Seek Verdict

seek-verdict: Independent Second-Opinion Verification

When NOT to Use

• General code review (use /codex-review-fast)
• Architecture debates (use /codex-brainstorm)
• Nit findings with no dismiss intent (use [NIT_DEFERRED])

Intent x Severity

Default intent: dismiss (backward compatible with v1).

Dismiss Authorization

P0/P1 dismiss produces a candidate, not a final authorization. See Policy Mapping for human gate protocol.

3-Phase Protocol

sequenceDiagram
    participant C as Claude
    participant S as seek-verdict
    participant X as Codex (fresh thread)

    C->>S: Finding + intent + severity
    Note over S: Phase A: Candidate Packaging
    S->>S: Extract finding_packet<br/>(no Claude conclusions)
    S->>X: Phase B: Blind Independent Verdict
    Note over X: Independent research<br/>git status / grep / cat
    X-->>S: verdict + confidence + evidence
    Note over S: Phase C: Policy Mapping (intent x severity)
    alt intent=dismiss + P2/Nit
        S-->>C: DISMISS_VERIFIED / FIX_REQUIRED (automated)
    else intent=dismiss + P0/P1
        S-->>C: DISMISS_CANDIDATE + Need Human (human gate)
    else intent=confirm/clarify
        S-->>C: [SEEK_VERDICT] informational result
    end

Phase A: Candidate Packaging

Extract finding artifact from review output:

Critical: Record Claude's hypothesis locally. Never send it to Codex.

Phase B: Blind Independent Verdict

Use the prompt template in Verdict Prompt.

Phase C: Policy Mapping

Apply thresholds from Policy Mapping.

Dismiss intent: graduated thresholds by severity (P0: 0.95/4, P1: 0.90/3, P2: 0.80/2, Nit: 0.70/1).

Confirm intent: ACTIONABLE->CONFIRMED, NON_ACTIONABLE->DISPUTED, low confidence->UNCERTAIN.

Clarify intent: HIGH_IMPACT / LOW_IMPACT / UNCERTAIN.

Output audit trail per Policy Mapping.

Rebuttal

When Phase C maps the verdict to FIX_REQUIRED (raw ACTIONABLE at or above the confidence floor — FIX_REQUIRED is the mapping's word, never Codex's) and Claude has objective counter-evidence:

• 1 round max via § Resume (same verdict thread)
• Only objective artifacts (tests, specs, language semantics)
• The rebuttal names Codex's own verdict, ACTIONABLE, not the mapped result
• After rebuttal: map the new answer again — still FIX_REQUIRED -> fix; ambiguous -> NEED_HUMAN

Anti-Abuse Guard

See Policy Mapping for full rules.

• Dismiss: 3 consecutive DISMISS_VERIFIED -> [DISMISS_PATTERN_WARN] + heightened thresholds
• Confirm/Clarify: per-finding cap (1 confirm + 1 clarify per finding per commit)
• Session end or branch switch resets all counters

Output

Dismiss intent:

[DISMISS_VERDICT] key=<file|canonical_issue> | severity=<P0-Nit> | verdict=<DISMISS_VERIFIED|DISMISS_CANDIDATE|FIX_REQUIRED|NEED_HUMAN> | confidence=<0..1> | codex_thread=<id> | evidence=<brief> | timestamp=<ISO8601> | intent=dismiss | authorization=<automated|human-required|human-confirmed>

Confirm/Clarify intent:

[SEEK_VERDICT] key=<file|canonical_issue> | severity=<P0-Nit> | intent=<confirm|clarify> | verdict=<CONFIRMED|DISPUTED|HIGH_IMPACT|LOW_IMPACT|UNCERTAIN> | confidence=<0..1> | codex_thread=<id> | evidence=<brief> | timestamp=<ISO8601>

Verification

• Intent is valid (dismiss / confirm / clarify)
• Codex prompt contains no Claude conclusions (anti-anchoring)
• Fresh Codex thread used (not review session thread)
• Correct audit trail token output ([DISMISS_VERDICT] or [SEEK_VERDICT])
• P0/P1 dismiss produces DISMISS_CANDIDATE + human gate (never auto-dismiss)
• Anti-abuse streak tracked (dismiss) / per-finding cap enforced (confirm/clarify)

References