How do I install the Nn Audit skill?

Run npx skillmds@latest add zedarvates/nn-audit in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Nn Audit safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Nn Audit?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Nn Audit free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Nn Audit?

zedarvates (@zedarvates) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Nn Audit by zedarvates — Agent Skill

Name: Nn Audit
Author: zedarvates

Nn Audit

Audit the micro-NNs — is each model grounded in REAL data, or a synthetic copy of a hand-coded rule? Scans skills/botte_nn and reports, per model, the training data source (real/synthetic/unknown), whether the model file records provenance (trained_on/eval_accuracy), whether a test guards a real-world output, and a grounded/synthetic verdict. Deterministic, 0 cloud tokens. Use to tell which learned components are real vs placeholder, and which should be grounded or replaced by the rule they imitate.

zedarvates f666233 5 files · 14.2 KB Updated Sep 17, 2026

File contents

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.

---name: nn-auditdescription: Audit the micro-NNs — is each model grounded in REAL data, or a synthetic copy of a hand-coded rule? Scans skills/botte_nn and reports, per model, the training data source (real/synthetic/unknown), whether the model file records provenance (trained_on/eval_accuracy), whether a test guards a real-world output, and a grounded/synthetic verdict. Deterministic, 0 cloud tokens. Use to tell which learned components are real vs placeholder, and which should be grounded or replaced by the rule they imitate.---# nn_audit — grounded NNs, or synthetic copies of rules?A learned component only earns its place if it's trained on **real data a rulecan't capture**. A net trained on `np.random` + hand-coded labels justapproximates a deterministic function we already wrote — strictly worse than therule (adds error + opacity, learns nothing). This audits which micro-NNs are real.```bashpython -m skills.nn_audit.cli # audit skills/botte_nnpython -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.

Frequently asked questions