← Complete research archive
Architecture researchPreregistered71 lines

PCF17: Ministral Publication Confirmation Successor

PCF16 closed during infrastructure qualification and never submitted a publication graph. Its unmocked CPU allocation passed all 41 sandbox probes. Two simultaneous H100 workers on evc28 then loaded the exact dense host and PCF15 revision adapter and each completed 27 worst-case …

docs/research/SHOHIN_PCF17_MINISTRAL_PUBLICATION_CONFIRMATION.mdOpen original Markdown ↗

PCF17: Ministral Publication Confirmation Successor

Status: closed terminal-null on 2026-08-12. The sole graph completed every scientific generation stage, then precompute custody failed before score authorization or assessor open. Formal result is null; see SHOHIN_PCF17_TERMINAL_INFRASTRUCTURE_RECEIPT_20260812.json. This terminal status does not alter the frozen contract below.

PCF16 closed during infrastructure qualification and never submitted a publication graph. Its unmocked CPU allocation passed all 41 sandbox probes. Two simultaneous H100 workers on evc28 then loaded the exact dense host and PCF15 revision adapter and each completed 27 worst-case generation-plus-sandbox cycles. Both stopped in lockstep on cycle 28 because the separate outer launcher wall watchdog expired. Jobs 752844 and 752845 were FAILED 1:0, with zero restarts; scratch teardown completed, the queue is empty, protected access is zero, and no formal scientific result exists. Preserve SHOHIN_PCF16_TERMINAL_QUALIFICATION_RECEIPT_20260812.json (SHA-256 0f11c1a3...cec3) and the frozen qualification root. Do not reuse its receipt as admission evidence.

PCF17 changes one infrastructure constant only. The candidate's exact CPU soft/hard limits remain 3:4 seconds, address-space limit remains 1 GiB, file-size limit remains 1 MiB, and candidate wall parameter remains 3 seconds. The parent-side launcher watchdog grace increases from 2 to 30 seconds. This grace covers host scheduling plus posix_spawn -> prlimit -> Bubblewrap namespace setup/teardown; it does not grant candidate CPU time, alter trusted termination codes, or turn a timeout into a wrong answer. Any expiry remains a terminal PCF1SandboxError infrastructure event.

The explicit os.posix_spawn FD projection, empty environment and signal mask, exact reset signals, pinned /usr/bin/prlimit, Bubblewrap namespace, minimal runtime/ELF closure, sealed candidate/assessor transport, candidate policy, and bootstrap remain unchanged. Bootstrap SHA-256 remains 726af7a5...a60; generated-candidate policy remains f27124db...f1e. The prospective sandbox source SHA-256 is 7e97c28d...f1c4, and derived config SHA-256 is 70a37245...1802.

Before a PCF17 graph may exist, fresh Newton evidence must prove:

  1. one CPU allocation passes all 41 unmocked sandbox probes using the exact new source/config and zero protected access;
  2. two simultaneous H100 allocations on one eligible two-GPU dense node each load the exact pinned host plus frozen PCF15 revision adapter and complete 1,000 greedy-generation-plus-sandbox cycles on exact calibration MBPP identity e23c78d9...a8dc;
  3. both workers have zero infrastructure failures, zero restarts, exact model/adapter/sandbox hashes, and no confirmation access;
  4. exact postchecks prove queue empty, scratch absent, and durable headroom of at least 128 GiB plus 150,000 inodes.

Any qualification divergence closes PCF17 before science. A passing qualification authorizes exactly one complete graph at a fresh root, with no reuse of any PCF14/PCF15 artifact or PCF16 qualification state.

The pinned dense host/tokenizer, source banks, identity order, prompts, arms, seeds, 256-update training geometry, generation settings, candidate CPU semantics, candidate policy, custody, sealed board, thresholds, and sole 1,289-row gate remain byte-for-byte unchanged:

  • unchanged reaches at least 387/1289 with every domain nonzero;
  • revision exceeds unchanged by at least 65 and self-refinement by at least 39, with no per-domain loss;
  • learned commit exceeds revision by at least 13, retains at least 95% of revision-correct and unchanged-correct identities, and loses no domain;
  • all arms have exact ordered coverage, zero truncation/malformed evidence, complete custody, one assessor open, and zero holdout/public/product access.

Requeue, retries, partial continuation, automatic successors, and protected split access are disabled. Infrastructure failure, formal PASS, or formal FAIL ends PCF17; a formal result authorizes no successor.