application: on tau2 airline task 9 it cancels a partially-flown reservation in 7 of 8 trials, every failure with the identical signature — cancel_reservation called without ever checking flight status. It applies the permissive business-class exception without gating on the hard already-flown precondition. Under self-play the task passed consistently — the twin customer steered around the trap — so the defect was invisible until an independent simulator was pinned.

verified 2026-08-11 · volatility medium
evidence run-0111

Note — the record's own working

Stage G: --num-trials 8, seed 42, pinned Haiku simulator, temperature-0 serving; 7/8 failures, 7/7 signature matches, detection logic validated against the original F2 failure transcript first. The passing trial is the outlier (simulator-side sampling variance). The error class — a permissive exception short-circuiting a hard precondition — generalises beyond airline policy and joins MTP output-verification on the production-restore checklist: any agentic deployment of this model wants a hard-precondition guard pattern in its policy prompts or tooling. Candidate for a targeted guard probe in the standard screen.

Cited by — computed at build time, never stored

model pages qwen35-122b