Olympus

An agent skill that answers software questions with a decision argued to consensus by a panel of agents, then ships that decision through a gated pipeline with two independent reviews. It handles decisions like redesigning dialogs, moving a flow to a new architecture or revoking sessions when a role changes.

Optionally. Jev runs a verdict meta-check on Arena rulings: a Noul per verdict asked with both candidates in both orders, consulted only after the ruling is recorded. It ships in shadow mode, where it calls and logs only, and can add a judgment but never remove a check.

No. Everything works without it, and any Jev failure becomes jev: UNAVAILABLE while the process continues. The optional tools accept a pinned jev-1.13.0 model only; jev-latest and jev-preview are rejected, and the question wording is frozen with a sha256.

A smaller panel used Olympus to decide the integration. As a meta-judge Jev flagged exactly the two most ambiguous verdicts; as a judge it was order-inconsistent and favored the lowest-ranked proposal, so that role was dropped. The J1 evidence is AUC 0.81 on 24 verdicts.

Copy the directory into your agent's skills folder, for example .claude/skills/olympus/ or .agents/skills/olympus/. Its commands name one specific harness, so map each to your own equivalent or remove it before use; the process itself is harness-independent.

A single case study in which Olympus produced its own SKILL.md: 6 seats, 15 pairs judged twice, rounds 3 to 20, and a pre-mortem that caught gaps a unanimous draft missed. The README lists it as a single case, not a controlled experiment.

At least 30 independently authored seed pairs from 2 authors, 20 wrong verdicts, a cluster-bootstrap AUC whose 2.5th percentile is at least 0.70 and beats the word-count baseline, and stable probabilities across orders; tools/jev_calibrate.py prints PROMOTE, HOLD or RETIRE.