model-evallisted

# Model Eval ## Overview The deep read for an AI model — a public HuggingFace model, **or a model folder already on your disk** (private, air-gapped, or already downloaded). The model-side counterpart of a git-repo eval. It answers: *what can this model's code do, does it describe itself honestly, and is it safe to run?* — with evidence, not vibes, and in plain English. One discipline governs everything: **an LLM-authored claim is a hypothesis, never a proof.** The value is not the model's say-so — it is catching the claim *contradicted* by the code's actual call graph, and refusing to blend correlated signals into false confidence (**No Clean Evaluator**: agreement among mechanisms that all read the same source is not corroboration). It **executes nothing** from the artifact, and it **never clears** it — a live custom-code surface always routes to a scoped human read. **Load `references/rubric.md` before scoring — it is the scoring contract.** ## Workflow 1. **Parse input & pick the mode.** If the argument is an existing **directory path**, run in **LOCAL mode** (a model folder on disk). Otherwise treat it as an HF `owner/repo` or URL and run in **HF mode**. Map any stated intended use to a risk tier (1 casual · 2 internal tooling · 3 production/security-sensitive; default 2). 2. **Collect signals** (deterministic, executes nothing): ``` # HF mode scripts/collect_signals.sh <owner/repo> # LOCAL mode (a model folder on disk) MAE_LOCAL_DIR=<fold