The auditable council. Every run can now produce a structured, agreement-scored, fully auditable verdict plus a redacted execution manifest — on top of the existing synthesize / raw / debate / adversarial modes.
conclave ask "<decision>" now returns a verdict: a headline + recommendation, a consensus score computed by reproducible arithmetic over the models' own clustering (never an LLM-emitted number), conflicts citing which providers disagreed (with evidence), and minority reports — alongside a manifest recording which providers were called and how the verdict was made. Streaming and --json carry it too.
Highlights
- CouncilResult v2 —
verdict,consensus_score/method/label,conflicts,provider_votes,minority_reports,manifest(all additive, backward-compatible). - ModelHarnessManifest with verdict-extraction provenance — no secrets, ever.
- Deterministic consensus (
position_cluster_ratio_v1) — the auditability fix: the number is arithmetic over the model's clustering, not an LLM figure. - Native structured output across OpenAI / Anthropic / Gemini, threaded through
call_model. - Static provider capability catalog; CLI verdict rendering +
--jsonmanifest; streaming verdict parity.
The verdict is optional — open-ended prompts, single-member councils, and extraction failures degrade gracefully to synthesis-only. Default-on; opt out with Council(extract_verdict=False).
Full notes in CHANGELOG.md. Install: pip install -U conclave-cli