# mif-rh-cli
Command-line interface for [`mif-rh`](../mif-rh), the compiled ontology
resolution/review engine for
[research-harness-template](https://github.com/modeled-information-format/research-harness-template)
corpora.
Drop-in replacement for rht's `scripts/resolve-ontology.sh`
(`mif-rh-cli resolve`) and `scripts/ontology-review.sh`
(`mif-rh-cli review`) — same flags, same exit codes, same
`TOPIC BOUND FIND STAMPED DISCOVERY UNTYPED INVALID` table/summary output,
same `ontology-map.json`/`--followup` backlog output.
`review` also acquires an exclusive lock (`<reports-dir>/_meta/.review.lock`)
for the duration of the run, and — with `--build-index` — rebuilds the
corpus-wide search index (`<reports-dir>/_meta/search-index.sqlite`) that
`mif-rh-mcp`'s `search`/`find_similar` tools read.
`suggest-type <TEXT> --topic <T>` (or `suggest-type --finding <path>`)
prints a JSON array of entity-type hypotheses ranked by embedding
similarity, each annotated with a confidence tier
(`auto_classify_eligible`/`flag_for_review`/`trigger_expansion`, MIF
ADR-020) under the corpus's calibration artifact
(`reports/_meta/confidence-calibration.json` by default; absent means
built-in thresholds and `calibrated: false`). Hypotheses only — it never
writes a finding's `entity_type`. With `--record` (requires `--finding`),
a tier-3 miss is persisted in the index for `expansion-candidates`.
The full ADR-020 tier routing:
- `review --suggest` writes tier-annotated suggestion queues to
`<reports-dir>/_meta/suggestions/<topic>.json` for this review's
not-durably-stamped findings (confirmed/rejected verdicts from
`/ontology-review --enrich` are preserved on re-runs), and records
tier-3 misses.
- `suggest-type` (and `review --suggest`) apply the negative-demotion-v1
gate for types carrying curated `negative_examples`: a candidate whose
query sits at least as close to one of its negatives as to its positive
document is barred from tier 1 (`negative_demoted: true` on the wire);
corpora whose packs carry no negatives score byte-identically to the
pre-gate engine, and the calibration artifact records participation as
`negatives_active`.
- `calibrate` derives the corpus's calibration artifact from its stamped
findings (`stamped-quantile-v1`: loosest floor+margin gate meeting
`--target-precision`, tier-2 floor from gold-recall quantiles). With
`--confusions <path>` it also exports the ranked confusable type pairs
(`confusions-v1` JSON: gold type, top-1 type, count, representative
finding ids) — the grounding input for `negative_examples` curation,
written even when the sweep itself fails to calibrate. Derived data:
regenerate it, never commit it.
- `expansion-candidates` clusters recorded misses (mutual similarity,
minimum cluster size, minimum distinct runs) into ontology-expansion
candidates, as JSON for `author-ontology.sh --from-clusters`.
`--relationship-script` is Unix-only: it spawns the given script directly
and relies on its `#!` shebang, which Windows does not honor. Leave it
unset on Windows (the default auto-detection already no-ops when the
script isn't found) or run under a POSIX-compatible shell.
## License
MIT