# Astronaut Space Station Repair Scoring

## Rubric

### 40 - Visual authenticity (VQ)

- VQ1. Rendering quality: clarity, detail density, and visible artifact control
- VQ2. Lighting logic: coherent key light, shadow direction, and space-appropriate lighting
- VQ3. Material realism: believable metal, glass, suit, and station surfaces without obvious plastic AI gloss
- VQ4. Scene integration: foreground and background feel like one scene instead of pasted layers

### 25 - Physics and consistency (PC)

- PC1. Motion logic: drift, movement, and pose should feel consistent with microgravity
- PC2. Mechanical relation: tether force, astronaut-to-station relation, and spark or tool contact should make sense
- PC3. Dynamic consistency: Earth, debris, astronaut, and station motion should remain stable across frames without visible jumpiness

### 20 - Semantic restoration (SR)

- SR1. Element coverage: astronaut, space station, Earth, drifting debris, tool sparks, and safety tether
- SR2. Behavior alignment: the astronaut should read as actively performing maintenance, not just floating for a portrait
- SR3. Scene correctness: the result should clearly be outside a space station, not indoors, on the ground, or in a fantasy environment

### 15 - Composition and readability (CR)

- CR1. Composition clarity: the main subject is clear and the frame is not visually muddled
- CR2. Information hierarchy: the viewer can tell what matters most in the scene
- CR3. Story readability: the maintenance action is understandable at a glance

## Fixed Band Guide

Use these bands before assigning an exact total. The goal is to make close comparisons between four runs more stable.

### Visual authenticity (VQ) bands

- `0-10`: broken, extremely crude, or visually incoherent
- `11-20`: readable but clearly cartoonish, flat, or materially simplified
- `21-30`: solid and pleasant, but still stylized or visibly game-like rather than realism-first
- `31-36`: strong finish, better lighting logic, stronger station/suit material read
- `37-40`: exceptional coherence, realism, and integration across astronaut, station, Earth, and background

### Physics and consistency (PC) bands

- `0-6`: obvious jumpiness, broken drift, or nonsense element relations
- `7-12`: some motion works, but tether/contact/microgravity logic is weak
- `13-18`: motion is mostly stable, but at least one critical physical cue remains weak
- `19-22`: strong drift, tether logic, and stable frame-to-frame motion
- `23-25`: excellent microgravity read and highly believable mechanical relation

### Semantic restoration (SR) bands

- `0-5`: many required prompt elements are missing or scene is substantially wrong
- `6-10`: partial prompt coverage; several required elements are absent or unclear
- `11-14`: most required elements are present, but one prompt-critical behavior is still weak
- `15-17`: strong and explicit prompt coverage with only minor ambiguity
- `18-20`: complete, explicit restoration of all required elements and behavior

### Composition and readability (CR) bands

- `0-4`: cluttered or confusing
- `5-8`: readable enough, but weak hierarchy or muddy story
- `9-11`: clear subject, clear scene, acceptable story read
- `12-13`: strong framing and immediate comprehension
- `14-15`: excellent hierarchy and instantly readable maintenance story

## Astronaut-Specific Caps

Apply these caps before finalizing the score.

### Tether cap

- If the tether is missing, `SR` should stay at `12` or below.
- If the tether exists but its attachment or force relation is unclear, `PC` should usually stay at `18` or below.

### Maintenance-action cap

- If the astronaut reads mainly as a portrait or idle float and not as active maintenance, `SR` should stay at `14` or below.
- If sparks exist but do not feel connected to a believable repair contact, `PC` should not enter the top band.

### Earth-motion cap

- If Earth is nearly static or visually negligible in motion review, `PC3` should be treated as partial rather than strong.
- Visible Earth motion alone does not justify a high total, but it should raise `PC` when the motion is genuinely readable.

### Station-context cap

- If the station is too sparse to support a clear "outside a space station" read, `SR` should stay at `14` or below.
- If the astronaut is isolated in open space without a convincing work target, `CR` should stay at `11` or below.

### Realism cap

- A clean cartoon or game-like look can still score well, but it should not reach the top `VQ` band unless materials, lighting, and integration start to read realism-first.

## Tie-Break Rule

When two astronaut runs land within `3` total points:

1. Prefer the higher `PC`
2. Then prefer the higher `SR`
3. Then prefer the higher `VQ`
4. Use `CR` last

This prevents a prettier still from automatically beating a run with better tether logic or better maintenance behavior.

## Practical Anchor Rule

When adding a new astronaut run, place it relative to the current anchors first:

- `GLM 80`
- `Qwen 77`
- `DeepSeek 67`
- `Doubao 53`

Ask:

- Is it clearly below `DeepSeek 67`?
- Is it roughly tied with `DeepSeek 67` but stronger on motion or weaker on realism?
- Is it closer to `Doubao 53` because it misses maintenance semantics or station context?

Only after that should the exact score be assigned.

## Evidence Expectations

- screenshot per run
- sampled video frames or clip review for temporal scoring
- notes on realism, prompt coverage, and visible physical failures