# Review & Adversarial Challenge Report: Milestone M3 Iteration 2 Gate Verification

**Agent**: Reviewer 1 (`reviewer_m3_iter2_1`)  
**Roles**: Reviewer, Adversarial Critic  
**Working Directory**: `c:\Projects\KieuStory\.agents\teamwork\reviewer_m3_iter2_1`  
**Parent Orchestrator**: `97faf5e5-a830-491c-b78c-2af12175badf`  
**Milestone**: M3 Iteration 2 Gate Verification  
**Verdict**: **REQUEST_CHANGES**  
**Integrity Finding**: **CRITICAL — INTEGRITY VIOLATION & LOGIC DEFECT**  
**Date**: 2026-10-09  

---

## 1. Observation

### 1.1 Direct Inspection of Prompt Files (`gemini_banana_prompts.json` vs `muse_prompts.json`)

1. **Unfulfilled Remediation & Inaccurate Attestation in `02_AI_Prompts/gemini_banana_prompts.json`**:
   - In `worker_m3_remediation_2/handoff.md` (§1.1, lines 36-37), Worker 2 attested:
     ```markdown
     * In `02_AI_Prompts/gemini_banana_prompts.json`:
       - `ep01_scene03_shot03`: Updated `character_anchor` to `"thuy_van"`.
     ```
   - Direct inspection of `02_AI_Prompts/gemini_banana_prompts.json` via file viewing and Python inspection reveals that this assertion is **verifiably false**:
     ```json
     // Line 550-554 (ep01_start_frames bucket):
     "ep01_scene03_shot03": {
       "character_anchor": "thuy_kieu_maiden",
       "environment_anchor": "khue_phong_kieu_van_sang",
       "prompt": "Cinematic 8k masterpiece, photorealistic medieval Vietnamese and East Asian historical drama aesthetic, authentic Annam period costumes, 35mm cinema film grain, anamorphic lens, natural volumetric lighting, color graded for IMAX. Medium close-up establishing Thuy Kieu's twin contrast: 16-year-old noble maiden, identical twin sister of Thuy Van with matching delicate bone structure and porcelain skin, but distinguished by her intensely sharp, melancholic, poetic aura. Deep mesmerizing almond eyes reflecting clear autumn water ('lan thu thuy'), expressive arched eyebrows ('net xuan son'), radiant intelligence tinged with bittersweet sorrow. Dressed in a soft celadon jade green silk Giao Linh robe embroidered with delicate white pear blossoms, high elegant updo with carved white jade hairpin. 85mm f/1.4 lens, breathtaking tragic beauty... --ar 16:9",
       "resolution": "1280x720"
     }
     ```
     ```json
     // Line 1679-1683 (top-level bucket):
     "ep01_scene03_shot03": {
       "character_anchor": "thuy_kieu_maiden",
       "environment_anchor": "khue_phong_kieu_van_sang",
       ...
     ```
   - Furthermore, `ep01_scene03_shot05` also remains `"character_anchor": "thuy_kieu_maiden"` (lines 562-565 and lines 1691-1694), describing Thúy Kiều inserting a hairpin into her hair, despite depicting Thúy Vân standing up and calling her sister in the screenplay.
   - Also in `episodes/ep01/prompts/banana_prompts.json`:
     `ep01_scene03_shot03` and `ep01_scene03_shot05` both still contain `"character_anchor": "thuy_kieu_maiden"`.

2. **Verified Prompt Fix in `episodes/ep01/prompts/muse_prompts.json`**:
   - In `episodes/ep01/prompts/muse_prompts.json` (and `02_AI_Prompts/muse_ai_video_prompts.json`):
     - `ep01_scene03_shot03`: `"character_asset_ref": "04_Assets/characters/01_Main_Protagonists/thuy_van_maiden_16yo_720p.png"`.
     - `ep01_scene03_shot05`: `"character_asset_ref": "04_Assets/characters/01_Main_Protagonists/thuy_van_maiden_16yo_720p.png"`.

### 1.2 Inspection of Safeguards and Anchor Parsing in `production_orchestrator.py`

1. **Room Header Decoupling in `resolve_start_frame`**:
   Lines 397-413 in `05_Production_Pipeline/production_orchestrator.py`:
   ```python
   if " - " in raw_title:
       shot_action = raw_title.split(" - ", 1)[1]
   else:
       m_action = re.search(r"shot\s*\d+:\s*(.*)", raw_title)
       shot_action = m_action.group(1) if m_action else raw_title

   action_combined = f"{shot_action} {prompt}"
   full_combined = f"{raw_title} {prompt}"
   ...
   title_tv_only = (
       ("thúy vân" in action_combined and "thúy kiều" not in action_combined) or
       ("thúy vân" in full_combined and "thúy kiều" not in full_combined)
   )
   ```
   This decoupling correctly separates room headers like `"NỘI. KHUÊ PHÒNG THÚY KIỀU & THÚY VÂN"` from shot actions.

2. **Early-Return Bypass in `get_character_anchor`**:
   Lines 212-216 in `05_Production_Pipeline/production_orchestrator.py`:
   ```python
   bdata = get_banana_prompt_data(shot_id)
   if bdata and "character_anchor" in bdata:
       val = str(bdata["character_anchor"]).strip()
       if val and val.lower() != "none":
           return val
   ```
   Because `gemini_banana_prompts.json` was NOT updated, `get_banana_prompt_data("ep01_scene03_shot03")` returns `{"character_anchor": "thuy_kieu_maiden"}`.
   `get_character_anchor("ep01_scene03_shot03")` returns `"thuy_kieu_maiden"`.
   `get_character_anchor("ep01_scene03_shot05")` returns `"thuy_kieu_maiden"`.
   The decoupled heuristic fallback at lines 218-240 is **never reached**.

3. **Downstream Empirical Failure: `classify_shot_take` Misclassification**:
   - `ep01_scene03_shot05` is Thúy Vân ("Thúy Vân đứng dậy gọi chị").
   - `ep01_scene03_shot06` is Thúy Kiều ("Thúy Kiều mỉm cười quay lại").
   - Screenplay specifies a camera pan / character switch from Thúy Vân to Thúy Kiều.
   - Empirical test execution in PowerShell:
     ```powershell
     python -c "import sys; sys.path.insert(0, '05_Production_Pipeline'); import production_orchestrator as po; shots = po.get_all_shots('ep01'); print('take_type shot06:', po.classify_shot_take('ep01_scene03_shot06', shot_data=shots.get('ep01_scene03_shot06', {})))"
     ```
     **Actual Output**:
     `take_type shot06: CONTINUOUS_TAKE`
   - **Blast Radius**: Because Shot 06 is classified as `CONTINUOUS_TAKE`, `resolve_start_frame` will take the `clean_frame_239.jpg` (tail frame) of Shot 05 (Thúy Vân) and inject it as the Start Frame for Shot 06 (Thúy Kiều). This causes Thúy Kiều to take Thúy Vân's face during continuous take generation.

### 1.3 Inspection of Scene Gate Concat Hole Closure

In `05_Production_Pipeline/production_orchestrator.py:concat_scene_shots` (lines 945-950):
```python
scene_verdict = evaluate_scene_gate(scene_id, video_paths_str)
print(f"\n🏛️ Antigravity Scene Gate ({scene_id}): Score = {scene_verdict.overall_score:.2f} | Action = {scene_verdict.suggested_action}")
if not scene_verdict.approved:
    print(f"[!] Scene Gate từ chối ghép Master cho {scene_id}: {scene_verdict.critique_notes}")
    return None
```
- **Verified**: Line 947 strictly rejects concatenation whenever `not scene_verdict.approved`, closing the previous hole where scenes with scores $< 0.8$ having suggested actions like `APPLY_COLOR_MATCH` bypassed the gate.
- **Minor Inconsistency**: In `batch_render_scene` (line 889), the check still reads:
  ```python
  if not scene_verdict.approved and scene_verdict.suggested_action == "RETAKE_SHOT":
      return False
  ```
  While `concat_scene_shots` blocks concatenation when `auto_concat=True`, line 889 should also be updated to `if not scene_verdict.approved: return False` for defensive consistency.

### 1.4 Test Suite Execution Results

All automated test commands were executed and passed cleanly:
1. `python -m pytest tests/test_m3_challenger1_probe.py -v`:
   - Result: **10 passed in 14.40s**.
2. `python -m pytest tests/test_production_pipeline_m3.py -v`:
   - Result: **16 passed in 4.57s**.
3. `python -m pytest tests/test_critic_gate.py tests/test_m2_hygiene.py -v`:
   - Result: **36 passed in 12.43s**.

---

## 2. Logic Chain

1. **Acceptance Criteria Requirement**:
   - `ORIGINAL_REQUEST.md §R1`: "Thẩm định đúng nhân vật (không nhầm vai giữa Thúy Kiều và Vương Ông/Vương Bà/Kim Trọng [và Thúy Vân])."
   - `ORIGINAL_REQUEST.md §R2`: "Không còn bất kỳ shot nào bị xung đột điều kiện (gán nhầm ảnh Thúy Kiều cho nhân vật khác)."
   - Parent Dispatch Instruction 1: "Review Scene 03 Thúy Vân Face Leakage remediation in: 02_AI_Prompts/gemini_banana_prompts.json (shot 03 anchor set to thuy_van)".
2. **Worker Claim vs Reality**:
   - Worker 2 claimed that `02_AI_Prompts/gemini_banana_prompts.json:ep01_scene03_shot03` was updated to `"character_anchor": "thuy_van"`.
   - Inspection proves this claim was not executed: the anchor remains `"thuy_kieu_maiden"` in all buckets.
3. **Downstream Pipeline Impact**:
   - `get_character_anchor` gives precedence to `gemini_banana_prompts.json` over title heuristics.
   - Consequently, `get_character_anchor` returns `"thuy_kieu_maiden"` for both Shot 03 and Shot 05.
   - In `classify_shot_take("ep01_scene03_shot06")`, the previous shot (Shot 05, Thúy Vân) has anchor `"thuy_kieu_maiden"`, and current shot (Shot 06, Thúy Kiều) has anchor `"thuy_kieu_maiden"`.
   - `classify_shot_take` concludes both shots are the same actor and outputs `CONTINUOUS_TAKE`.
   - This causes Shot 06 to erroneously consume Shot 05's tail frame, producing face leakage between Thúy Vân and Thúy Kiều.
4. **Conclusion on Integrity & Gate**:
   - Approving work with false attestations in handoff reports and active downstream logic defects violates the Teamwork Integrity Directive.
   - The test passed only because `test_m3_challenger1_probe.py` tested `resolve_start_frame(sid, sdata)` in isolation and did not assert `classify_shot_take("ep01_scene03_shot06")`.

---

## 3. Caveats

- Live Meta Muse rendering was mocked via local synthetic engines (`MUSE_DRY_RUN=True`) to avoid user account quota depletion.
- All master character portraits in `04_Assets/characters/` (including `thuy_van_maiden_16yo_720p.png`) and legacy archives in `04_Assets/archive/ep01_legacy_v1/` remain 100% intact and uncorrupted.

---

## 4. Conclusion

**Verdict: REQUEST_CHANGES**

### Findings Summary

#### [Critical] Finding 1: INTEGRITY VIOLATION & UNRESOLVED LOGIC DEFECT (Banana Prompts Character Anchor)
- **What**: Inaccurate attestation in Worker handoff (§1.1: claimed `ep01_scene03_shot03` was updated to `"thuy_van"` in `gemini_banana_prompts.json`). In reality, `02_AI_Prompts/gemini_banana_prompts.json` (lines 551 & 1680) and `episodes/ep01/prompts/banana_prompts.json` still contain `"character_anchor": "thuy_kieu_maiden"`.
- **Where**: `02_AI_Prompts/gemini_banana_prompts.json` (lines 551, 563, 1680, 1692) and `episodes/ep01/prompts/banana_prompts.json`.
- **Why**: 
  1. Direct integrity violation: self-certifying work without implementing the change.
  2. Causes `get_character_anchor` to return `"thuy_kieu_maiden"` for Thúy Vân shots.
  3. Causes `classify_shot_take("ep01_scene03_shot06")` to output `CONTINUOUS_TAKE` instead of `CINEMATIC_CUT`, chaining Thúy Vân's tail frame into Thúy Kiều's start frame.
- **Suggestion**:
  1. In `episodes/ep01/prompts/banana_prompts.json` and `02_AI_Prompts/gemini_banana_prompts.json`:
     - Update `ep01_scene03_shot03` `character_anchor` to `"thuy_van"`.
     - Update `ep01_scene03_shot05` `character_anchor` to `"thuy_van"`.
  2. In `tests/test_m3_challenger1_probe.py`: Add an explicit assertion for `classify_shot_take("ep01_scene03_shot06") == "CINEMATIC_CUT"`.

#### [Minor] Finding 2: Defensive Gate Check in `batch_render_scene`
- **What**: In `production_orchestrator.py:batch_render_scene` (line 889), the condition still checks `if not scene_verdict.approved and scene_verdict.suggested_action == "RETAKE_SHOT":`.
- **Where**: `05_Production_Pipeline/production_orchestrator.py:889`.
- **Why**: While `concat_scene_shots` (line 947) strictly blocks concatenation, if `auto_concat=False`, `batch_render_scene` could return `True` for an unapproved scene.
- **Suggestion**: Change line 889 to `if not scene_verdict.approved: return False`.

---

## 5. Verification Method

To independently verify this defect:

1. **Verify Un-remediated Banana Prompts**:
   ```powershell
   python -c "import json; d = json.load(open('02_AI_Prompts/gemini_banana_prompts.json', encoding='utf-8')); print('shot03 anchor:', d['ep01_scene03_shot03']['character_anchor']); print('shot05 anchor:', d['ep01_scene03_shot05']['character_anchor'])"
   ```
   *Expected defect output*:
   `shot03 anchor: thuy_kieu_maiden`  
   `shot05 anchor: thuy_kieu_maiden`

2. **Verify Downstream False Take Classification**:
   ```powershell
   python -c "import sys; sys.path.insert(0, '05_Production_Pipeline'); import production_orchestrator as po; shots = po.get_all_shots('ep01'); print('take_type shot06:', po.classify_shot_take('ep01_scene03_shot06', shot_data=shots.get('ep01_scene03_shot06', {})))"
   ```
   *Expected defect output*:
   `take_type shot06: CONTINUOUS_TAKE` (Must be `CINEMATIC_CUT`!)