PCF15: Ministral Publication Confirmation Successor

Status: prospectively frozen on 2026-08-12 before admission or scientific compute.

PCF14 is terminal-null. It completed preparation, mechanics, B1, all 16 draft shards, the exact draft merge, materialization, and revision training. During revision calibration, shard 2 stopped on MBPP identity 8d17a4...eb7f7 because pinned Python failed before trusted READY. All peer allocations and descendants were cancelled. The prepared confirmation assessor board exists as required but has zero semantic reads; no confirmation candidate, score authorization, score, final comparison, or formal scientific result exists. Preserve SHOHIN_PCF14_TERMINAL_INFRASTRUCTURE_RECEIPT_20260812.json and its clarifying amendment. Do not replay PCF14.

PCF15 is a separately named infrastructure successor under the user's standing authorization. Its only change removes Python preexec_fn from the per-candidate launcher. Python explicitly documents that callback as unsafe in the presence of threads because the child can deadlock before exec; PCF14's failure occurred at exactly that boundary in a resident PyTorch/CUDA process. PCF15 instead executes exact host binary /usr/bin/prlimit (SHA-256 2c1c7948...41d5, util-linux 2.32.1), which applies the unchanged CPU 3:4 seconds, address-space 1 GiB:1 GiB, and file-size 1 MiB:1 MiB limits and then execs the unchanged Bubblewrap command. No Python code runs between fork and exec.

The bootstrap SHA-256 remains 726af7a5...a60, the generated-candidate policy remains f27124db...f1e, and the three-second scientific candidate limit is unchanged. Fresh Newton job 752701 passed all 41 sandbox probes. H100 job 752702 loaded the exact dense host and PCF14 revision adapter on the same node, then completed 2,000 consecutive executions of the exact failed MBPP assessor context with zero infrastructure failures and no model generation or confirmation access. Bind admission to SHOHIN_PCF15_SANDBOX_LAUNCH_QUALIFICATION_20260812.json.

PCF15 must start from a fresh root and replay the complete graph. It may not reuse any PCF14 draft, adapter, calibration, or partial output. The pinned dense host and tokenizer, source banks, split and identity order, prompts, unchanged/revision/self-refinement/commit arms, seeds, 256-update training geometry, generation settings, candidate policy, sandbox filesystem and network isolation, three-second candidate limit, custody rules, sealed board, and sole 1,289-row conjunctive gate remain byte-for-byte unchanged. Requeue, retries, partial continuation, automatic successors, and protected-split access are disabled.

The sole falsifiable gate remains:

• unchanged reaches at least 387/1289 with every domain nonzero;
• trained revision exceeds unchanged by at least 65 correct and self-refinement by at least 39, with no per-domain loss against either;
• learned whole-trajectory commit exceeds revision by at least 13, retains at least 95% of both the revision-correct and unchanged-correct identity sets, and loses no domain against revision;
• all four arms cover the exact 1,289 identities in frozen order, with zero truncation, zero malformed selections, exact runtime/model/data/checkpoint/ scheduler custody, one consumed assessor authorization, and zero holdout/public/product access.

Any infrastructure failure, formal PASS, or formal FAIL ends PCF15. A formal result must be preserved exactly and authorizes no successor.