← Complete research archive
Architecture researchPreregistered53 lines

PCF8: Ministral Publication Confirmation Successor

Terminal update: preparation, no-score mechanics, and B1 training completed. Draft task 751823 3 then failed on evc33 before generator entry because nvidia-smi reported no devices despite one allocated H100 GRES. No draft file or partial was published. All downstream jobs were ca…

docs/research/SHOHIN_PCF8_MINISTRAL_PUBLICATION_CONFIRMATION.mdOpen original Markdown ↗

PCF8: Ministral Publication Confirmation Successor

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

Terminal update: preparation, no-score mechanics, and B1 training completed. Draft task 751823_3 then failed on evc33 before generator entry because nvidia-smi reported no devices despite one allocated H100 GRES. No draft file or partial was published. All downstream jobs were canceled, the confirmation board was not semantically read, and no score or scientific PASS/FAIL was emitted. Preserve SHOHIN_PCF8_TERMINAL_INFRASTRUCTURE_RECEIPT_20260811.json.

Admission update: the required no-model H100 scratch qualification completed as job 751818 on evc23 in four seconds. It reported 183,262,756,864 bytes and 357,942,594 inodes available, verified the H100 identity and the existing write/rename/fsync boundary, and completely deleted a nested read-only mock model tree. The exact scratch path is absent. Model, data, and assessor access were all false. Preserve SHOHIN_PCF8_SCRATCH_QUALIFICATION_20260811.json; its remote receipt SHA-256 is 2cb17ef5...cb17.

PCF1 through PCF7 remain closed with formal scientific result null. PCF7 passed preparation and its complete no-score mechanics payload: the pinned model loaded, one ephemeral optimizer update completed, all 24 mechanics rows passed, and the immutable mechanics report says capability_scored=false. The allocation then failed only during its EXIT cleanup because the staged model retained read-only source modes and shutil.rmtree could not unlink it. B1 and all capability stages never started. The exact node-local tree was subsequently validated and permanently deleted by cleanup job 751816; all durable evidence was frozen. Preserve SHOHIN_PCF7_TERMINAL_INFRASTRUCTURE_RECEIPT_20260811.json.

PCF8 is the separately named, prospectively declared infrastructure successor under the user's standing authorization. It inherits every scientific and security clause of PCF1 through PCF7. The host, model and tokenizer semantics, source bytes, split, prompts, arms, seeds, training geometry, candidate policy, thresholds, assessor custody, and terminal gate remain unchanged. In particular, the observed Transformers tokenizer-regex warning is preserved as a disclosed risk; PCF8 does not change tokenizer flags or bytes.

Its sole repair is bounded scratch deletion. After validating the exact private direct-child scratch root, cleanup walks only its same-owner, same-device, nonsymlink directories without following links and restores owner write/search permission before deletion. No durable model/source tree is ever modified. Before a graph, run one no-model H100 scratch canary that creates a nested read-only mock model tree and proves complete cleanup plus the existing 128-GiB/150,000-inode and write/rename/fsync gates.

If that canary passes, preserve its receipt, replay the full live preflight, and submit one PCF8 graph. Disable requeue, retries, and automatic successors. Any infrastructure failure, formal PASS, or formal FAIL ends PCF8. Do not change or open an alternate host, tokenizer, data split, prompt, arm, seed, threshold, or protected board.