ontology-field-trial

# Field trial — does the vault survive being the only thing you have? This project sells a meaning layer that an agent consumes. The only test of that claim is to **take the source away** and see whether the vault still answers. Everything else — node counts, green checks, a map that looks full — measures effort, not usefulness. Use the [Atlas meta-model specification](../../../docs/ONTOLOGY-ATLAS-SPEC.md#2-the-five-authorable-node-kinds-and-reserved-reader-kind) as the shared answer key for kind, relation, direct `is_a`, and unsupported inference claims. The trial measures whether agents can apply that contract; it must not invent a second rubric from the vault it is judging. Run this as four measurements, in order. Each one produces a number or a list that goes in the report. **Do not skip a phase because the previous one looked good** — phase 1 has looked good every time, and phases 2–4 are where the defects were. ## Before you start Pick a repository **nobody in this session knows**, permissively licensed (Apache-2.0 / MIT / BSD), in a domain you cannot bluff. That is the point: a familiar repository lets the agent — and you — fill gaps from memory, and the whole trial then measures your memory instead of the vault. ⚠️ **Never name the chosen repository in this repo's code, tests, identifiers, fixtures, or commit messages** (`.claude/rules/forbidden.md` — no third-party brands in identifiers). Everything the trial produces lives outside this checkout — a scratch dir...