# Coverage Pass Strategy — A/B/C Formats

Status: LOCKED (2026-04-05)
Source: JT + Claude session 2026-04-04/05

## The Three Formats

### Format A — Same Moment, Multiple Angles
**What:** Generate the same beat from 3 different camera angles (~4s each).
**Like:** Shooting a master + coverage on set. Three cameras, one take.
**Use for:** Reaction beats, dialogue, any moment where editorial choice matters.
**Output:** 3 clips of the same 4-second moment. Editor picks the best angle or cuts between them.

### Format B — Sequential Cuttable Shots
**What:** Generate the next 12-15 seconds as a multi-shot sequence with cuts.
**Like:** A pre-edited scene. The model is the cinematographer AND editor.
**Use for:** Narrative progression, scene flow, exposition.
**Output:** One video with internal cuts. The blueprint image test used this format.

### Format C — Oner (Continuous Take)
**What:** Generate one continuous 10-15 second shot. Extract edit points in post.
**Like:** Shooting a long take on a Steadicam. Find the cuts later.
**Use for:** Atmospheric moments, set pieces, climactic sequences where performance continuity matters.
**Output:** One continuous clip. Editor finds the gold inside it.

## Tier Mapping

| Tier | Score Range | Formats | Rationale |
|------|------------|---------|-----------|
| **Climax** (Tier 3) | >= 0.90 | **A + C** | Maximum editorial options. Multi-angle coverage for the key moment + a continuous performance take to mine. |
| **Rising/Peak** (Tier 2) | 0.65 – 0.89 | **A + B** | Coverage angles for the important beats + sequential flow for narrative momentum. |
| **Valley/Rising** (Tier 0-1) | < 0.65 | **B + C** | Scene progression to move the story + atmospheric oners for mood. Angles less critical. |

## Cost Model

At Kling V3 Standard ($0.084/sec):

| Tier | Formats | Duration | Takes | Cost per scene |
|------|---------|----------|-------|---------------|
| Climax | A (12s) + C (15s) | 27s | ×3 takes | ~$6.80 |
| Rising/Peak | A (12s) + B (15s) | 27s | ×2 takes | ~$4.54 |
| Valley | B (12s) + C (15s) | 27s | ×1 take | ~$2.27 |

Per episode (~5-8 shots, assume 2 scenes): ~$7-15
Per season (16 episodes): ~$110-240

## Implementation Requirements

### Multiple Takes
Peak/Climax tiers generate N variations of the same pass. N = tier-dependent:
- Valley: 1 take
- Rising/Peak: 2 takes
- Climax: 3-5 takes

### Blueprint Image Anchor
Every pass in a scene feeds from the SAME start image (character hero or scene establishing shot). Add `blueprint_image_path` field to `CoveragePass` dataclass. The blueprint is the casting hero for character scenes, or the location hero for ENV scenes.

### Scene-Level Emotional Arc
Wrap per-shot segment prompts with a scene-direction preamble:
```
[SCENE DIRECTION: This scene starts quiet and cracks open. The routine breaks.
Something is wrong and she can't name it yet.]

Shot1 2s: ...
Shot2 3s: ...
```

The preamble is derived from the episode script's action/emotion for the scene. Not per-shot — per-pass.

### Director's Wildcard
One optional segment per pass with minimal prompting. Mark it in the coverage planner:
```python
CoverageSegment(
    prompt="Something catches her attention. A shift.",
    is_wildcard=True,
)
```

### Model Routing by Tier
| Tier | Model | Mode |
|------|-------|------|
| Valley | Veo 3.1 | T2V (atmospheric, slow) |
| Rising/Peak | Kling V3 Standard | Multi-shot |
| Climax | Kling V3 Pro | Multi-shot + multiple takes |

### Motion Control Presets
Map shot types to Kling camera presets:

| Shot Type | Preset |
|-----------|--------|
| WS/EST establishing | Slow orbit or jib down |
| MS/MCU dialogue | Subtle push-in |
| OTS | Gentle track |
| INSERT | Static or slow dolly |
| Climax reveal | Whip pan or dramatic push |

### Transition Variety
Drive transitions from emotion tier:
- Valley: dissolves, slow fades
- Rising: hard cuts, match cuts
- Peak: whip pans, smash cuts
- Climax: hard cuts or continuous (Format C)

## Leaving on the Table (Priority Order)

1. **Multiple takes** — trivial to implement (loop existing call), huge editorial impact
2. **Blueprint image anchor** — one field addition to CoveragePass, high spatial consistency impact
3. **Scene-level emotional arc prompt** — prompt engineering in pass builder, high narrative momentum
4. **Director's wildcard segment** — optional slot in pass builder, medium impact (happy accidents)
5. **Coverage variation pass (shot cycling)** — moderate effort (concept exists), huge impact (real editorial coverage)
6. **Motion Control preset mapping** — config table + lookup, medium impact (professional camera feel)

## Open Questions
- How many reference image slots to use for blueprint scenes (Kling max 3 with start frame)?
- Should Format A share the SAME start frame for all 3 angles, or use angle-specific start frames?
- Can Wan 2.7 first+last frame control replace Format B entirely?
- Does Format C benefit from Motion Control presets, or should oners be fully model-directed?

## References
- Coverage planner: `recoil/pipeline/orchestrator/coverage_planner.py`
- Coverage density: `recoil/pipeline/orchestrator/coverage_density.py`
- Video-first experiments: `recoil/pipeline/docs/experiments/video-first-patterns.md`
- Blueprint test: `projects/afterimage/output/refs/characters/sadie/_exploration/blueprint_test/`