PCF14: Ministral Publication Confirmation Successor

Status: terminal-null on 2026-08-12; no formal scientific gate was reached.

PCF14 completed preparation, mechanics, B1, all 16 draft shards, the exact draft merge, materialization, and revision training. Revision-calibration shard 2 then stopped on MBPP identity 8d17a4...eb7f7 because the isolated pinned Python process failed before trusted READY. All peer allocations and descendants were cancelled. No complete calibration shard, commit training, confirmation candidate, assessor semantic read, score, or final comparison exists. Preserve the immutable receipt SHOHIN_PCF14_TERMINAL_INFRASTRUCTURE_RECEIPT_20260812.json; do not replay this run or report it as a scientific FAIL.

PCF13 is closed with formal scientific result null. Preparation, qualified mechanics, B1, all 16 source-only draft shards, the exact draft merge, materialization, 256-update revision training, all four revision-calibration shards, and their merge completed with zero restarts. Unchanged-calibration shard 1 then reached MBPP identity 51405a...646f7; its isolated process exceeded the five-second outer watchdog after 416 of 1,456 local rows. The sandbox correctly refused to count that ambiguous termination as a wrong answer. The other three unchanged shards and all descendants were cancelled. No complete unchanged-calibration shard, commit checkpoint, confirmation candidate, assessor semantic read, score, or formal gate result exists. Preserve SHOHIN_PCF13_TERMINAL_INFRASTRUCTURE_RECEIPT_20260811.json.

PCF14 is the separately named infrastructure successor under the user's standing authorization. Its sole repair makes the already-frozen three-second candidate execution limit observable inside the trusted PID-1 bootstrap. Only after assessor bytes are compiled and dropped, the exact runtime descriptor is attested, PCF1_PYTHON_READY is emitted, and candidate policy is accepted, the bootstrap installs both SIGXCPU and SIGALRM handlers and arms ITIMER_REAL for the hash-bound candidate timeout. Either trusted signal exits with reserved resource-limit code 79 and therefore records scientific candidate_resource_limit. The outer timeout + 2 seconds watchdog remains an infrastructure-only fallback for namespace, launcher, or host failures. The candidate cannot import or reach signal, change the timer, or forge code 79; admission tests and the live allocation probe must prove those invariants.

No PCF13 artifact may be reused. PCF14 starts from a fresh root and replays the complete graph. The pinned dense host and tokenizer, sources, split, prompts, arms, seeds, 256-update training geometry, candidate language policy, three- second resource limit, generation settings, custody, thresholds, sealed board, and sole 1,289-row terminal gate remain unchanged. Disable requeue, retries, and automatic successors. Any infrastructure failure, formal PASS, or formal FAIL ends PCF14.