/adjudicator-synthesis

name: adjudicator-synthesis
description: "Use when you have run 2 or more agent-council deliberator skills on the same artifact (Skeptic, Voice & Identity, Evidence & Calibration, Strategy & Stakes) and want their structured critiques synthesized into a single SHIP / REVISE / HOLD verdict with a concrete revision brief. Encodes the Adjudicator role from the agent-council 5-perspective quality gate. The Adjudicator does not produce a new critique — it produces the decision document the operator acts on."
version: "0.1.1"
type: "codex"
tags: ["Quality Gate", "Verdict", "Synthesis", "Council Adjudicator"]
created: "2026-05-29"
valid_until: "2026-11-29"
derived_from: "prompts/adjudicator.md in Avyayalaya/agent-council"
tested_with: ["Claude Sonnet 4.6", "Claude Opus 4.6", "GPT-4o"]
license: "MIT"
composes_with:
  - package: "agent-council"
    skill: "skeptic-review"
    relation: "consumes_output_of"
    reason: "Adjudicator consumes Skeptic's structured critique (would_block, irreducible, top_3_failure_modes) and applies the verdict policy."
  - package: "agent-council"
    skill: "voice-identity-review"
    relation: "consumes_output_of"
    reason: "Adjudicator consumes Voice & Identity's structured critique (voice_violations, register_match, cxo_test, would_block) and integrates it into the verdict reasoning."
  - package: "agent-council"
    skill: "evidence-calibration-review"
    relation: "consumes_output_of"
    reason: "Adjudicator consumes Evidence & Calibration's claim_tier_map + calibration_issues + would_block into the verdict reasoning + revision brief."
  - package: "agent-council"
    skill: "strategy-stakes-review"
    relation: "consumes_output_of"
    reason: "Adjudicator consumes Strategy & Stakes's goal_alignment + opportunity_cost + kill_check into the verdict reasoning."
capability_summary: "Synthesizes 2-4 deliberator critiques (Skeptic, Voice & Identity, Evidence & Calibration, Strategy & Stakes) into a single decision document: SHIP/REVISE/HOLD verdict applying the verdict policy, ≤3-sentence reasoning naming what converged and what diverged, numbered revision_brief if REVISE (operator-actionable in ≤60 minutes), dissent_summary if HOLD, convergence_notes surfacing where deliberators caught the same underlying issue from different angles. Output is fenced JSON, voice-gated."
input_schema:
  artifact: "string or path — the text artifact under review (for grounding the revision brief)"
  deliberator_critiques: "object — required, the structured outputs of 2-4 deliberator skills, keyed by role (skeptic, voice_identity, evidence, strategy)"
  verdict_policy: "object — optional, override of the default SHIP/REVISE/HOLD rules"
  prior_verdicts_on_artifact_type: "array — optional, up to 5 prior verdicts on the same artifact_type for compounding memory"
output_schema:
  verdict: "SHIP | REVISE | HOLD | INCOMPLETE"
  reasoning: "≤3 sentences explaining the verdict, which deliberators blocked, what their root issue was"
  revision_brief: "Numbered list as string. Each item: deliberator + concrete change + line/paragraph. Null when SHIP. Required when REVISE or HOLD."
  dissent_summary: "≤4 sentences. Which deliberators blocked, what converged in R2, what diverged, any irreducible flags"
  convergence_notes: "≤2 sentences naming where two or more deliberators caught the same underlying issue from different angles"
example_invocation: "examples/adjudicator-on-pitch-bundle.md"

Important: this Skill is asymmetric

Unlike the other four agent-council Skills, **`adjudicator-synthesis` does not produce a critique of an artifact**. It is a synthesis Skill, not a deliberator. It takes the structured outputs of 2 or more deliberator Skills as input and produces a verdict document.

**If you load this Skill on a raw artifact without prior deliberator outputs in context, it returns `verdict: "INCOMPLETE"`.** Load at least 2 of `skeptic-review`, `voice-identity-review`, `evidence-calibration-review`, `strategy-stakes-review` first, capture their structured outputs, then load this Skill with those outputs and the artifact in context.

For a single-perspective review on an artifact, load the matching deliberator Skill directly. The Adjudicator never runs first.

Purpose

You have run 2 or more deliberator skills on the same artifact. Each produced a structured critique. Now you need one synthesized decision: **SHIP**, **REVISE**, or **HOLD**. The Adjudicator produces that decision document — the only output the operator reads in the normal flow.

The Adjudicator is **not** a fifth deliberator. It does not produce a new critique. Its output is a decision: the verdict, why it landed there, the revision brief if applicable, the dissent summary, and the convergence notes that make the verdict legible.

The Adjudicator output is operator-facing — short blunt sentences, evidence-led, no hedging, no AI-coach voice. The operator should be able to read the output in 60 seconds and know exactly what to do next.

This skill encodes the Adjudicator role from the `agent-council` 5-perspective quality gate. Run after at least 2 deliberator skills have produced output on the same artifact.

When to Use / When NOT to Use

**Use this skill when:**

• You have run 2 or more deliberator skills (`skeptic-review`, `voice-identity-review`, `evidence-calibration-review`, `strategy-stakes-review`) on the same artifact and have their structured JSON outputs
• You want a SHIP / REVISE / HOLD verdict with a concrete revision brief, not another perspective
• You need a verdict document the operator can act on in 60 seconds
• You are running an interactive Council-style review (in the analytics product, Claude, Cursor) and want the same synthesis the CLI's automated Council produces

**Do NOT use this skill when:**

• You have only one deliberator critique (verdicts on single perspectives are meaningless; load at least 2 deliberators first)
• You want a new critical perspective — Adjudicator synthesiz