PCF16: Ministral Publication Confirmation Successor
Status: prospectively frozen on 2026-08-12 before admission or scientific compute.
PCF15 is terminal-null. Preparation, mechanics, B1, all 16 source-only draft
shards, exact draft merge, materialization, and the frozen 256-update revision
training completed with zero restarts. All eight calibration allocations then
ran beyond PCF14's earlier failure point. Revision-calibration shard 3 stopped
after 768/1456 rows on evc28 because the isolated Python bootstrap did not
emit trusted READY. The seven peer allocations and every descendant were
cancelled. Commit training and confirmation generation never started; the
prepared confirmation assessor board has zero semantic reads; no score
authorization, score, normalized report, compute custody, final comparison,
or formal scientific result exists. Preserve
SHOHIN_PCF15_TERMINAL_INFRASTRUCTURE_RECEIPT_20260812.json (SHA-256
a252dfc7...ff8e) and the frozen remote PCF15 run. Do not replay or continue
PCF15.
PCF15 removed Python preexec_fn, but its subprocess call retained
pass_fds, which forces Python away from its safe posix_spawn fast path and
through a fork/exec implementation in the resident multithreaded 9B
PyTorch/CUDA process. PCF16 removes that last parent-process fork boundary.
It invokes exact /usr/bin/prlimit through os.posix_spawn with five explicit
POSIX_SPAWN_DUP2 actions: assessor stdin, bounded stdout/stderr, sealed raw
candidate, and Bubblewrap namespace-info FD. Child candidate/info FDs are
chosen deterministically in [32,63]; every other Python-opened descriptor is
close-on-exec. The spawned process receives an empty environment, and the
existing wall-time monitor waits and fail-closes exactly as before.
The exact resource chain remains prlimit -> bubblewrap -> pinned minimal Python: CPU 3:4 seconds, address space 1 GiB:1 GiB, file size 1 MiB:1 MiB, private PID/proc/dev/tmp, no network, read-only root/runtime/candidate,
sealed assessor transport closed before trusted READY, bootstrap exit
attestation, and infrastructure/scientific termination separation. Bootstrap
SHA-256 remains 726af7a5...a60; generated-candidate policy remains
f27124db...f1e; the three-second scientific candidate limit is unchanged.
The prospective sandbox source SHA-256 is 448730d6...cf2d, and its derived
config SHA-256 is d8f2060a...3733.
Before a PCF16 graph may exist, all of these infrastructure-only gates must pass on fresh Newton allocations and be frozen with exact source/runtime/model hashes and zero protected access:
- one CPU allocation runs the complete unmocked sandbox qualification with
all 41 probes true using the
os.posix_spawnproduction path; - two simultaneous H100 allocations are co-located on one eligible two-GPU dense node, each loads the exact pinned Ministral host plus the frozen PCF15 revision adapter and repeatedly performs greedy generation followed by the exact late-failure calibration MBPP assessment;
- each H100 allocation completes at least 1,000 generation-plus-sandbox iterations with zero infrastructure failures, zero restarts, exact environment/model/adapter/sandbox hashes, and no confirmation access;
- an independent postcheck verifies the queue empty, private scratch absent, and at least 128 GiB plus 150,000 inodes of durable headroom.
Any qualification divergence closes PCF16 before science. A passing qualification authorizes exactly one fresh complete graph at a new immutable root. PCF16 may not reuse any PCF14/PCF15 source view, B1 adapter, draft, revision adapter, calibration output, or partial artifact in that graph.
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, generated-candidate policy, sandbox visibility boundary, 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
387/1289with every domain nonzero; - trained revision exceeds unchanged by at least
65correct and self-refinement by at least39, with no per-domain loss against either; - learned whole-trajectory commit exceeds revision by at least
13, retains at least 95% of both 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 PCF16. A
formal result must be preserved exactly and authorizes no successor.