Other

/adversarial-council

Run and validate a multi-tier anonymized-peer-ranking council. Triggers: council, peer ranking, judge panel, anonymized review, adversarial verification, spec hardening, score candidates.

$ golems-cli skills install adversarial-council

Updated 1 week ago

The multi-tier anonymized-peer-ranking council — N agents each produce a candidate, then judge the anonymized set on merit, no authorship leak, ending with a sentinel line — is a KEEP-the-win discipline that decays to folklore unless it is gate-shaped. This skill codifies the protocol and ships a mechanical validator (council_lint.py) for the ballot a judge produces, with RED→GREEN fixtures (fixtures/, tests/).

The protocol (judge-on-merit)

Read the required inputs first. Acknowledge each spec/design/diff in an ## Inputs read section before any verdict — read-required-inputs-before-opining.
Anonymize. Candidates are presented without authorship; judges must not reference a participant engine (…Claude/Codex/Cursor/Gemini) or claim first-person authorship. Rank tracks quality, never identity.
Score each candidate on a comparable numeric scale (Score: N), with a merit critique (strengths + concrete risks).
End with the sentinel line (default COUNCIL-BALLOT-COMPLETE) so a truncated or streamed ballot is detectable.
Synthesize from the winner, grafting the best ideas from runners-up.

Pre-build adversarial-verification gate

Before a finding or brief is handed off, resolve every cited symbol/line against the actual code (the symbol exists, the line says what the finding claims). A council verdict that cites file:line or a function name must be checkable — pair this skill with /never-fabricate (read-before-cite) so a verdict is never folklore.

Validate a ballot

python3 council_lint_cli.py ballot.md --require specs/spec.md --require designs/design.md
python3 -m pytest tests/      # the RED→GREEN gate (10 cases)

Rule	Catches
`authorship-leak`	A candidate critique names a participant engine or claims first-person authorship (`my proposal`, `I wrote this`).
`missing-sentinel`	The ballot's final non-empty line isn't the agreed sentinel.
`unread-required-input`	A required input is not acknowledged in the `Inputs read` section.
`missing-score`	A ranked candidate has no numeric `Score:`.

The RED fixture trips all four; the GREEN fixture (merit critiques, scored, inputs read, sentinel) is clean. That delta is the gate that keeps the council from decaying.

Full SKILL.md source — includes LLM directives, anti-patterns, and technical instructions stripped from the Overview tab.

The protocol (judge-on-merit)

Read the required inputs first. Acknowledge each spec/design/diff in an ## Inputs read section before any verdict — read-required-inputs-before-opining.
Anonymize. Candidates are presented without authorship; judges must not reference a participant engine (…Claude/Codex/Cursor/Gemini) or claim first-person authorship. Rank tracks quality, never identity.
Score each candidate on a comparable numeric scale (Score: N), with a merit critique (strengths + concrete risks).
End with the sentinel line (default COUNCIL-BALLOT-COMPLETE) so a truncated or streamed ballot is detectable.
Synthesize from the winner, grafting the best ideas from runners-up.

Pre-build adversarial-verification gate

Validate a ballot

python3 council_lint_cli.py ballot.md --require specs/spec.md --require designs/design.md
python3 -m pytest tests/      # the RED→GREEN gate (10 cases)

Rule	Catches
`authorship-leak`	A candidate critique names a participant engine or claims first-person authorship (`my proposal`, `I wrote this`).
`missing-sentinel`	The ballot's final non-empty line isn't the agreed sentinel.
`unread-required-input`	A required input is not acknowledged in the `Inputs read` section.
`missing-score`	A ranked candidate has no numeric `Score:`.

The RED fixture trips all four; the GREEN fixture (merit critiques, scored, inputs read, sentinel) is clean. That delta is the gate that keeps the council from decaying.