Direct training from VFS Federation

Version 0.0.11 can train the embedded final linear head from an explicit corpus stored inside the root Federation Image v2 owner. Training reads the image through normal bounded exact-path resolution. It does not crawl unrelated files, follow READ fallback rules, or mutate the opened image.

Corpus layout

Document	Purpose
Training manifest	Names the run, recipe, ordered record paths, epochs, ownership mode, and byte/step budgets.
Spindle recipe	Maps named record signals to every model feature in exact attachment order.
Training record	Contains signals, bounded output targets, verifier identity, evidence path, evidence digest, and `verified=true`.
Evidence file	Host-defined supporting material whose exact bytes must match the record SHA-256.

Run training

vfs-federation-neural training.vfsbin /ai/training/manifest.json \
  --proposal-out proposed-monk-loom.json \
  --report-out training-report.json

The report is also written as canonical JSON to standard output. Progress records go to standard error so automation can keep result data separate from human-visible status.

Results

The candidate attachment increments model lineage once per record and epoch. The report binds the manifest, dataset, recipe, starting model, final model, every extraction, every before/after evaluation, each adaptation proposal, and the number of parameter deltas. Both carry SHA-256 identities.

Nothing is committed: the source image remains byte-for-byte unchanged. A developer must review the ledger, place the candidate attachment into a new image, deep-verify it, and use the ordinary promotion path.

Ownership rule

Schema version 1 requires ownership_scope=root-image. The manifest, recipe, records, and evidence must all be files owned by the root image that carries the model. A path resolving into an embedded child fails with an authority error, even when it is otherwise readable.