PCF9: Ministral Publication Confirmation Successor

Status: closed terminal-null on 2026-08-11.

Terminal update: preparation completed. Mechanics had one visible H100 and completed its ephemeral update, then rejected its valid seven-node compute-host receipt because a duplicated Python validator still required the former six-node exclusion list. This occurred before mechanics generation, B1, candidates, assessor read, or scoring. Preserve SHOHIN_PCF9_TERMINAL_INFRASTRUCTURE_RECEIPT_20260811.json.

PCF8 is closed with formal scientific result null. Preparation, no-score mechanics, and the frozen B1 training completed. Draft task 751823_3 received one Slurm H100 GRES on evc33, but its first nvidia-smi invocation reported No devices were found and exited 6 before entering the generator. No draft file or partial was published. The other 15 tasks and every downstream stage were canceled once success was impossible. Confirmation assessors were never semantically read and no score was emitted. Preserve SHOHIN_PCF8_TERMINAL_INFRASTRUCTURE_RECEIPT_20260811.json.

PCF9 is the separately named infrastructure successor under the user's standing authorization. Its sole change is to add proven-broken node evc33 to the existing scheduler exclusion set. The pinned host family, model, tokenizer behavior, source bytes, split, prompts, arms, seeds, training geometry, candidate policy, sandbox, thresholds, assessor custody, and final gate are unchanged. PCF9 starts from a fresh run root and executes the entire graph; it does not reuse the PCF8 B1 checkpoint or any partial output.

The successful H100 allocations in PCF8 established device visibility on the remaining nodes used by the draft array, while the PCF8 mechanics and scratch cleanup passed exactly. Replay the full live preflight, require the queue empty and at least 128 GiB/150,000 inodes of durable headroom, disable requeue and retry, and submit exactly one PCF9 graph. Any infrastructure failure, formal PASS, or formal FAIL ends PCF9. No automatic successor is authorized.