consensus-synthesis

name: consensus-synthesis description: Synthesize N per-validator verdicts into a single consensus verdict with confidence scoring based on agreement level. triggers: - "synthesize consensus" - "consensus verdict" - "validator agreement" - "confidence score"

Consensus Synthesis

Tabulate N per-validator verdicts into a single consensus verdict per journey, then assign a confidence tier based on the agreement ratio. This is the voting and confidence-scoring skill used by the consensus-synthesizer agent. It turns independent per-validator reports into one synthesized report without silently discarding any dissent.

The authoritative contract for synthesis states and the confidence formula is rules/consensus-engine.md §Synthesis States and §Confidence Formula. In any conflict between this SKILL.md and the rule, the rule wins.

This skill is read-only against validator evidence directories. It never writes to validator-N/ — it writes only to e2e-evidence/consensus/report.md.

Input

N validator reports, one per spawned validator:

e2e-evidence/consensus/
  validator-1/report.md   ← or verdict.md, per the validator's emission convention
  validator-2/report.md
  validator-3/report.md
  ...
  validator-N/report.md

Preconditions before running this skill:

• All N validators have completed. A missing validator-N/report.md is a blocking condition — partial synthesis is forbidden. If any validator is missing, stop and signal the coordinator; do NOT compute an agreement ratio over fewer than N inputs.
• Each report is non-empty. A zero-byte report is invalid evidence of a verdict; treat it as validator error (see rules/consensus-engine.md §Iron Rules case d).
• Each report cites evidence files in its own validator-N/ directory. The synthesizer re-reads those evidence files during disagreement analysis; reports without citations cannot be validated and must be treated as validator error.

The skill operates per-journey. For a plan with J journeys, the skill runs the tabulation → synthesis → confidence cycle J times, then aggregates an overall run verdict (see Output Schema below).

Vote Tabulation

For each journey in the plan, extract each validator's PASS/FAIL verdict and each validator's per-criterion verdicts (from the Criteria Assessment table defined in agents/verdict-writer.md). Build a tabulation matrix:

Per-journey tabulation

Per-criterion tabulation (within a journey)

When a journey has K PASS criteria, build a K-row sub-table so the disagreement protocol can isolate the diverging criterion rather than re-running the whole journey:

The per-criterion view is what feeds skills/consensus-disagreement-analysis when the journey-level state is not UNANIMOUS.

Synthesis Rules

Apply these rules per journey. See rules/consensus-engine.md §Synthesis States for the authoritative table.

Never shortcut the majority path. A MAJORITY_PASS verdict is not safe to emit until the disagreement protocol has looked at the dissenting validator's evidence and either validated the majority's position or flipped it. Promoting a MAJORITY to a final PASS/FAIL without analysis is equivalent to silently dropping the minority — see Anti-Patterns below.

Overall run verdict. After each journey has a synthesized verdict, the overall run verdict is the weakest journey verdict (weakest-link rule):

• If every journey is UNANIMOUS_PASS → overall PASS (HIGH)
• If any journey is DISAGREEMENT_UNRESOLVED → overall DISAGREEMENT_UNRESOLVED (LOW)
• Otherwise the weakest per-journey confidence sets the overall confidence tier

Confidence Formula

Confidence is a direct function of the agreement ratio. See rules/consensus-engine.md §Confidence Formula.

majority_count  = max(pass_count, fail_count)
agreement_ratio = majority_count / total_validators

confidence = HIGH    if agreement_ratio == 1.0               (unanimous)
confidence = MEDIUM  if 0.67 <= agreement_ratio < 1.0        (supermajority, after analysis resolves)
confidence = LOW     if agreement_ratio < 0.67               (split, unresolved)

Worked examples (total_validators = 3):

Worked examples (total_validators = 5):

Confidence degrades monotonically. Evidence quality cannot substitute for agreement; no amount of compelling screenshots moves a MEDIUM verdict to HIGH. HIGH requires unanimity. A re-run that flips a validator's vote does not retroactively upgrade confidence — only the final tuple counts. If a re-run is needed, record it as a separate synthesis pass with its own tuple and timestamp.

Edge case: total_validators < 2. The consensus engine requires ≥2 validators by contract. If only one validator's report is present, abort synthesis and return to the coordinator with a CONSENSUS_ABORTED_INSUFFICIENT_VALIDATORS signal. Do NOT emit a HIGH confidence verdict from a single validator — that is single-validator mode, not consensus, and misusing this skill that way defeats its purpose.

Anti-Patterns

Treat every item below as a blocking violation. If you are tempted by any of these, stop and re-read rules/consensus-engine.md.

Output Schema

The synthesizer emits e2e-evidence/consensus/report.md matching templates/consensus-report.md. At minimum, the report contains the following sections for each journey:

## Journey: {NAME}

**Synthesis State:** UNANIMOUS_PASS | UNANIMOUS_FAIL | MAJORITY_PASS | MAJORITY_FAIL | SPLIT
**Final Verdict:** PASS | FAIL | DISAGREEMENT_UNRESOLVED
**Confidence:** HIGH | MEDIUM | LOW
**agreement_ratio:** 0.00 – 1.00
**Validators:** N

### Vote Tabulation

| Validator | Verdict | Evidence Directory |
|-----------|---------|--------------------|
| Validator-1 | PASS | `e2e-evidence/consensus/validator-1/` |
| Validator-2 | PASS | `e2e-evidence/consensus/validator-2/` |
| Validator-3 | FAIL | `e2e-evidence/consensus/validator-3/` |

### Per-Criterion Tabulation

| # | Criterion | V1 | V2 | V3 | Agreement |
|---|-----------|----|----|----|-----------|
| 1 | {criterion} | PASS | PASS | PASS | UNANIMOUS |
| 2 | {criterion} | PASS | PASS | FAIL | MAJORITY |

### Dissenting Opinions

{For each dissenting validator: which criterion, what evidence they cited, what their
reasoning was. Never drop this section — if there is no dissent, write "None (UNANIMOUS)".}

### Disagreement Analysis (non-UNANIMOUS only)

{Summary of `skills/consensus-disagreement-analysis` output: diverging criterion,
sequential-analysis conclusion, resolution taken. Cite
`e2e-evidence/consensus/disagreement-analysis/`.}

### Final Verdict Reasoning

{Why the final verdict follows from the tuple and the synthesis state. Cite specific
evidence files from the majority validators' subdirectories.}

The report closes with:

## Overall Run Verdict

**Verdict:** PASS | FAIL | DISAGREEMENT_UNRESOLVED
**Confidence:** HIGH | MEDIUM | LOW
**Journeys:** J total; {counts per state}
**Weakest-link journey:** {name} ({synthesis state})

The one-line stdout summary, for pipeline consumers:

ValidationForge CONSENSUS: J/K journeys PASS. Overall: {verdict} ({tier}). Report: e2e-evidence/consensus/report.md

Integration

Iron Rules (carry-over)

1. No silent dissent. Every minority verdict is recorded, analyzed, and either resolved or escalated.
2. No partial synthesis. All N validator reports must be present and non-empty before this skill runs.
3. No writing to validator-N/. Synthesis is strictly read-only against validator subdirectories.
4. No inconclusive verdicts. Use DISAGREEMENT_UNRESOLVED.
5. No confidence inflation. HIGH requires unanimity; no amount of evidence quality moves MEDIUM to HIGH.
6. No synthesis from a single validator. Consensus requires ≥2 inputs; single input aborts with CONSENSUS_ABORTED_INSUFFICIENT_VALIDATORS.
7. Every PASS/FAIL in the synthesized report cites evidence from at least one validator's directory. The synthesizer does not invent citations.