A sequence of gates, not a schedule to be executed.

Freeze the audio/video exclusion rule, the named data owner and two independent approvers, immutable storage prefixes, the reviewed container image digest, per-job spend and wall-clock limits, and the deterministic Batch concurrency, retry and capacity limits.

Exit: every identity, prefix, digest, limit, reviewer and stop procedure is exact. Readiness still reports execution disabled, but no planning ambiguity remains.

Run the already-selected canary — 512 records covering all 212 populated extension/size strata — using version-specific S3 reads. Verify signatures, active-content risk, encryption, corruption, unsupported formats and extractor routing.

Exit: all 512 inputs reconcile to a terminal disposition; raw has zero writes; every result binds its exact source version and pipeline digest; no audio/video body is decoded.

Deterministically select a representative pilot across DOCX, PPTX, native-text PDF, plain text, XML/HTML/RTF/CSV and selected legacy Office strata. Exercise malware scanning, bounded container handling, extraction, normalization, lineage, resume and reconciliation.

Exit: one terminal disposition per input; repeated deterministic stages produce matching digests; resource ceilings are measured; failure and quarantine samples get human review.

Process every eligible native-text object through malware screening, bounded extraction, deterministic normalization, initial language/subject/level labelling and immutable derived shards.

Exit: the full accepted inventory reconciles exactly; every object has an explicit terminal result; nothing is overwritten; output-volume and curriculum distributions are published.

Exact, normalized-text and bounded near-duplicate clustering; development and evaluation families frozen; overlap screened; rights, privacy, cultural and quality review completed. Deduplication weights or selects representatives and never deletes lineage.

Exit: train/dev/eval manifest digests are disjoint; unresolved contamination is zero; restricted or unknown-rights items are absent from approved partitions; exact record, byte, unit and token counts are frozen.

Train and compare a 48K SentencePiece unigram tokenizer against a 48K byte-level BPE, using only the approved training partition. Measure Chinese and English fertility by education stage, mathematical and scientific symbol round-trip fidelity, byte fallback and token-count stability.

Exit: one candidate selected from frozen evidence; trainer, config, seed, input manifest and artefact digests exact; development and evaluation content never inspected; shard order and token counts reproduce.

A randomly initialised model on deterministic synthetic data only, structurally prohibited from referencing the research dataset freeze. This validates packaging, BF16 execution, logging, heartbeats, finite-loss stops, checkpoint publication, resume and teardown.

Exit: no research data was readable by the job; checkpoint/resume preserves run lineage; emergency stop and zero-capacity teardown are demonstrated; no training-quality claim is made. GPU launch requires a new explicit instruction.

Train the smallest selected dense model, not exceeding 200M parameters, on a frozen approved curriculum slice using one H100. Record throughput, utilisation, memory, loss, held-out progression, restart behaviour and exact cost.

Exit: all inputs, code, environment, image, tokenizer, seed and outputs immutable; loss finite; curriculum-retention evidence complete; no leakage or unapproved input; all billable compute returns to zero capacity.

Reconcile budget, artefacts, quality, failures, utilisation and effective tokens per dollar against the pre-registered promotion criteria.

Exit: publish exactly one decision — go (prepare a separately budgeted 500M proposal), hold (name the defect to correct before replay), or stop (preserve artefacts and close the cycle). M8 cannot launch 500M training.