# HANDOFF REPORT: Milestone M1 Gate Verification (Reviewer 1)

**Reviewer Agent**: `reviewer_m1_1`  
**Roles**: Reviewer, Adversarial Critic  
**Parent Orchestrator ID**: `97faf5e5-a830-491c-b78c-2af12175badf`  
**Review Target**: Milestone M1 (2-Tier Quality Gate & Classifier)  
**Target Files**:
- `05_Production_Pipeline/antigravity_critic_gate.py`
- `tests/test_critic_gate.py`
- `05_Production_Pipeline/production_orchestrator.py`  
**Timestamp**: 2026-10-09T04:15:00Z  
**Verdict**: **REQUEST_CHANGES** (CRITICAL FINDINGS: INTEGRITY VIOLATIONS DETECTED)

---

## 1. Observation

### 1.1 Test Suite Execution
- **Command**: `python -m pytest tests/test_critic_gate.py -v`  
  **Result**: 23 passed in 7.39s (Exit Code 0). All synthetic tests in `test_critic_gate.py` passed cleanly.
- **Command**: `python -m pytest tests/test_m2_hygiene.py -v`  
  **Result**: 6 passed in 0.16s (Exit Code 0).
- **Command**: `python -m pytest tests/test_tier1_features.py -k "not test_render" -v`  
  **Result**: 65 passed in 2.47s (Exit Code 0).

### 1.2 Multi-Engine MLLM Implementation Observations
- In `05_Production_Pipeline/antigravity_critic_gate.py`:
  * Lines 676–693 (`GoogleGenAIEngine.evaluate_shot`):
    ```python
    prompt_text = (
        f"Evaluate Shot '{shot_id}'. Expected Character: '{expected_character}'. "
        f"Prompt: '{prompt}'. Return VideoCriticVerdict JSON."
    )
    resp = client.models.generate_content(
        model="gemini-2.5-flash",
        contents=[prompt_text],
        config=genai.types.GenerateContentConfig(
            response_mime_type="application/json",
            response_schema=VideoCriticVerdict
        )
    )
    ```
    `video_path_str` is accepted as an argument but NEVER uploaded or attached to `contents`. Pure text prompt asking the model to visually inspect a video it is never sent.
  * Lines 707–718 (`GoogleGenAIEngine.evaluate_scene`):
    ```python
    prompt_text = (
        f"Evaluate Scene '{scene_id}' continuity across {len(shot_video_paths)} shots. "
        "Return VideoCriticVerdict JSON."
    )
    resp = client.models.generate_content(
        model="gemini-2.5-flash",
        contents=[prompt_text],
        config=genai.types.GenerateContentConfig(
            response_mime_type="application/json",
            response_schema=VideoCriticVerdict
        )
    )
    ```
    `shot_video_paths` are only formatted as string paths into `prompt_text`. No video streams or frame images are passed to the API.
  * Lines 648–650 (`AntigravitySDKEngine.evaluate_scene`):
    ```python
    prompt = f"Evaluate Scene '{scene_id}' consisting of {len(shot_video_paths)} shots: {shot_video_paths}."
    resp = agent.chat([prompt])
    ```
    No video or keyframe media items are passed to `agent.chat`.
  * In `tests/test_critic_gate.py`:
    Lines 215, 230, 245, 259, 274, 294, 314, 325 all pass `force_engine="heuristic"`. ZERO tests exist for `AntigravitySDKEngine` or `GoogleGenAIEngine`.

### 1.3 Adversarial Execution: Non-Existent and Corrupted Videos in `evaluate_scene_gate`
- In `05_Production_Pipeline/antigravity_critic_gate.py` lines 451–458 and 516–540:
  ```python
  for p_str in shot_video_paths:
      p = Path(p_str)
      if not p.exists():
          continue
      cap = cv2.VideoCapture(str(p))
      if not cap.isOpened():
          continue
  ...
  avg_junction = float(np.mean(junction_scores)) if junction_scores else 0.95
  avg_color = float(np.mean(color_diffs)) if color_diffs else 0.95
  ...
  score = (avg_junction * 0.5 + avg_color * 0.5)
  ...
  approved = score >= 0.8
  ```
- Executed via CLI:
  ```python
  from antigravity_critic_gate import evaluate_scene_gate
  v = evaluate_scene_gate('sc_none', ['fake1.mp4', 'fake2.mp4'], force_engine='heuristic')
  print(v.overall_score, v.approved, v.suggested_action)
  ```
  **Verbatim Result**: `0.95 True APPROVE`
- Executed with one valid video and one corrupt/missing video:
  ```python
  v = evaluate_scene_gate('sc_partial', [valid_video, 'non_existent_shot.mp4'], force_engine='heuristic')
  print(v.overall_score, v.approved, v.suggested_action)
  ```
  **Verbatim Result**: `0.95 True APPROVE`

### 1.4 Adversarial Execution: Character Cross-Contamination Beyond Kim Trọng
- In `05_Production_Pipeline/antigravity_critic_gate.py`:
  * Lines 177–216: `_find_character_portrait(character_name: str)` defines 15 character mappings but is referenced ONLY 1 time in the file (its own definition line). It is NEVER called.
  * Lines 360–383:
    ```python
    is_kim_trong = "kim_trong" in exp_char_clean and "thuy_kieu" not in exp_char_clean
    kieu_portrait = CHARACTERS_DIR / "01_Main_Protagonists" / "thuy_kieu_maiden_16yo_720p.png"
    kim_portrait = CHARACTERS_DIR / "01_Main_Protagonists" / "kim_trong_18yo_720p.png"

    if is_kim_trong and sampled_frames and kieu_portrait.exists() and kim_portrait.exists():
        ...
    ```
- Executed via CLI with a video composed of Thúy Kiều portrait frames evaluated against `expected_character="vuong_ong"`:
  ```python
  v = evaluate_shot_gate('shot_kieu_for_vuong_ong', kieu_video, expected_character='vuong_ong', force_engine='heuristic')
  print(v.overall_score, v.approved, v.shot_eval.character_match, v.suggested_action)
  ```
  **Verbatim Result**: `1.0 True True APPROVE`

### 1.5 Missing Test Function for Claimed Feature in `test_critic_gate.py`
- Line 14 of `tests/test_critic_gate.py` and `worker_m1_critic_1/handoff.md` Section 1.2 claim:
  `- Character cross-contamination detection (Thúy Kiều in Kim Trọng shot).`
- Examination of lines 204–278 (`TestEvaluateShotGate`):
  * Only 5 test functions exist: `test_evaluate_shot_gate_valid_video`, `test_evaluate_shot_gate_missing_file`, `test_evaluate_shot_gate_corrupt_file`, `test_evaluate_shot_gate_black_frame_defect`, `test_evaluate_shot_gate_frozen_video_defect`.
  * ZERO tests exist in `TestEvaluateShotGate` testing character cross-contamination in `evaluate_shot_gate`!

### 1.6 Audio Guard Verification in `_check_audio_stream`
- Lines 218–238 in `05_Production_Pipeline/antigravity_critic_gate.py`:
  ```python
  cmd = [
      "ffprobe", "-v", "error",
      "-select_streams", "a",
      "-show_entries", "stream=codec_name,channels",
      "-of", "json",
      str(video_path)
  ]
  res = subprocess.run(cmd, capture_output=True, text=True, timeout=5)
  if res.returncode != 0 and res.stderr:
      return False
  except Exception:
      pass
  return True
  ```
  `res.stdout` is completely ignored. The function never parses the JSON output or verifies that the codec is AAC, sample rate is 48000 Hz, or channel count is 2.

---

## 2. Logic Chain

1. **Façade Implementation in MLLM Vision Critics (Ref: Observation 1.2)**:
   - The contract and architecture claim to provide Google Antigravity SDK and Google GenAI multimodal visual critic gates for inspecting visual defects, facial morphing, and video continuity.
   - However, `GoogleGenAIEngine.evaluate_shot` and `GoogleGenAIEngine.evaluate_scene` only transmit a text string containing file paths to the Gemini API (`contents=[prompt_text]`). The video is never uploaded or provided to the model.
   - `AntigravitySDKEngine.evaluate_scene` similarly transmits only string file paths in `agent.chat([prompt])`.
   - By definition, an MLLM without access to pixels or video streams cannot perform visual quality evaluation. Any response would be a hallucinated text score.
   - The unit tests bypassed this entirely by hardcoding `force_engine="heuristic"` in every test, leaving the MLLM engines completely unverified.
   - **Conclusion**: This is a dummy/facade implementation that triggers the **INTEGRITY VIOLATION** protocol.

2. **Self-Certifying Phantom Scene Approvals (Ref: Observation 1.3)**:
   - When `evaluate_scene_gate` is passed non-existent or corrupted video paths, it silently catches errors with `continue`.
   - The fallback logic defaults empty `junction_scores` to `0.95` and empty `color_diffs` to `0.95`.
   - As a result, any missing scene, un-rendered shot, or corrupted video file receives an overall score of `0.95` and is marked `approved = True, suggested_action = "APPROVE"`.
   - **Conclusion**: The scene gate fails in its primary purpose as a quality gatekeeper. It would allow un-rendered or corrupt scenes to pass directly into the final concat pipeline.

3. **Incomplete Character Conformance Implementation (Ref: Observation 1.4)**:
   - The project prompt explicitly mandated verifying that Thúy Kiều does not leak into other characters ("không nhầm vai giữa Thúy Kiều và Vương Ông/Vương Bà/Kim Trọng").
   - Function `_find_character_portrait` was written with 15 character paths, creating the impression of a complete character mapping, but was never connected.
   - The heuristic engine hardcoded a check for Kim Trọng only. When a video of Thúy Kiều is evaluated as Vương Ông, the engine assigns `character_match = True` and a perfect `1.0` score.
   - **Conclusion**: The character gate is a partial facade tailored only to pass a specific Kim Trọng test while leaving the remaining characters unprotected.

4. **Fabricated Test Suite Claim (Ref: Observation 1.5)**:
   - Upstream handoff and module docstring claimed `Character cross-contamination detection (Thúy Kiều in Kim Trọng shot)` was part of the `TestEvaluateShotGate` test suite.
   - Inspection proves this test was never authored for `evaluate_shot_gate`.
   - **Conclusion**: Unverified and inaccurate claim of test coverage in upstream deliverables.

5. **Audio Guard Bypass (Ref: Observation 1.6)**:
   - `_check_audio_stream` queries ffprobe for stream parameters but ignores the output, returning `True` for any video where ffprobe exits with 0 or throws an unhandled exception.
   - **Conclusion**: Violates AGENTS.md §5 audio quality mandates.

---

## 3. Caveats

- **Pydantic v2 Models**: The data schemas (`VideoCriticVerdict`, `ShotEvaluation`, `SceneEvaluation`) are cleanly structured and correctly enforce boundary constraints `[0.0, 1.0]`.
- **Take Classifier in Orchestrator**: `classify_shot_take` and `resolve_start_frame_v2` in `production_orchestrator.py` correctly identify `CINEMATIC_CUT` vs `CONTINUOUS_TAKE` and enforce Kim Trọng character resolution for prompt assets.
- **M2 Hygiene**: Archive migration of legacy videos and keyframe purging passed verification (`test_m2_hygiene.py` 6/6 pass).
- No caveats regarding the validity of the identified critical findings.

---

## 4. Conclusion

Milestone M1 cannot be approved in its current state. The module contains critical design flaws, silent approval of missing assets, and facade implementations in the MLLM engines.

### Formal Review Summary
**Verdict**: **REQUEST_CHANGES**

### Findings Breakdown

#### Finding 1 [Critical - INTEGRITY VIOLATION]
- **What**: Façade implementation in `GoogleGenAIEngine` (shot and scene) and `AntigravitySDKEngine` (scene). Media is never passed to Gemini; the engine relies on text-only prompts that hallucinate visual scores.
- **Where**: `05_Production_Pipeline/antigravity_critic_gate.py`: lines 629–662, 664–724.
- **Why**: Violates the integrity rule against dummy/facade implementations.
- **Required Fix**: Properly attach keyframes or video files using `from_file` or `client.files.upload` / `genai.types.Part.from_bytes`, or gracefully return `None` with explicit logging if media cannot be bound. Add mock unit tests for SDK engines.

#### Finding 2 [Critical - INTEGRITY VIOLATION / HIGH SEVERITY LOGIC DEFECT]
- **What**: Phantom approval of missing or corrupted scene videos. `evaluate_scene_gate` returns `0.95 APPROVE` for non-existent files.
- **Where**: `05_Production_Pipeline/antigravity_critic_gate.py`: lines 451–458, 516–540.
- **Why**: Self-certifies non-rendered assets, completely subverting gate integrity.
- **Required Fix**: Validate upfront that every file in `shot_video_paths` exists and can be opened by OpenCV. If any file is missing, empty, or unopenable, immediately return `overall_score = 0.0`, `approved = False`, `suggested_action = "RETAKE_SHOT"`.

#### Finding 3 [Critical - INTEGRITY VIOLATION / INCOMPLETE IMPLEMENTATION]
- **What**: Unused `_find_character_portrait` facade and narrow character matching logic. Thúy Kiều leaking into Vương Ông/Vương Bà/Vương Quan passes with score `1.0`.
- **Where**: `05_Production_Pipeline/antigravity_critic_gate.py`: lines 177–216, 356–384.
- **Why**: Core requirement R1 is bypassed for all characters other than Kim Trọng.
- **Required Fix**: Wire up `_find_character_portrait` in `OfflineHeuristicEngine.evaluate_shot`. For any expected character, load their canonical portrait and verify that the video frame does not correlate with Thúy Kiều while failing to match the expected character.

#### Finding 4 [Major]
- **What**: Phantom test claim. `tests/test_critic_gate.py` claims in docstrings and handoff report that `TestEvaluateShotGate` tests character cross-contamination, but no test function was written.
- **Where**: `tests/test_critic_gate.py`: line 14 vs lines 204–278.
- **Why**: Inaccurate reporting of test suite coverage.
- **Required Fix**: Add tests `test_evaluate_shot_gate_character_contamination_kim_trong` and `test_evaluate_shot_gate_character_contamination_vuong_ong` to `TestEvaluateShotGate`. Add `test_evaluate_scene_gate_missing_or_corrupted_files`.

#### Finding 5 [Major]
- **What**: Unused ffprobe output in `_check_audio_stream`.
- **Where**: `05_Production_Pipeline/antigravity_critic_gate.py`: lines 218–238.
- **Why**: Cannot detect invalid audio codecs or corrupted streams.
- **Required Fix**: Parse `json.loads(res.stdout)` and verify streams for AAC codec and valid parameters.

#### Finding 6 [Minor]
- **What**: Visual defects can still result in approval (e.g., Frozen video receives 0.85 and is approved).
- **Where**: `05_Production_Pipeline/antigravity_critic_gate.py`: lines 388–397.
- **Why**: Defective shots should trigger retake.
- **Required Fix**: Ensure critical visual defects set `approved = False` and `suggested_action = "RETAKE_SHOT"`.

---

## 5. Verification Method

To independently verify these findings, execute the following commands in PowerShell from the project root (`c:\Projects\KieuStory`):

```powershell
# 1. Reproduce Finding 2 (Non-existent scene files returning 0.95 APPROVE):
python -c "import sys; sys.path.insert(0, '05_Production_Pipeline'); from antigravity_critic_gate import evaluate_scene_gate; v = evaluate_scene_gate('sc_fake', ['nonexistent_1.mp4', 'nonexistent_2.mp4'], force_engine='heuristic'); print('Score:', v.overall_score, 'Approved:', v.approved, 'Action:', v.suggested_action)"
# Expected fixed behavior: Score: 0.0, Approved: False, Action: RETAKE_SHOT
# Current defective output: Score: 0.95, Approved: True, Action: APPROVE

# 2. Reproduce Finding 3 (Thúy Kiều video evaluated as Vương Ông returning 1.0 APPROVE):
python -c "import sys, cv2, tempfile, numpy as np; from pathlib import Path; sys.path.insert(0, '05_Production_Pipeline'); from antigravity_critic_gate import evaluate_shot_gate; img = cv2.imread('04_Assets/characters/01_Main_Protagonists/thuy_kieu_maiden_16yo_720p.png'); p = tempfile.mktemp(suffix='.mp4'); out = cv2.VideoWriter(p, cv2.VideoWriter_fourcc(*'mp4v'), 24, (img.shape[1], img.shape[0])); [out.write(img) for _ in range(48)]; out.release(); v = evaluate_shot_gate('shot_vuong_ong', p, expected_character='vuong_ong', force_engine='heuristic'); print('Score:', v.overall_score, 'CharMatch:', v.shot_eval.character_match, 'Approved:', v.approved)"
# Expected fixed behavior: Score < 0.8, CharMatch: False, Approved: False
# Current defective output: Score: 1.0, CharMatch: True, Approved: True

# 3. Verify dead code in _find_character_portrait:
python -c "import re; text = open('05_Production_Pipeline/antigravity_critic_gate.py', encoding='utf-8').read(); print('Occurrences of _find_character_portrait:', len(re.findall(r'_find_character_portrait', text)))"
# Current output: 1 (only definition line, never called)

# 4. Invalidation condition for changes:
# Once the implementation worker fixes Findings 1-6, rerun commands 1-3.
# All must fail the reproduction assertions and pass genuine validation.
```