Converts a model family's upstream checkpoint into a reference-dtype GGUF that transcribe.cpp's loader can ingest. Produces ONLY the reference-dtype artifact (F32 / F16 / BF16 per the intake); the full quant matrix is a Stage 5 (porting-5-quants) concern. Use this after porting-2-oracle has emitted the contract tensor dumps. Input: intake.json, per-family converter, and the upstream checkpoint. Output: models/<slug>/<slug>-<REFDTYPE>.gguf, reports/convert/<variant>-<refdtype>.json, and Preflight Gate B green.
First stage for a new speech-model port. Produces reports/porting/<family>/<variant>/intake.json, drafts the family capability table, seeds the golden manifest skeleton, and clears Preflight Gate A.
Brings up the C++ implementation for a new model family at the reference dtype. Produces src/arch/<family>/*, finalized tolerances, resolved capability checks, batch parity, and full ref-dtype WER against the Oracle reference baseline. Use after porting-3-convert. Quant generation is Stage 5, not here.
Produces user-facing and upload-ready documentation for a completed port. Use after porting-7-wer has produced the ref-dtype gate pass and per-quant WER table. Output: filled family doc, model card, HF card YAML, rendered HF README, and private-repo docs upload.
Produces the shipped quant matrix from the reference-dtype GGUF, smoke-tests each, publishes the matrix to a private HF repo, and takes a tentative WER read (Modal or local) for human review. Use after porting-4-cpp has finalized tolerances and passed validate.py + the full ref-dtype WER gate. Input: models/<variant>/<variant>-<REFDTYPE>.gguf. Output: F16, Q8_0, Q6_K, Q5_K_M, Q4_K_M alongside the reference-dtype GGUF; a CLI smoke pass per file; quants pushed to a private HF repo; tentative per-quant WER. Authoritative quant WER is Stage 7. No tensor-level numerical comparison is required for quant acceptance — that is intentional.
Builds the oracle packet Stage 4 implements against: tensor dumps, reference transcripts, measured reference WER, and provisional tolerances. Use after porting-1-intake clears Gate A and before porting-3-convert. Output: tests/golden/<family>/<variant>.manifest.json, build/validate/<family>/<variant>/, reports/wer/<variant>-REF.<dataset>.{jsonl,score.json}, and tests/tolerances/<family>.json.
Runs the publication performance benchmark for a ported model variant and scripts the hypothesis → change → bench → accept-or-revert loop. Use after porting-5-quants has produced the full shipped quant matrix. Input: full quant matrix at models/<variant>/, reference machine matrix. Output: reports/perf/<machine>/<name>_<variant>_<backend>.json per bench run, scoped to the cells that ship in docs/models/<variant>.md. Every accepted performance iteration is followed by a validate.py all gate so a perf change cannot land while breaking ref-dtype numerics.
Full release WER sweep for a ported variant. Scores the full acceptance manifest across the reference dtype and every shipped quant. Ref-dtype C++ is hard-gated against the measured Oracle reference baseline; quant WERs are human-reviewed, not auto-gated. Use after porting-6-bench. Output: reports/wer/<variant>-<preset>.<dataset>.score.json per shipped preset, reports/wer/<variant>.<dataset>.summary.md, and a ship-gate decision.