Verification

This is the phase that earns the quorum. The retrospective is unusually clear about it:

The five strongest runs all had a separate adversarial Verify phase. The mediocre runs also had named lenses — what they lacked was a phase whose only job was to try to break each finding.

Measured: 31/36 findings confirmed, with lens survival uniform at 83-89% (~86%).

The uniformity is the finding. If lens naming were doing the work, survival would vary by lens. It does not — the returns come from having a refutation step, not from perfecting the roster.

If you cut a phase, never cut this one. Cut a judge instead.

The contract: verifiers refute, they do not assess

A verifier asked "is this finding valid?" will mostly say yes. The prompt must invert the burden.

A miner claims:

  TITLE:    Token budget is not enforced on subagent spawns
  CLAIM:    Subagents inherit no budget cap, so a fan-out can exceed the run budget
  EVIDENCE: workflow.js:142 spawns agents in a loop with no budget check;
            run 2026-08-14 shows 41 agents against a stated cap of 16

YOUR JOB IS TO REFUTE THIS.

Independently re-open the cited evidence. Do not trust the miner's reading of it.

Refute if ANY of these hold:
  - The cited evidence does not exist, or does not say what is claimed
  - The evidence exists but does not support the conclusion drawn
  - A mechanism elsewhere already prevents the problem
  - The finding is real but its stated impact is materially overstated

Uphold ONLY if you re-opened the evidence yourself and it reproduces.

Default to refuted when uncertain. A finding that cannot be independently
reproduced must not reach the report as confirmed.

Return: {refuted: boolean, reason: string, whatIChecked: string, correctedImpact?: string}

Four load-bearing elements:

1. "YOUR JOB IS TO REFUTE THIS" — stated as the task, not as a perspective to consider.
2. "Independently re-open" — the verifier must go to the source, not reason about the summary. A verifier that only reads the finding is an opinion generator.
3. "Default to refuted when uncertain" — this is what makes the phase asymmetric, and the asymmetry is the whole point.
4. whatIChecked — makes a lazy verification visible. A verifier that returns upheld with an empty whatIChecked did not verify anything.

What counts as reproducing evidence

A finding whose evidence is a plausible narrative refutes. Plausibility is what the miner already supplied; the verifier's job is to test whether anything underneath it holds.

Verdict tallying

Deterministic — do it in code, not with another agent:

import { tallyVerdicts, ruleFinding } from './tools/quorum-lib.mjs';
const tally = tallyVerdicts(verdicts);          // {survives, upheld, refuted, abstained, contested}
const ruled = ruleFinding(finding, tally);      // adds CONFIRMED | PLAUSIBLE | REFUTED

Three rules, each guarding a way runs quietly go wrong:

Abstentions are not support. A verifier that errored, timed out, or returned a malformed verdict has produced no evidence for the finding. Counting it as a non-refutation manufactures confidence from a failure.

A tie fails. Two refute, two uphold means the evidence is contested. Contested evidence is reported PLAUSIBLE, never CONFIRMED. Rounding a tie toward confirmation is the single easiest way to put a wrong claim in a report.

Zero valid verdicts does not survive. Nothing verified it, so nothing confirms it.

Verdict meanings

Keeping refuted findings with their reasons is cheap and stops a requorum rediscovering the same dead end.

How many verifiers

Verification is the run's dominant cost, and it multiplies: total verifier agents = deduped findings × verifiers per finding — 30 findings at the 3-verifier tier is 90 agents. Size the tier with that product in view, before it is spent.

Perspective-diverse verification beats redundancy when a finding can be wrong in several ways. Three identical refuters catch the same class of error three times; a correctness refuter, a reproduction refuter, and an impact refuter catch three classes.

Reading survival rates

The corpus baseline is ~86%, uniform across lenses. Use it as a calibration signal:

A 100% survival rate is the one to distrust hardest. It is far more often a broken verify phase than a flawless mining phase.

Verifying a verification

For anything genuinely high-stakes, spot-check the verifiers themselves:

• Sample two upheld findings and re-check them yourself.
• Look for whatIChecked that merely restates the claim.
• Look for verdicts returned faster than the evidence could have been opened.

A verify phase that is not itself occasionally audited drifts toward rubber-stamping, which returns the run to the mediocre baseline while looking like the strong one.

Related skills

• quorum — the orchestrator
• quorum-lenses — the refuter seat, and permitting empty miner results
• quorum-judging — what happens to the surviving dossier
• quorum-degradation — verification when parallel agents are unavailable