Visual Scene

Astronaut Space Station Repair

A scored space-maintenance visual case comparing four runs on realism, microgravity logic, and prompt fidelity.

GameCanvasSpaceRepairVisual review

Watch episode

Published video

Astronaut four-way comparison

The primary video slot points to the four-way astronaut comparison. Scoring for this case is now driven by a realism-first visual rubric rather than the earlier gameplay-oriented checklist.

10.6s

Watch full episode Watch comparison short Watch best-run short Watch worst-run short

Result matrix

One task, 4 recorded outcomes

The table shows visible results first. Detailed logs and clips can be attached once real runs are recorded.

Agent	Model	Status	Score	Breakdown	Time	Cost	Failure
Claude Code	GLM	Strong	A-80	Visual authenticity (VQ)30/40 Physics and consistency (PC)21/25 Semantic restoration (SR)17/20 Composition and readability (CR)12/15	10.5s clip reviewed	unknown	The scene still reads as stylized illustration rather than near-photoreal simulation, and the repair action plus tool sparks are present only modestly.
Claude Code	Qwen	Strong	B+77	Visual authenticity (VQ)32/40 Physics and consistency (PC)17/25 Semantic restoration (SR)15/20 Composition and readability (CR)13/15	10.5s clip reviewed	unknown	The tether is not clearly readable, the frame is visually busier than GLM, and the maintenance action is implied more than cleanly demonstrated.
Claude Code	DeepSeek	Partial	B-67	Visual authenticity (VQ)26/40 Physics and consistency (PC)18/25 Semantic restoration (SR)12/20 Composition and readability (CR)11/15	10.8s clip reviewed	unknown	Earth presence is weak, debris and spark behavior are limited, and the clip reads more like a floating station vignette than an active maintenance sequence.
Claude Code	Doubao	Weak	C53	Visual authenticity (VQ)19/40 Physics and consistency (PC)13/25 Semantic restoration (SR)11/20 Composition and readability (CR)10/15	10.7s clip reviewed	unknown	The image is the least realistic of the set, station structure is sparse, the red tether or tool relation feels artificial, and the maintenance story remains under-specified.

Claude Code

GLM

A-

Status: Strong
Score: 80
Breakdown: Visual authenticity (VQ)30/40
Physics and consistency (PC)21/25
Semantic restoration (SR)17/20
Composition and readability (CR)12/15
Time: 10.5s clip reviewed
Failure: The scene still reads as stylized illustration rather than near-photoreal simulation, and the repair action plus tool sparks are present only modestly.

Claude Code

Qwen

B+

Status: Strong
Score: 77
Breakdown: Visual authenticity (VQ)32/40
Physics and consistency (PC)17/25
Semantic restoration (SR)15/20
Composition and readability (CR)13/15
Time: 10.5s clip reviewed
Failure: The tether is not clearly readable, the frame is visually busier than GLM, and the maintenance action is implied more than cleanly demonstrated.

Claude Code

DeepSeek

B-

Status: Partial
Score: 67
Breakdown: Visual authenticity (VQ)26/40
Physics and consistency (PC)18/25
Semantic restoration (SR)12/20
Composition and readability (CR)11/15
Time: 10.8s clip reviewed
Failure: Earth presence is weak, debris and spark behavior are limited, and the clip reads more like a floating station vignette than an active maintenance sequence.

Claude Code

Doubao

Status: Weak
Score: 53
Breakdown: Visual authenticity (VQ)19/40
Physics and consistency (PC)13/25
Semantic restoration (SR)11/20
Composition and readability (CR)10/15
Time: 10.7s clip reviewed
Failure: The image is the least realistic of the set, station structure is sparse, the red tether or tool relation feels artificial, and the maintenance story remains under-specified.

What this tests

The hidden engineering inside a simple game

Scene rendering

Builds a readable space scene instead of a blank or broken canvas.

Movement and control

Supports visible player motion and direction changes in a constrained environment.

Objective clarity

Communicates station repair goals, hazards, or mission state clearly enough to evaluate.

Readable game state

Keeps score, progress, damage, or repair cues legible during play.

Visual evidence quality

Produces captures that can support later scoring, clips, and side-by-side comparison.

Unified prompt

Create a single-file HTML page. In a space background, an astronaut is floating outside a space station performing maintenance. The astronaut is connected to the space station via a safety tether. The Earth slowly rotates in the background. Small pieces of space debris continuously drift around. The tool in the astronaut's hand periodically emits sparks. The entire scene should convey a realistic sense of weightlessness and floating in microgravity.

Acceptance criteria

The page renders a visible playable scene without fatal runtime errors.
The player can control an astronaut or equivalent repair avatar.
The game communicates a repair objective, mission state, or survival loop clearly.
Movement and interaction are readable on screen during normal play.
The implementation preserves enough visible state to judge progress or failure.

Scoring

Readable rubric for video and reproduction

Visual authenticity (VQ)

Judges rendering quality, lighting logic, material realism, and how well the astronaut, station, Earth, and background feel like one coherent scene.

Physics and consistency (PC)

Judges microgravity motion, tether or contact logic, spark or tool interaction, and whether motion remains temporally stable across the clip.

Semantic restoration (SR)

Judges whether the run actually follows the prompt by covering the astronaut, station, Earth, debris, tether, sparks, and a believable maintenance action.

Composition and readability (CR)

Judges composition clarity, visual hierarchy, and whether a viewer can understand the maintenance story at a glance.

Run details

What each agent got right and wrong

GLM gameplay clipWatch clip

Screenshots

Captured run

Claude Code

GLM

A-80

Visual authenticity (VQ)30/40

Physics and consistency (PC)21/25

Semantic restoration (SR)17/20

Composition and readability (CR)12/15

Status: Strong
Strength: Most balanced overall result: clear tether logic, readable astronaut-station relationship, stable motion across the clip, and the cleanest high-level maintenance story.
Failure: The scene still reads as stylized illustration rather than near-photoreal simulation, and the repair action plus tool sparks are present only modestly.

Qwen gameplay clipWatch clip

Screenshots

Captured run

Claude Code

Qwen

B+77

Visual authenticity (VQ)32/40

Physics and consistency (PC)17/25

Semantic restoration (SR)15/20

Composition and readability (CR)13/15

Status: Strong
Strength: Richest station detail and debris field in the set, with strong scene density and a convincing sense of working outside a larger orbital structure.
Failure: The tether is not clearly readable, the frame is visually busier than GLM, and the maintenance action is implied more than cleanly demonstrated.

DeepSeek gameplay clipWatch clip

Screenshots

Captured run

Claude Code

DeepSeek

B-67

Visual authenticity (VQ)26/40

Physics and consistency (PC)18/25

Semantic restoration (SR)12/20

Composition and readability (CR)11/15

Status: Partial
Strength: Best after GLM for readable tether attachment and stable drift, with a restrained composition that stays easy to parse frame to frame.
Failure: Earth presence is weak, debris and spark behavior are limited, and the clip reads more like a floating station vignette than an active maintenance sequence.

Doubao gameplay clipWatch clip

Screenshots

Captured run

Claude Code

Doubao

C53

Visual authenticity (VQ)19/40

Physics and consistency (PC)13/25

Semantic restoration (SR)11/20

Composition and readability (CR)10/15

Status: Weak
Strength: The main character silhouette is immediately readable and the oversized planet backdrop creates an instant space cue.
Failure: The image is the least realistic of the set, station structure is sparse, the red tether or tool relation feels artificial, and the maintenance story remains under-specified.

Postmortem

Why The Astronaut Case Needed A Different Scoring System

Realism beats pure prettiness here

A nice-looking space scene is not enough. The strongest runs are the ones that keep tether logic, astronaut placement, and station context believable while still looking composed.

Prompt fidelity is a hard requirement

The rubric now rewards explicit coverage of the tether, Earth, debris, sparks, and a real maintenance action. Rich detail without those prompt-critical cues no longer scores as highly.

Video matters because physics matters

Static screenshots help with composition and materials, but the video pass is what reveals whether microgravity motion, tether behavior, and frame-to-frame consistency actually hold up.

Case package

starter/
tests/
prompt.md
task.md
scoring.md
runs.json
screenshots/
clips/
logs/

Manifest Prompt Scoring Runs JSON

Commands

npm install
npm run dev
npm test

Use this scaffold to collect prompt, scoring, screenshots, clips, logs, and final run summaries before marking the case as published.

Astronaut Space Station Repair

Astronaut four-way comparison

One task, 4 recorded outcomes

Claude Code

Claude Code

Claude Code

Claude Code

The hidden engineering inside a simple game

Scene rendering

Movement and control

Objective clarity

Readable game state

Visual evidence quality

Readable rubric for video and reproduction

Visual authenticity (VQ)

Physics and consistency (PC)

Semantic restoration (SR)

Composition and readability (CR)

What each agent got right and wrong

Claude Code

Claude Code

Claude Code

Claude Code

Why The Astronaut Case Needed A Different Scoring System

Realism beats pure prettiness here

Prompt fidelity is a hard requirement

Video matters because physics matters

Ready for the next visual task?