Varimu Neural Training

Vari-Mu neural training

Treat training as an evidence pipeline, not as permission to recursively consume audio. The exact dataset allowlist is the sole eligibility authority. Preserve raw hashes and route every consumed source through the repository training ledger.

Required order

1. Read references/current-manley-policy.md and resolve the exact batch, device state, target variant, objective ID, and claim tier. Stop on any unresolved eligibility conflict.
2. Validate the capture package with the hardware-compressor-capture skill. Completed capture payload may have been deleted locally after verified Release publication; in that case restore only the pinned Release members into the claimed workspace and verify every source hash before creating the READY receipt. Create the local READY receipt with tools/mc_capture_training_ledger.py claim before optimization. Do not infer eligibility from folders, names, or split labels.
3. For a newly captured batch with no candidate algorithm, run the paired Original/Target baseline in references/forensic-validation.md. This is a two-signal characterization; do not invent a Current render.
4. Read references/training-plan.md, freeze the split, metric registry, normalization floors, baselines, model budget, and acceptance gates, then train in a local claimed workspace.
5. For every usable checkpoint or algorithm version, render the frozen Original through it and run the complete three-way Original/Target/Current forensic comparison plus audit_forensic_metric_completeness.py. Report every comparable statistic and all missing or non-comparable fields. Never select from one aggregate score alone.
6. Validate the completed run directory, seal the local training ledger, retain the complete local folder plus verified package parts, publish the large immutable package as create-only GitHub Release assets, then publish only hashes, annotations, the ledger receipt, and the Release link under varimu-training/release-index/ in one create-only Git commit.

Multi-machine rule

Read references/multi-machine-protocol.md before installing or operating a worker. Each machine runs watch-once every 600 seconds. A worker may claim only when it has no local active run. The canonical queue/<batch-id>/claim.json is created with create-only semantics; concurrent losers must not overwrite, rename, delete, or take over the claim.

Run the following relative scripts/ commands from the skill directory that contains this SKILL.md, or resolve their paths explicitly first.

Use:

python3 scripts/varimu_training_coordinator.py watch-once --config /absolute/config.json

Install the ten-minute monitor with:

python3 scripts/install_monitor.py install \
  --repo-root /absolute/path/to/mc-patcher \
  --repo shawnxxxqazwsx/mc-patcher \
  --branch main

The Git repository Code page is the lightweight coordination and index plane. Large capture/training audio, derived data, checkpoints, reports, and package parts must not be committed there and must not use Git LFS. They are hashed, compressed, split below GitHub's per-asset limit, and uploaded to a unique create-only GitHub Release. varimu-training/release-index/ contains only the completion receipt, detailed annotation, manifests, ledger receipt, SHA-256 inventory, and Release link. Packaging and publication copy from the complete local run and verified local parts; they never move or delete unique local evidence. The removed trained Data/ namespace must never be recreated.

Stop conditions

Stop before training or publishing when any of these is true:

• the exact allowlist and file annotations disagree;
• a validation/withheld/rejected/superseded file is offered to optimization;
• primary and comparison-only target variants are mixed in one objective;
• raw sources or split manifests changed after the READY ledger claim;
• the batch already has another machine's claim;
• the local worker already has a CLAIMED, TRAINING, or FINALIZING run;
• a candidate lacks the complete three-way forensic artifact set;
• completion annotations omit the machine name, machine ID, UTC timestamps, source hashes, code/environment identity, model artifacts, or known limits;
• the GitHub Release tag/assets or Code-page completion index would overwrite or disagree with existing evidence;
• the local complete folder or verified split-package backup is missing.

Evidence and claim language

Follow references/completed-run-contract.md. Training success establishes only the frozen gates that actually passed. Hardware-fit or product-ready claims additionally require independent withheld captures, controlled probes, listening, realtime engineering gates, and the declared plug-in/DAW path.