Judging and synthesis

By this point the dossier is confirmed. Judging is not "check it again" — it is deciding what to do about it, under the constraints that actually govern.

And synthesis carries the single most important rule in the pack:

The only family-invariant across ~21 runs is an explicit dissent-honoring convergence step. It survives every specimen, including judge-less and verify-less runs. Everything else — phase count, verify presence, judge count (0 to 3 observed), evidence mechanism — varied.

Judges are constraints, not lenses

A lens asks how could this fail? A judge asks would this pass the thing that actually kills proposals like this?

Name the 2-4 hard constraints that would kill or approve this class of decision, from the subject's real stakes.

If a proposed judge would have made a good miner, it is a lens. Move it.

Zero judges is a real answer

The corpus observed 0 to 3 judges. Use zero when nothing is contested:

• Generation with checkable ground truth — citations resolve, tests pass. Mine → Verify, then synthesize. A panel over an uncontested check adds cost and no information.
• A quick review where the confirmed dossier already implies its own priority order.

Adding a panel because a quorum "should have one" is exactly the invented-structure failure the corpus rejected elsewhere.

What judges receive, and may do

Judges get the confirmed dossier as data — JSON, findings with verdicts, evidence, corroboration counts.

• They may not re-mine. A judge that goes hunting produces unverified findings that bypass the verify phase entirely.
• They may not launch tools, with one exception: a targeted re-read of a single cited artifact.
• They rule on what is in front of them.

Each judge returns:

{
  judge: "token economist",
  recommendations: [{title, gain, effort, rationale}],
  rejected:        [{title, reason}],
  novelInsight:    "one thing nobody asked about"
}

The rejected array matters as much as the recommendations — it is what produces a genuine dispute rather than silent divergence, and disputes are what the synthesis must rule on.

novelInsight is where a panel repeatedly earns its cost: the judge sees the whole dossier at once, which no miner did.

Consensus, computed not narrated

import { judgeConsensus } from './tools/quorum-lib.mjs';
const { consensus, singleJudge, disputed } = judgeConsensus(judgeOutputs);

Matching is fuzzy by design: judges phrase the same recommendation differently, and exact matching would report three single-judge picks where there was one three-judge consensus — inverting the point of the panel.

Dissent is the deliverable

The rule that survived every run:

A disagreement is reported with an explicit ruling. It is never silently dropped.

Dropping dissent is tempting because it produces a cleaner document. It also destroys the only information the panel uniquely generated — that informed perspectives disagree, and on what.

Ruling on a dispute means stating:

1. What each side held, in one line each
2. Which position governs here, and why — usually because one judge's constraint is binding in this context and the other's is not
3. What would change the ruling — the condition under which the other side wins

### Disputed: delete the unused MCP servers

- **Token economist (for):** 2 servers, 0 calls in 34 sessions; their tool
  definitions are resident on every request.
- **Reliability judge (against):** one is the emergency rollback path; removing
  it trades a recurring small cost for a rare large one.

**Ruling:** disconnect the analytics server, keep the rollback server. The
economist's argument is binding for the server with no failure role; the
reliability constraint governs the other.

**Would change this:** if the rollback path gets a non-MCP equivalent, the
second server goes too.

The synthesis output

Fixed section order. Do not improvise a variant:

1. Root causes — ranked, quantified. Not a restatement of findings; the why underneath them
2. Top actions — a table with gain and effort, consensus first
3. Stop-doing list — what to cease. Consistently the most-acted-on section, and the most often omitted
4. Disagreements and rulings — per above

Then, briefly: what was not covered, and what remains unverified. A synthesis claiming total coverage is asserting something the run did not establish.

Quality checks on a ruling

• Every quantitative claim traces to a CONFIRMED finding
• PLAUSIBLE findings are labelled as unverified where used
• Every dispute has a ruling and a would-change-this condition
• The stop-doing list is present and non-empty (if genuinely empty, say so explicitly)
• Actions are specific enough to start today — the corpus's strong runs shipped 6 of 10 recommendations in the same session; vague actions are the tell of a thin run
• The novel insights from judges appear somewhere rather than being discarded

Related skills

• quorum — the orchestrator and the output contract
• quorum-verification — what produced the confirmed dossier
• quorum-lenses — why judges are not lenses
• decision-quorum — the adapter where judging carries the most weight