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)"
  }
}

When to use

• Before adding ANY new quality gate — verify it slots into the schema instead of adding a 7th parallel system.
• When the orchestrator's stop_reason doesn't match the user's perception of "did this pass."
• When iteration_memory learns from VLM but ignores policy findings (or vice versa).
• Before unifying agent-facing language across the brand-orchestrator / brand-critic / brand-explorer surface.

Inputs

• Repo path (default $PWD).
• Optional --include-naming — also emit a synonym normalization table for "review/critique/critic/feedback".

Procedure

1. Enumerate verdict-bearing fields on the manifest. Read brand_gen/generation_flow.py (manifest entry assembly) and card_engine.py (_verify_render). Find every field whose value is a quality decision string or struct: visual_review_status, decision, policy_status, vlm_decision, agent_review_status, etc.

1. Locate every emitter. For each manifest field, walk back to the function that computes it. Capture: emitter function file:line, output shape (what fields the emitter sets), what scope it covers (file / rubric / image / policy / scratchpad / orchestration), whether it is blocking (gates the run) or advisory.

1. Detect implicit combination. Search pipeline_runner.py, generation_flow.py for sites that read multiple verdict fields and produce a derived state (stop_reason, next_action, decision). Flag any site where the precedence is not explicit (no documented rule, no comment, no test).

1. Find conflicting paths. Pairwise check each emitter combination and look for cases where the system would emit divergent verdicts. Read tests under tests/test_critique_policy.py, tests/test_generation_flow.py to confirm whether documented behavior covers the conflict.

1. Audit naming. Grep for review, critique, critic, feedback, verdict, decision, status, judgment. Build a synonym map keyed by which module owns each term. Highlight cases where two terms refer to the same concept (e.g. agent_review_status vs vlm_decision).

1. Recommend the unified schema. Define a Verdict dataclass with explicit composition rules. Recommend the precedence order (typically: blocking-policy > rubric > image > render > pre-flight). Recommend the manifest extension (verdicts: List[Verdict] plus a computed resolved_verdict). Identify the minimum migration path that doesn't break existing tests.

Reference files

• brand_gen/card_engine.py — _verify_render, the file-level emitter.
• brand_gen/scoring/program.py, signatures.py, rubric_registry.py — the rubric emitter.
• brand_gen/vlm_critique.py — the image emitter (legacy path).
• brand_gen/generation_flow.py:1016 — build_structural_auto_critic (pre-flight emitter).
• brand_gen/agent_review.py — the orchestration packet.
• brand_gen/critique_policy.py — the policy emitter.
• tests/test_critique_policy.py — test surface that already constrains some combinations.

Don't

• Don't replace existing emitters. Wrap their outputs in Verdict and let the resolved-verdict logic compose.
• Don't overreach into iteration_memory entry shape — that's brand-gen-intent-preservation-trace's territory. Only state what the emitter writes there.
• Don't propose deleting vlm_critique even if it's deprecated — its output may still be the only image-level signal in legacy flows. Recommend deprecation path, don't shortcut it.