Tool: consensus_answer
Purpose
Aggregate branch outcomes and produce final answer with route confidence and route decision.
Current implementation mode: heuristic. The server ranks existing frontier nodes and applies verification policy gates; it does not perform model-backed semantic synthesis or calibrated confidence.
Input
deliberation_idmethod:self_consistency | weighted_vote | verifier_weightedpolicy_override(optional):min_independent_groups: integer (1-8)allow_tier3_for_independence: booleanrequire_tier1_unambiguous_for_critical: booleanfail_closed_on_critical_unresolved: boolean
Successful output
final_answerconfidencesemantic_confidence:nullunless a future model-backed provider supplies itconfidence_basisroute_decisiondisagreement_reportanalysis: provenance and limitations for the consensus run
When governance blocks convergence, the MCP tool returns an application error with code: "consensus_blocked", the safe route, blocker summary, and a recommended next action. That blocker response deliberately omits final_answer; clients must not present an unapproved candidate answer as a decision.
Behavior
- Fails closed on unresolved critical verification issues or assumptions.
- Requires at least one agent-authored frontier candidate and excludes heuristic scaffold/root nodes from an allowed final answer.
- Authorship requires a Thought node with server-recorded
agent_authoredprovenance; a caller-added tag on an Action or other heuristic node cannot satisfy the gate. Consensus provenance lists only the candidates actually used in the answer. - Blocks when a pipeline stage failed its gate or the latest checkpoint still reports blocking gaps.
- Requires a route decision in governed mode.
- A
policy_overrideapplies only to that call; it does not silently rewrite the session policy. confidenceis a route heuristic derived from verification status, not a model self-rating or historical calibration score.
Errors
- No eligible branches
- Policy, authorship, pipeline, or checkpoint gate failure