Reference

Commands

Command	What it does	Needs a model
rdlab map check <map.yaml>	Validates a dataset map and prints its canonical SHA-256.	No
rdlab project fingerprint <dir> --map <map.yaml>	Hashes the five project files, the map and canon/.	No
rdlab models	Lists the models a Pilot B build pins.	No
rdlab models --verify-live --licence-registry <path>	Checks the live Ollama manifests and licence digests against a licence registry.	Ollama on loopback
rdlab variety	Computes the deterministic variety report for a Pilot B build.	No
rdlab pilot-b --stub N	A calibration run of N triads on the real models; seals nothing.	Yes
rdlab pilot-b calibrate	Measures total eval_count for one operation; writes calibration.json.	Yes
rdlab pilot-b	Builds the Pilot B critic test set.	Yes
rdlab pilot-b certify-values	Certifies the generator’s value sets with two blind readers.	Yes
rdlab pilot-b compare-arms	Compares two certification runs.	No
rdlab --version	Prints the version.	No

Run rdlab <command> --help for every flag.

Exit codes

• 0: success.
• 1: refused. A structured {code, message, hint} JSON error goes to stderr.
• 2: a usage or file error, in the same structured form.

Licence registry

rdlab pilot-b, rdlab pilot-b calibrate and rdlab models --verify-live require --licence-registry <path> pointing at a licences/registry.json file. Omitting it refuses the run with PILOT_B_LICENCE_REGISTRY_REQUIRED. The build report records the registry path, its SHA-256 and the read date of each licence entry the model chain used.

What a build records

pilot_b_report.json and provenance.jsonl record the registry, each adapter’s class, num_ctx and max_num_ctx, the num_predict values and where each came from, and for every call the model, its digest, the prompt’s SHA-256, the token count, thinking_words, and the wall time.