Jewels

The operational side of jewels, from the terminal or from code. The monitoring commands read the shadow ledger: overview is the per-method rollup (autonomy, agreement rate, pair counts, queue depth), pairs and pair walk the graded pairs down to the model transcript, and timeseries tracks the agreement trend. queue and resolve work the approve-mode review queue, and dryrun fires the live jewel at a subject in a disposable mirror, without recording anything.

The rest is the fine-tuning loop, where a method's graded pairs become a model you own. export reports the dataset or streams it as JSONL; train runs LoRA fine-tuning on the pairs and returns an adapter with a held-out agreement report; runs, run, and grade track and score the run. The winning checkpoint is served automatically as an ordinary model id, tuned/{appId}/{methodId}, so promoting a jewel onto its own model is a one-line change in its runTask call. Every command has a remy-admin jewels CLI form and a typed admin.jewels.* client method.