# MotionLab Rounds 2–100 — Diagnosis

## What I can and cannot do here

I am a coding sub-agent with **no filesystem, no shell, no tool access**. I cannot run `run_round.py`, `motion_score.py`, or `delegate_agent`. So I cannot literally execute rounds 2–100. What I *can* do is deliver the exact, ready-to-run plan: the prompt texts, the round schedule, the seed discipline, the logging format, and the decision rules for hill-climbing — so the orchestrator executes without further design work.

**Missing context I would need to actually run it:** the clip URLs and motion_score JSON for rounds 2–100, plus the sub-agent 0–10 scores per clip. Only round 1's numbers are in the bundle.

---

## Diagnosis of the round-1 baseline (the thing to beat)

Round 1 is a **descriptive caption**, not a motion instruction. That is the whole finding:

- `motion_energy 854`, `active_area_mean 0.121`, `col_shift_px 0`, `static_fraction 0.007`.
- Sub-agent: subject 2, env 4, cam 1, coherence 7, yield 3.
- The only sustained motion is **fog drift** — i.e. the model animated the one element whose *noun* implies motion ("fog drifts"). The monolith, described with stative verbs ("stands", "glowing"), stayed still. Camera got nothing because nothing in the prompt asked for it.

**Working hypothesis for rounds 2–100:** i2v models under a turbo LoRA at 8 steps respond to *explicit, verb-first, per-layer motion assignment* and to *camera language*, and are largely deaf to adjectives and to negation. The lever is **grammatical mood and layer decomposition**, not vocabulary richness. Round 1's `col_shift_px 0` is the cleanest signal available: zero camera motion was requested, zero was produced.

---

## Round schedule (2–100)

### Phase 1 — broad sweep, rounds 2–10 (one per family, seed 3617007189758648082 for all, matching the base frame's seed lineage so cross-family comparison is prompt-effect only)

| R | Scheme | Family | Prompt file | Core construction |
|---|---|---|---|---|
| 2 | B1 | B explicit verbs | `prompts/B1_verbs.txt` | "The blue fissure pulses and flares brighter, then dims. Loose rocks tumble down the slope and settle. Fog surges across the canyon floor. The monolith cracks and shifts." |
| 3 | C1 | C camera | `prompts/C1_dolly.txt` | "The camera slowly dollies in toward the monolith, pushing forward through drifting fog. Parallax: foreground rocks slide past frame left as distant cliffs and the suspension bridge grow closer." |
| 4 | D1 | D layered | `prompts/D1_layers.txt` | "Foreground: loose rocks tumble and roll toward camera. Midground: the monolith's fissure pulses and flares, stone fragments break away. Background: fog banks surge between the cliffs, the suspension bridge sways." |
| 5 | E1 | E sequencing | `prompts/E1_beats.txt` | "First the fissure ignites and flares. Then the monolith cracks and stone shards fall. Then fog surges over the rocks. Finally the camera pushes in as debris settles." |
| 6 | F1 | F intensity | `prompts/F1_intensity.txt` | "VIOLENT, CONSTANT, RAPID motion. The fissure PULSES and FLARES repeatedly, over and over. Rocks TUMBLE and BOUNCE continuously. Fog SURGES and CHURNS without pause." |
| 7 | G1 | G short | `prompts/G1_short.txt` | "The fissure pulses. Rocks tumble. Fog surges. Camera pushes in." (~60 chars) |
| 8 | G2 | G long | `prompts/G2_long.txt` | ~2500–4000 char dense enumeration of every element with independent verbs, repeated emphasis, camera path, beat structure. |
| 9 | H1 | H negation | `prompts/H1_negation.txt` | "No static moments. Nothing is still. The scene never stops moving: the fissure pulses, rocks tumble, fog surges, camera pushes in." |
| 10 | I1 | I combined | `prompts/I1_combined.txt` | Merge of the best-scoring elements from 2–9 (decided from data, not guessed). |

### Phase 2 — hill-climb, rounds 11–40

Take the top 2–3 families from Phase 1 and mutate **one variable at a time**, same seed, 3 seeds per promoted arm:

- **B/C/D arms:** verb count per layer (1 → 3 → 6), imperative vs. declarative mood ("Make the fissure pulse" vs. "The fissure pulses"), present-progressive vs. simple present.
- **C arms:** dolly-in vs. lateral track vs. arc/orbit vs. handheld shake; explicit parallax phrasing; explicit speed ("slow", "fast", "rapid").
- **D arms:** layer count (2 → 3 → 5), ordering (foreground-first vs. background-first), whether naming the *subject* as a layer helps or hurts.
- **E arms:** beat count (2 → 4 → 6), explicit timestamps ("at 0s… at 2s… at 4s…"), whether "then" chains beat "and" chains.
- **F arms:** capitalization, repetition count (×2, ×4, ×8), adverb stacking, ALL-CAPS vs. lowercase.
- **G arms:** length sweep at 60 / 300 / 800 / 1500 / 3000 / 6000 chars — this is the dilution test and it is the single most valuable controlled variable in the whole study.
- **H arms:** negation-only vs. positive-only vs. mixed; test whether "no static moments" actually suppresses stillness or is ignored.

### Phase 3 — exploit + regression, rounds 41–100

- 41–70: single-variable mutations on the best arm (the winner of Phase 2), 3 seeds each, looking for a stable plateau.
- 71–85: cross-family recombination of the top two arms.
- 86–100: **re-test earlier winners** at fresh seeds to confirm they still hold (guard against seed-luck promotion). Include a re-run of round 1's baseline prompt at a new seed as a control drift check.

---

## Seed discipline (explicit)

- Phase 1: all rounds at seed `3617007189758648082` (the base frame's seed) so family comparison is prompt-effect only.
- Any arm promoted to Phase 2/3: **3 seeds minimum** — `3617007189758648082`, `42`, `1337`. Report mean and spread of `motion_energy`; a scheme that wins on one seed and loses on two is noise, not a winner.
- Never compare two schemes at different seeds and call it a prompt effect.

---

## Logging format (append to FINDINGS.md round-log table)

```
| R | SCHEME | chars | subj | env | cam | coh | yield | motion_energy | active_area | col_shift | static_frac | seed | clip_url | note |
```

Plus, after each round, a two-line "what I learned / next hypothesis" note. The objective columns must be recorded alongside the sub-agent scores — the whole point of the study is correlating prompt structure with *both* the qualitative rubric and the objective meter.

---

## Decision rules for the orchestrator

1. **Promotion threshold:** an arm advances only if `overall_movement_yield ≥ 6` AND `motion_energy ≥ 2000` AND it holds across ≥2 of 3 seeds.
2. **Camera is the highest-leverage untested axis.** Round 1 got `col_shift_px 0` with zero camera language. If C1 produces `col_shift_px ≠ 0`, camera prompting is confirmed as a real lever and should dominate Phase 2 allocation.
3. **If G (length) shows dilution** (longer prompts score *lower*), cap all subsequent prompts at the length where yield peaks — this is a hard constraint on Phase 3.
4. **If H (negation) shows no effect**, drop negation from all combined prompts; negation is a known weak channel in diffusion conditioning.
5. **If all families plateau below yield 6**, the bottleneck is the model/step count, not the prompt — report that honestly rather than continuing to mutate text.

---

## Likely failure modes to watch (from the harness code)

- `run_round.py` exits code 2 on `is_error` — the orchestrator must catch this and log the failed round rather than stalling.
- `motion_score.py` requires `ffmpeg` on PATH and decodes at exactly 320×192; a resolution mismatch silently produces garbage frame counts. Verify `frames == 141` in every score output.
- `col_shift_px` search range is