Role red team critique

Role: Red-Team Critique

Purpose

Red-Team critique provides an adversarial second opinion on baton artifacts and governance documents by routing to a cross-family fleet model. The cross-family invariant (non-Anthropic reviewer) eliminates same-family echo-chamber effects where the primary model overlooks its own blind spots.

This skill is the canonical umbrella over:

• scripts/global/fleet-red-team-dispatch.js — HAMR-wrapped Ollama dispatcher
• scripts/global/baton-fleet-review-comment.js — structured comment formatter
• scripts/global/red-team-evidence-quality.js — validator for evidence blocks
• .claude/commands/fleet-review.md — slash command entry point
• config/red-team-model-matrix.yml — A1 model-selection function

Dispatch Protocol

When to invoke

Red-Team is ADVISORY on all lanes. Invoke before posting the-collaborator-handoff on tickets labeled area:governance, area:instructions, or area:scripts. For lane:code-change touching deployed runtime artifacts, invoke before the-admin-handoff. Invocation is always optional; skipping must be noted in the-collaborator-handoff as red_team_review: skipped — <rationale>.

How to invoke

# High-stakes review (32B model, ~3-5 min)
node scripts/global/fleet-red-team-dispatch.js \
  --artifact-type collaborator-handoff \
  --ticket 2317 \
  --model qwen2.5-coder:32b

# Fast fallback (7B model, ~30-60 sec)
node scripts/global/fleet-red-team-dispatch.js \
  --artifact-type collaborator-handoff \
  --ticket 2317 \
  --model qwen2.5-coder:7b

Model selection at dispatch time uses config/red-team-model-matrix.yml via selectModel(taskContext, opts) in fleet-red-team-dispatch.js. Pass --stakes high to force 32B; omit for automatic selection per the matrix.

Cross-family invariant

The dispatched model MUST differ from the primary authoring model family. Claude Code sessions use Anthropic models; therefore all red-team dispatches MUST target Ollama-hosted models (Qwen, Llama, Mistral, etc.) or an OpenRouter non-Anthropic provider. Dispatching to another Claude model violates the cross-family rule and produces findings that must be discarded.

Enforce via config/red-team-model-matrix.yml cross_family_required: true field.

Output Format

Each red-team run produces a structured comment posted to the linked GitHub issue. The comment format is governed by scripts/global/baton-fleet-review-comment.js.

Per-finding structure

Every finding carries three required fields:

Field	Values	Description
verdict	ACCEPT / REJECT / PARTIAL	ACCEPT = valid; REJECT = false positive; PARTIAL = needs iter
text	string	One-sentence finding description
hallucination_risk	low / medium / high	Likelihood the model invented this finding

hallucination_risk: high findings MUST be discarded or independently verified before acting. The primary author classifies risk per these heuristics:

• low: finding cites a specific line, file, or label that exists in the diff
• medium: finding references a pattern or convention that is plausible
• high: finding cites an arxiv URL, a non-existent function, or implausible claim

arxiv URLs are auto-stripped by stripArxivHallucinations() in the dispatcher. Any stripped reference automatically elevates finding risk to high.

Rubric output

The comment includes a summary rubric block:

## Red-Team Rubric (iteration N)
| Goal | Score | Rationale |
|---|---|---|
| G1 Governance | N/10 | ... |
| G2 Quality   | N/10 | ... |

Rubric scores are deterministic: each goal has a binary checklist and the score is (boxes_checked / boxes_total) * 10. Do NOT inflate scores to avoid uncomfortable findings. A perfect 10/10 across all goals on a non-trivial artifact is a red flag indicating the model failed to engage adversarially.

Score-honesty discipline: if the model returns 9s across orthogonal goals for a multi-file change, treat the output as hallucination_risk: high for the rubric and re-run with a more specific prompt or escalate to the 32B model.

Iteration protocol

Red-Team converges in 2-3 iterations per the memory feedback-red-team-iteration-pattern:

1. Iteration 1: broad sweep — primary author posts findings to the issue
2. Iteration 2: primary explicitly classifies each finding as ACCEPT/REJECT/PARTIAL and re-prompts with yes/no question format for each PARTIAL item
3. Iteration 3 (only if needed): targeted re-run on unresolved PARTIAL items

Stop after 3 iterations regardless of unresolved partials. Remaining partials become recommended_follow_ups in the-collaborator-handoff.

Score-Honesty Discipline (normative)

• Rubric scores must reflect the deterministic checklist output only.
• Do not add subjective positive framing that inflates scores.
• If a goal has no applicable checklist items for this artifact type, score it N/A.
• Record boxes_checked, boxes_total, and score for audit traceability.
• A closeout citing red-team rubric MUST include the raw boxes_checked/boxes_total for at least two goals.

Cross-Runtime Loadability (AC6)

This skill must be loadable in:

• Claude Code: via CLAUDE.md include reference
• Codex AGENTS.md: via skills/role-red-team-critique/SKILL.md reference block
• Copilot Agent HQ: as a skill card with the frontmatter name field as identifier
• Antigravity SDK: via skill registry JSON entry pointing to this file

The skill body MUST NOT contain Claude-Code-specific syntax. Cross-runtime loadability is validated by the loadability smoke test in the test spec.

Must Not Do

• Do not use another Anthropic/Claude model as the red-team reviewer.
• Do not act on hallucination_risk: high findings without independent verification.
• Do not run more than 3 iterations on a single artifact.
• Do not block the baton on unresolved red-team partials — they become follow-ups.
• Do not inflate rubric scores to signal "all good" when the checklist disagrees.