adversarial-verify

name	adversarial-verify
description	The ensemble verification harness: split a conclusion, fix, or design into checkable claims, classify each to a pre-written probe row, spawn N independent voters (N from the tier: 1/3/5), and aggregate mechanically. Trigger: a conclusion/fix/design is about to ship, a core path changed, or a pipeline step names an adversarial pass. Manual invocation: /adversarial-verify.

adversarial-verify — the ensemble harness

This skill RUNS the harness; it performs no verification itself. Voters execute pre-written probes — never invented ones — and a script, not a model, merges the votes. Mechanics: docs/architecture/mechanisms/M2-ensembles.md (votes) and M8-refute-by-probe.md (probes).

Step 1 — claims checklist

Break the subject into individually checkable claims with file/line anchors: "the cache at foo.py:42 isn't taking effect", "this module has no callers". A vague "it looks right overall" cannot be verified — split it until each claim names one artifact and one expected fact.

Step 2 — classify each claim to a probe row (lookup, not judgment)

Load references/probe-tables.md — one file, the whole catalog. Match each claim to a row by the row's match pattern. A claim matching no row → STOP for that claim and escalate per docs/architecture/escalation-ladder.md row 11 (probe-library gap → R2: a pinned model authors the probe once). Never improvise a probe inline.

Step 3 — read N from the tier, spawn N voters

N = prescription.votes in .maestro/evidence/<diffhash>/sizing.json (T1=1, T2=3, T3=5). Artifact missing → run bash ${CLAUDE_PLUGIN_ROOT}/hooks/lib/sizing.sh first. Voting is not optional and N is never chosen by feel.

Each voter is a fresh agent (Task tool, model: haiku) that receives:

A voter receives	A voter never receives
the identical rubric: claims + probe rows + the verdict schema	another vote's verdict or transcript
the LITERAL artifact under test — never a paraphrase	your working context (fresh eyes)
read access to .maestro/evidence/<diffhash>/	the aggregate of any prior round

Voter prompt (fixed — fill only the slots):

You verify; you never fix. Default every claim to `failed`.
For EACH claim, run its probe EXACTLY as written in its row. Capture the real
command line, output excerpt, and exit code into the evidence field.
A claim resting on a runnable artifact (a query, command, config, or snippet in a
doc) is verified by running the LITERAL shipped text — extract it from the file;
a re-typed "equivalent" silently repairs the defect you were sent to find.
Flip failed → met ONLY with the probe output attached. A probe you cannot run =
failed, with the error as the output excerpt.
Write your verdict to .maestro/evidence/<DIFFHASH>/votes/vote-<I>.json and output it.
CLAIMS + PROBE ROWS: <paste>
SCHEMA: <the verdict schema from references/probe-tables.md>

Step 4 — aggregate mechanically (no model in the loop)

bash ${CLAUDE_PLUGIN_ROOT}/hooks/lib/aggregate_votes.sh <diffhash>

The aggregator computes per-criterion strict majority, applies the evidenced-failed veto (one failed with attached probe output overrides any majority met), and writes .maestro/evidence/<diffhash>/aggregate.json with a split field.

Aggregate outcome	What you do
every criterion met, no split	claims are deliverable; the verdict stays in the bundle
split at T2 (2–1 on any criterion)	escalate R1: re-run with 5 voters (ladder row 2)
split at T3 (≥2 of 5 dissent)	escalate R2: one pinned sonnet-class pass supersedes (ladder row 1)
any evidenced failed	that claim is overturned — DROP it; never patch it up
more claims overturned than survived	the frame is wrong; rethink, don't patch claim by claim

Step 5 — record

The gates read the bundle, so after aggregation record the ledger line:

bash ${CLAUDE_PLUGIN_ROOT}/hooks/lib/ledger.sh mark adversarial pass <ref> "<claims that survived>"

votes/vote-<i>.json + aggregate.json stay in the bundle; the commit gate validates their existence, schema, and vote count against sizing.json.

What changed from the frontier version

• "Verify from multiple perspectives (if it matters)" is deleted: N always comes from the tier. Voting is the default, not a stakes call.
• "Assume 60% wrong and go invent a counterexample" is deleted: voters execute probe rows. Refute-first survives as the default-failed rule plus the evidenced veto.