Terra’s steadier physics beat Sonnet 5’s audio edge in Gravity Desk, 87–73
In one blind, frozen comparison, contestant B won with high confidence because stronger physics, control, and presentation outweighed A’s more responsive-feeling sound feedback.
The short versionTerra won on feel, control, and finish. Sonnet’s clearest edge was sound.One run. One frozen brief. No broader ranking is implied.
Verdictopenai/gpt-5.6-terra
Confidencehigh
Blind score73:87
Margin+14
Same brief.
Different finish.
Both models delivered the complete one-screen physics toy: clickable marbles, gravity, collisions, working controls, and a live object count. Completion did not decide this comparison.
Terra’s movement felt steadier and its controls and presentation felt more resolved. Sonnet added sound that made individual actions feel more responsive, but that advantage did not outweigh the gap in physics and finish.
“B's result is overall better because of the stabler physics, more control over the balls and better visuals overall.”Evaluation locked before model reveal
Contestant Aanthropic/claude-sonnet-5
Runner-up
Contestant A · preserved captureScore73
Wall time1:46
HTML start1:00
Cost$0.1229
Sound improved immediate feedback. Physics stability, robustness, and visual presentation scored below B.
Contestant Bopenai/gpt-5.6-terra
Winner
Contestant B · preserved captureScore87
Wall time1:32
HTML start0:51
Cost$0.1361
Stronger motion, control, and presentation won the run. The populated state also reveals footer overlap.
Publication-primary captures show reviewer-created populated states. A contains 44 marbles and B contains 31, so the images compare presentation—not capacity.
Both met the brief.
The quality was not equal.
Every requirement was judged before model identities were revealed. The weighted score then separated basic completion from how convincing the result felt in use.
Frozen requirements5 met / 1 partial each
| Requirement | A | B |
|---|---|---|
| Clicking spawns a colorful glass-like marble. | Met | Met |
| Marbles fall and bounce within the viewport. | Met | Met |
| Marbles collide convincingly enough for the intended toy. | Met | Met |
| Gravity and Clear controls work during active simulation. | Met | Met |
| The live object count remains accurate. | Met | Met |
| The single-file canvas experience remains smooth, responsive, and offline. | Partial | Partial |
Weighted criteriaA 73 / B 87
Requirement fit · 20%
How completely and accurately the artifact satisfies the frozen brief.
A / 1010
B / 1010
Physics & feel · 25%
Believability, responsiveness, collision behavior, and the satisfaction of motion.
A / 107
B / 1010
Technical robustness · 25%
Runtime correctness, edge-case handling, performance, responsiveness, and code reliability.
A / 106
B / 107
Visual design · 20%
Hierarchy, composition, typography, color, polish, and coherent visual judgment.
A / 106
B / 107
Interaction & usability · 10%
Discoverability, responsiveness, feedback, control quality, and interaction feel.
A / 108
B / 1010
What the code suggested
Both models implemented substantive collision systems. Terra used a fixed 1/120-second accumulator, three collision passes, low-speed restitution, and tangential friction. Those choices help explain the result, but the runtime judgment—not code aesthetics—determined the score.
The winner was stronger.
It was not spotless.
Terra’s populated capture shows a fixed footer crossing the settled marbles. It also lacks a recorded automated interaction probe. Its controls passed human evaluation, but the packet cannot independently confirm those paths.
Sonnet’s probe successfully exercised a click, drag, five-point burst, and Clear at a narrow viewport without runtime errors. That confirms the input paths executed; it does not prove visual correctness or performance near the 260-marble limit.
The limitNo motion telemetry or video was preserved. “Steadier physics” remains a bounded human judgment from this run, not a measured simulation claim.
Narrow viewport
536 x 520 viewport exercised
Single click
Dispatched one complete pointer and mouse click sequence
Short drag
Dispatched a bounded six-step diagonal drag
Five-point burst
Dispatched five bounded clicks across the interaction surface
Reset control
Activated 🧹 Clear
No automated interaction probe was preserved for this contestant.
Pairwise collision checks could become expensive near the 260-marble cap. The run did not establish an observed slowdown.
The canvas has no keyboard spawning route or accessible name, and viewport scaling is disabled.
The fixed footer competes with the settled marble row during ordinary play.
The claim, with its
conditions attached.
Limits of the claimRead the caveats
This was one blind comparison under frozen conditions, not a universal ranking. Static captures cannot establish bounce consistency, tunneling, pile stability, or collision satisfaction. Terra lacks a recorded automated probe; Sonnet’s probe was too limited to establish high-population performance. Audio, physical-device touch behavior, reduced-motion preferences, and low-power hardware were not systematically tested.
The frozen assignmentShow the exact prompt
Build a playful one-screen physics toy called GRAVITY DESK. Clicking spawns colorful glass-like marbles that fall, bounce off the viewport edges, and collide convincingly enough to feel satisfying. Include Gravity and Clear controls plus a live object count. Use a single offline HTML file with Canvas and vanilla JavaScript, no libraries or assets. Prioritize smoothness, polish, and immediate playability.
Method and publication recordInspect provenance
The public page uses a sanitized, allowlisted record. Credentials, raw responses, generated source, executable artifacts, and private reviewer notes remain in the private Studio.
Run recordPR-MSKKQ2K6-34
Protocolproofrun-v0.6
Test lineagev1 · curated:gravity-desk:v1
Identityrevealed
Evaluationlocked
Record hashsha256:ceb35e3680ce91601e1004a220a487ef6beea0fdb36e5448b8ef47e6f5a047e3
AI assistance
Download the hashed publication record ↓Evidence analysis used openai/gpt-5.6-sol while identities were hidden; 6 findings were accepted and 3 rejected. The article draft used openai/gpt-5.6-sol and was approved by the human editor. Recorded assistance cost: $0.4538.