Evaluate

evaluate judges an artifact against a rubric, spec, checklist, or acceptance criteria using two provider-CLI-backed AI peers, preserving unified findings, per-peer reasoning, and dissent in the deliberation log.

For an operator walkthrough of evaluation inputs, expected JSONL, sidecar output, and dissent review, see the consensus plugin README and skills/evaluate/references/operator-qa.md.

Evaluate an artifact against a rubric

node plugins/consensus/skills/evaluate/scripts/consensus-evaluate.mjs artifact.md --rubric rubric.md

Defaults

The evaluate wrapper defaults to:

• --cold-start shared_input (cold start)
• --iteration parallel_revision
• --agency minimal

parallel_revision with minimal agency is the default because independent judgment and visible disagreement are the point of evaluation. The wrapper requires --rubric <path> and rejects --cold-start independent_draft because evaluate compares an existing artifact against a rubric. independent_draft is for the create/decide/plan wrappers that start from briefs, options, or goals rather than an existing artifact.

Output

Without --output, the wrapper writes a sidecar named <artifact>.evaluation.md and reports it in the run_completed event; with --output <path>, that path receives the artifact. The evaluation artifact contains:

• unified findings from the converged or last-agreed evaluation document
• embedded consensus-verdict records for each peer turn
• ## Dissent for residual concerns after convergence, when present
• ## Unresolved dissent for impasse or escalation states

At minimal agency, unresolved peer disagreement is surfaced rather than silently decided — never hidden.

Guided rubric creation

If you want an evaluation but do not have a rubric yet — or you ask for help authoring one — the evaluate skill runs a host-model guided flow: it elicits your evaluation goals, adapts one of the bundled example rubrics, writes a draft to a path you approve, then invokes the wrapper with --rubric. The raw --rubric contract above is unchanged for users who already have a rubric.

Rubric criteria are the ##–###### headings and - / * bullets in the file, and the wrapper uses the first 12 distinct criteria, so keep the load-bearing ones near the top. Weights and scoring scales are peer-facing guidance, not machine-parsed structure.

Four ready-to-adapt example rubrics ship under skills/evaluate/references/examples/:

• general-purpose.md — annotated template for any written artifact
• code-review.md — pull request descriptions and implementation proposals
• technical-writing.md — documentation, tutorials, and API reference
• design-architecture.md — ADRs, system designs, and RFCs