nn_audit — grounded NNs, or synthetic copies of rules?

A learned component only earns its place if it's trained on real data a rule can't capture. A net trained on np.random + hand-coded labels just approximates a deterministic function we already wrote — strictly worse than the rule (adds error + opacity, learns nothing). This audits which micro-NNs are real.

python -m skills.nn_audit.cli                 # audit skills/botte_nn
python -m skills.nn_audit.cli <dir> --json

Per model it reports:

• data_source — real | synthetic | unknown, inferred from the training script's content (distill/corpus/labelled → real; np.random → synthetic).
• wired — does any production file (not tests/training/registry) consume it, and which (usage)? Tells synthetic-but-driving-behaviour (real risk) apart from synthetic-but-orphan (dead weight).
• has_provenance — does the .json record trained_on / eval_accuracy / data?
• has_test_guard — does a test assert a specific real-world output for it?
• verdict — grounded | synthetic — drives behaviour: ground it | synthetic + orphan: delete or wire | unknown; at_risk = synthetic AND wired.

The fix is decided by both axes: a wired synthetic net (e.g. binary_router in the routing belt) must be grounded on real data (distillation / active- learning, like error_classifier); an orphan synthetic net is dead weight to delete or wire. Exposed via [[llm_mcp]] as nn_audit; pairs with [[botte_nn]]. Pure file inspection, 0 cloud tokens.