Datasets:
Public validation question banks
106,324 questions come from91 existing groups:54,402 in training-replay.jsonl.gz overlap public Qwen prompts or belong to training replay;51,922 in observed-regression.jsonl.gz do not overlap those prompts but have been observed during development. They are not independent blind tests or proven disjoint from every unpublished historical consumption record.
Rows retain complete calls, fixed settings, English full prompts, reference labels, accident flags, groups and training overlap. Historical AI labels are not certified human truth. Full-prompt deduplication preserves group membership. Length-boundary questions are included; front-limit refusal is not model risk recognition.
All415,194 private inputs were decoded exactly and excluded by tool-call body across settings. Identifiable local paths/identities, nonreserved emails and credential literals were also withheld:204 of106,528 reference prompts are not public. Retained questions were not altered through redaction.
The483,680 real Codex/Claude command bodies are withheld for private paths, source, credentials and sessions. Only counts (Codex469,219,Claude14,461) and release rates are public. See the real-traffic review. Public validation, training replay and private traffic are distinct denominators.
Public-subset native matrices including front-limit refusals and file hashes are in validation/manifest.json. The original106,501 within-budget core remains separate; the reduced public set is not that original cohort.