PCF1: Ministral Publication Confirmation
Status: prospectively frozen before any PCF1 data mutation, model load, or compute on 2026-08-11. This is the only authorized confirmation in this phase. A completed gate, pass or fail, ends the phase.
Purpose and immutable boundary
PCF1 tests only the transferable content of the qualified dense Shohin release:
source -> model-owned draft -> trained same-family revision
-> learned whole-trajectory commit -> one coherent answer
The positive control is the immutable Qwen3.5-9B release whose protected
product result is learned commit 383/538, trained revision 374/538, and
unchanged continuation 316/538. PCF1 does not modify or retest that release.
It does not reopen NDR1, KCR1, VTE1, any natural-language-to-microcode bridge,
the Qwen3.6-35B-A3B edit cascade, a small-OLMoE route, or any closed prompt,
seed, rank, layer, duration, decoding, threshold, or data-mix variant.
The sole new host is the cached dense language path of
mistralai/Ministral-3-8B-Reasoning-2512 at exact revision
81eaece1948f3875421d9a45bc55487d10e2d894. The snapshot must receive a
complete verification of its pre-existing immutable 58-entry manifest,
SHA-256
46cc9203a18a414e08a53109662c3802b57c046896185ca9ab31875e8167cf1f,
and pass one no-score mechanics/admission replay before the capability graph
may start. The manifest was created by completed download job 747023, covers
35,706,515,534 bytes, and verified read-only before this implementation used
the model. No alternate host is a contingency.
Its pinned configuration is the multimodal
Mistral3ForConditionalGeneration wrapper with a dense 34-layer
Ministral3Config language path; every PCF1 program therefore requires the
explicit established multimodal loader and never falls back to a causal
auto-class. Its exact final-four trainable language-layer indices are
[30, 31, 32, 33]. The execution environment is likewise pinned to the
existing product-reasoning-b3a3603-r2 receipt: Python 3.13.13, Torch
2.6.0+cu124, Transformers 5.15.0.dev0, runtime receipt SHA-256
277b97fbd6b18760c9789cf3f3372bdb6b40ca87bf84a1df4b41ee3194c4e9dd,
and package freeze SHA-256
1d4dfd4a1dc11af9788b0bab072d262278db1814d3fca49465d4df5931b3b87a.
Every generated MBPP program, including training-only calibration programs,
must execute behind the same pinned Bubblewrap OS-isolation boundary. The
candidate sees only its anonymous read-only program, a read-only minimal
Python runtime, private /proc and /dev, and an empty temporary filesystem;
it receives a clean environment, no network namespace, no parent process, and
no campaign, source, assessor, model, evidence, or user filesystem. The
Bubblewrap binary is /usr/bin/bwrap 0.4.0 with SHA-256
eb767688b8224d8d3dbe1f8cb30ac3dff9ae8b02ff0452eaec9f94874d4e0011.
The no-score mechanics gate and every code-scoring allocation must reproduce
a hash-bound escape probe before executing any generated program.
The earlier decision to cancel an unused dense Ministral detour was a priority
decision made during the MoE phase, not a scientific result. The 2026-08-11
publication directive supersedes that priority decision for PCF1 alone.
At inference the runtime sees only the problem and model-owned trajectories. It receives no target, correctness bit, task label, benchmark route, solver, verifier, tool, retrieval result, or external proposal model. Every commit selects one complete candidate byte-for-byte; it never splices fields.
Frozen lineage and data custody
The draft role is a source-only final-four-token-mixer rank-8 LoRA trained for
256 AdamW updates on the existing verified B1 corpus, SHA-256
2461d6f70b44a142854d56c24e1fb42d600065e5788a2c4e055ba47b12696549.
It uses the qualified B1 settings: maximum sequence 1,024, batch one,
accumulation 16, learning rate 2e-4, alpha 16, model seed 2026080711, and
data seed 20260802.
The frozen source universe is the existing CVG1 pair bank and three verified, evaluation-disjoint source banks:
- pairs:
45f1d66ce5e87dc2a1f4c3594bdde2bae26e9417e879d16eb4eddb228b696afe; - math:
e0ede83257e441050a019f59fb13d9c85bd6cba1d6a755ab86fb7129966ddbe5; - logic/science:
5a96859fd9088cde598b61da60dd2c6cb7281323ee06c034742a1b4e0e237017; - executable MBPP:
0b6d068b4d71f407cb234579b9278dc640df09139ea906dd0f52a6ab71e05398.
NUL-delimited source identity and split seed 2026080811 remain unchanged.
A CPU-only custodian freeze writes a new nonsealed model-input root containing
only train and development source identities. It writes the 1,289 development
assessors as a distinct write-once CPU-only artifact outside that root; no
confirmation runtime row or GPU report contains an assessor, target,
correctness field, execution result, or per-arm score. The custodian records
only the count and sorted-identity digest of the existing 1,279-row holdout;
it must not copy holdout questions, answers, assessor fields, or prior model
trajectories. The protected 538-row product is not an input to any PCF1
program. After the exact hash-pinned CPU source freeze (whose historical bank
directory happens to be named product_reasoning), every model-visible data
path, output path, schema, and requested split fails closed if it contains
holdout, product, or public.
The exact B1 bytes and the unmodified qualified generic trainer receipt retain
their historical control-plane schema names. They are admitted only behind
the pinned B1 hash in a safe path; schema metadata is never a runtime field,
and no PCF1 evaluation, candidate, selection, or result uses a legacy schema.
The draft adapter generates one greedy, no-thinking, at-most-768-token draft
with seed 2026080818, batch size two, and exactly 16 ordered shards for every
nonsealed identity. The revision adapter
warm-starts from that exact draft checkpoint and receives source + exact draft. Its 5,824 unique training identities retain the qualified presentation
geometry: single-correct outcomes appear four times and both-correct or
both-wrong outcomes once, for exactly 9,655 presentations. It trains for 256
updates with batch one, accumulation eight, context 4,096, learning rate
2e-5, final-four-token-mixer rank 8, alpha 16, model seed 2026080815, and
data seed 2026080814. Verified complete targets are supervisor-only.
Commit calibration uses only training identities and candidate outputs from
the frozen trained revision and unchanged continuation. The learned commit is
the simpler independent shared-candidate scorer already qualified as the
matched AQC control; antisymmetry is not a PCF1 claim. Calibration is
source-disjoint from the 1,289-row confirmation board. The commit host is the
same pinned Ministral backbone plus the frozen draft adapter. Candidate order
is presented both ways and must be exactly consistent. The qualified
independent-control recipe remains exact: 80/20 calibration split seed
2026080820, training seed 2026080822, 128 updates, accumulation eight,
maximum sequence 3,072, head width 512, backbone LR 2e-6, head LR 2e-4,
and tie-loss weight 0.25.
Matched arms
All capability arms use the same 1,289 source identities, exact model-owned
drafts, source-plus-draft serialization, maximum 768 final tokens, scorer,
greedy evaluation seed 2026080816, batch size two, and exactly four ordered
shards per arm. Each of the two training-only commit-candidate calibration
arms, trained revision and unchanged continuation, also uses four ordered
shards; self-refinement is a confirmation control and is not an extra
calibration experiment.
They record
prompt/generated tokens, wall time, peak CUDA,
trainables, and charged GPU-seconds.
- Trained revision — the frozen warm-started revision adapter.
- Unchanged continuation — the frozen draft adapter reads the exact treatment prompt and receives the same generation budget.
- Self-refinement — that same draft adapter reads the exact draft under one frozen generic review-and-correct instruction and the same budget.
- Learned coherent commit — the frozen independent scorer chooses the complete trained-revision or unchanged-continuation trajectory.
There is no long-generation, best-of-K, router, hidden-draft, transaction, edit, microcode, or alternate-commit arm. An infrastructure failure is preserved as terminal infrastructure evidence; it is never counted as a wrong answer and does not authorize a shard replay, retry, requeue, or successor.
The single falsifiable gate
The 1,289-row development confirmation opens and scores exactly once, in one CPU process, after all three generated arms, the learned-commit application, their candidate hashes, scheduler accounting, and the pre-score custody authorization are immutable. That process is the sole reader of the separate assessor artifact and scores revision, unchanged, self-refinement, and the selected whole trajectory atomically. PCF1 passes only if every Boolean below is true:
- Capable host: unchanged continuation solves at least
387/1289(30.0%) and solves at least one math, one logic/science, and one executable code identity. - Causal revision margin: trained revision solves at least 65 more identities than unchanged continuation and at least 39 more than self-refinement.
- Revision retention: trained revision has nonnegative math, logic/science, and executable-code correct-count deltas against both matched controls.
- Useful learned commitment: coherent commit solves at least 13 more identities than trained revision.
- Conservative commitment: commit retains at least 95% of the revision-correct identities and at least 95% of the unchanged-correct identities, and has nonnegative per-domain correct-count deltas against revision.
- Complete custody: every arm covers the same
1289/1289identities; candidate and selection outputs are nonempty and well formed; candidate assessment truncation and malformed selections are zero; A/B semantic order consistency is1289/1289; all pinned model, data, runtime, execution-environment, code-sandbox/probe, checkpoint, result, and compute hashes verify; and the holdout, public, and product access counters remain zero.
PASS = capable_host AND causal_revision_margin AND revision_retention AND useful_learned_commitment AND conservative_commitment AND complete_custody.
Any false conjunct is the final PCF1 result. It closes exact PCF1 without a model, source, schedule, prompt, selector, threshold, seed, or decoding retry. A pass records that the separately sealed holdout and product may be considered by a later explicit authorization; PCF1 itself does not open them.
Execution and stop rule
- Freeze and test the CPU source firewall, runtime, model manifest, and all dispatch receipts without reading a score.
- After durable storage is within a verified safe quota, run one 24-presentation mechanics/admission job. It may validate loading, gradients, adapter restoration, serialization, masking, commit order, and the code-sandbox escape boundary, but may not emit a capability statistic.
- Run the one dependency-bound graph: draft fit, nonsealed draft collection, revision fit, training-only commit calibration, the three label-free matched development arms, learned commit application, pre-score custody, one atomic CPU score, and the gate.
- Mirror reports and compute receipts, verify hashes independently, update the master/status/runbook ledgers, and stop. Do not submit a holdout, product, public, alternate-host, or successor job.
At freeze time Newton had no queued/running jobs and sufficient idle H100s, but the account reported approximately 612 GiB and 1.326 million inodes over hard Lustre quota. Therefore no PCF1 remote write or job is authorized until storage safety is resolved without deleting scientific evidence.