brand-gen-verdict-unification

For Sage brand work in Pi, use the paste-ready prompt at docs/prompts/pi-sage-brand-gen-full-pipeline.md. Keep this link instead of copying the full prompt into skill bodies.

Brand-Gen Verdict Unification

Risk addressed: brand-gen has at least four parallel quality systems with non-fungible outputs, composed implicitly:

1. card_engine._verify_render — file-level checks (PNG magic, headline presence) → gates visual_review_status to "render_suspect".
2. DSPy BrandScorer (scoring/program.py) — rubric-based axis scoring with axis_scores, axis_rationales, decision: approved | iterate.
3. vlm_critique — VLM gate with approved | rejected decision (legacy path).
4. apply_generation_critique_policy (generation_flow.py) — policy mode (advisory vs strict), can block.
5. build_structural_auto_critic (generation_flow.py:1016) — structural pre-flight critic.
6. agent_review packet — orchestrator-level review record.

Each writes a different shape onto manifest entries. Naming is overloaded: "review" vs "critique" vs "critic" vs "feedback" are not consistently distinguished. There is no place where a downstream consumer can ask "what's the unified verdict for this version?" — they have to read 4-6 fields and apply implicit precedence.

What this skill produces

{
  "summary": {
    "verdict_emitters_total": 6,
    "manifest_fields_carrying_verdicts": 9,
    "implicit_combination_sites": 3,
    "conflicting_verdict_paths": 2,
    "naming_synonyms": ["review", "critique", "critic", "feedback", "verdict", "decision"]
  },
  "emitters": [
    {
      "emitter": "card_engine._verify_render",
      "site": "card_engine.py:1500-1540 (approx)",
      "verdict_shape": "{visual_review_status: 'pending' | 'render_suspect'}",
      "scope": "file-level (per PNG)",
      "blocking": false,
      "writes_to": ["manifest.entries[].visual_review_status"]
    },
    {
      "emitter": "BrandScorer.forward",
      "site": "scoring/program.py BrandScorer",
      "verdict_shape": "{axis_scores: {...}, axis_rationales: {...}, decision: 'approved' | 'iterate'}",
      "scope": "rubric-level (per version)",
      "blocking": false,
      "writes_to": ["review-packet.json", "manifest entry decision"]
    },
    {
      "emitter": "vlm_critique",
      "site": "vlm_critique.py",
      "verdict_shape": "{approved: bool, rationale: str}",
      "scope": "image-level (per version)",
      "blocking": false,
      "writes_to": ["iteration_memory positive/negative"]
    },
    {
      "emitter": "apply_generation_critique_policy",
      "site": "generation_flow.py apply_generation_critique_policy",
      "verdict_shape": "{policy_status: 'allow' | 'block', mode: 'advisory' | 'strict', findings: [...], bypassable: bool}",
      "scope": "policy-level (per run)",
      "blocking": true,
      "writes_to": ["pipeline_result.stop_reason", "run_ledger event"]
    },
    {
      "emitter": "build_structural_auto_critic",
      "site": "generation_flow.py:1016",
      "verdict_shape": "{P1: [...], P2: [...], P3: [...]}",
      "scope": "scratchpad-level (pre-execute)",
      "blocking": "indirect — drives iteration_memory negatives",
      "writes_to": ["iteration_memory negatives via auto_capture_generation_feedback"]
    },
    {
      "emitter": "agent_review",
      "site": "agent_review.py",
      "verdict_shape": "{orchestrator_review_packet: {...}}",
      "scope": "orchestration-level (per run)",
      "blocking": false,
      "writes_to": ["agent-review packet json"]
    }
  ],
  "implicit_combination_sites": [
    {
      "site": "manifest entry decision_field_resolution",
      "issue": "no documented rule for which verdict wins when render says render_suspect, scorer says approved, policy says allow"
    }
  ],
  "conflicting_verdict_paths": [
    {
      "scenario": "vlm_critique approved + apply_generation_critique_policy block",
      "current_behavior": "policy blocks the run; vlm verdict is recorded but does not gate",
      "documented": false
    }
  ],
  "recommended_verdict_schema": {
    "name": "Verdict",
    "fields": [
      "verdict_id",
      "version_id",
      "emitter: enum",
      "scope: enum",
      "decision: enum (approve | iterate | block | flag)",
      "blocking: bool",
      "rationale: str",
      "axis_scores: dict | null",
      "findings: list[Finding] | null",
      "supersedes: List[verdict_id]"
    ],
    "composition_rule": "blocking emitter (policy) > rubric (scorer) > image (vlm) > file (render verification) > pre-flight (auto-critic). Equal scope: most-recent supersedes older.",
    "manifest_field": "manifest.entries[].verdicts: List[Verdict]; manifest.entries[].resolved_verdict: Verdict (computed)"
  }
}