| # 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 They are not independent blind tests or proven disjoint from every unpublished historical consumption record.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. | |