Sol was the visual favorite; Opus 5 led Prism Garden’s scorecard, 77–70
In one blind run, the reviewer preferred Sol’s richer prism work with low confidence, while Opus 5 scored higher on interaction and robustness.
The short versionSol was the visual favorite. Opus 5 led the weighted scorecard.The reviewer marked the preference low-confidence. One run; no broader ranking is implied.
Reviewer pickopenai/gpt-5.6-sol
Confidencelow
Blind score70:77
OutcomeSplit decision
The preference
did not follow the score.
Contestant A won this run of Prism Garden, but not by the numbers. The locked human evaluation selected Sol with low confidence even though Opus 5 received the higher aggregate score: 77 to 70.
The deciding factor was taste. The reviewer preferred Sol’s visual quality and prism detail while describing Opus 5 as more stable. This is not a claim of comprehensive superiority; it is a polished visual choice overcoming weaker interaction and robustness scores in one frozen, blind comparison.
“While Contestant B's result is more stable, Contestant A's visual quality and detail in the prisms makes me prefer A.”Evaluation locked before model reveal
Contestant Aopenai/gpt-5.6-sol
Reviewer pick
Contestant A · preserved captureScore70
Wall time3:16
HTML start1:44
Cost$0.3885
More varied outlined petals, stems, leaves, and prism detail drove the reviewer’s visual preference.
Contestant Banthropic/claude-opus-5
Score leader
Contestant B · preserved captureScore77
Wall time3:56
HTML start2:32
Cost$0.5096
A more restrained star system earned stronger interaction and robustness scores, despite one recorded runtime error.
Publication-primary captures show reviewer-created interaction states with 34 Sol blooms and 33 Opus 5 blooms. They establish visible composition, not animation quality or performance.
Both met the brief.
The tradeoff was real.
Both contestants passed the five concrete planting, animation, counter, and Reset requirements. Both received partial credit for the combined responsiveness and offline requirement.
Frozen requirements5 met / 1 partial each
| Requirement | A | B |
|---|---|---|
| Clicking plants a luminous geometric flower. | Met | Met |
| Dragging plants multiple blooms without runaway creation. | Met | Met |
| Blooms animate and show meaningful visual variation. | Met | Met |
| The visible bloom counter remains accurate. | Met | Met |
| Reset returns the experience to a clean initial state. | Met | Met |
| The single-file experience remains responsive and works offline. | Partial | Partial |
Weighted criteriaA 70 / B 77
Requirement fit · 25%
How completely and accurately the artifact satisfies the frozen brief.
A / 1010
B / 1010
Visual design · 25%
Hierarchy, composition, typography, color, polish, and coherent visual judgment.
A / 108
B / 107
Interaction & usability · 25%
Discoverability, responsiveness, feedback, control quality, and interaction feel.
A / 105
B / 107
Technical robustness · 20%
Runtime correctness, edge-case handling, performance, responsiveness, and code reliability.
A / 105
B / 107
Originality · 5%
Useful creative choices that go beyond a generic first solution without harming clarity.
A / 105
B / 105
What decided it
Sol’s one-point advantage in visual design was enough to determine the winner because the reviewer’s final preference was not mechanically bound to the aggregate score. Opus 5 still led by two points each in interaction and technical robustness, producing the higher weighted total.
Both probes passed.
One error remained.
Opus 5’s runtime record contains an uncaught error caused by a NaN hue passed to CanvasGradient.addColorStop. Because its animation loop schedules the next frame only after drawing, the error could stop subsequent rendering.
The exact trigger was not reproduced, and Opus 5’s formal probe logged no errors. The run therefore establishes that the defect occurred at least once—not that the preserved session was confirmed to freeze.
The limitNo frame-rate or memory measurements were collected. Performance observations from code remain risks and explanations, not measured outcomes.
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 RESET
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 RESET
The recorded hsla(NaN,95%,60%,0.04) value is a reachable defect, although the trigger was not reproduced.
Sol caps bloom population, limits interpolation, constrains pixel ratio, and reduces blur above 65 flowers.
Opus 5 uses offscreen sprites and limits sprite baking to three flowers per frame, but no actual throughput was measured.
As many as four automatic seed blooms can increment the counter before the user plants anything.
Sol labels and focus-enables the canvas, supports keyboard planting, exposes counter updates, and respects reduced motion.
Static review found no external assets, fonts, or libraries. Offline operation was not independently exercised.
The claim, with its
conditions attached.
Limits of the claimRead the caveats
This was one quick blind comparison under a single frozen prompt, not a universal ranking. The manually created screenshots cannot establish animation quality or interaction feel. Offline operation was inferred, not exercised. Population-limit behavior, frame rate, and memory were not measured. The exact trigger for Opus 5’s error was not reproduced, and its controlled probe completed without errors.
The frozen assignmentShow the exact prompt
Build a one-screen interactive web experience called PRISM GARDEN. Use a dark near-black background. Users click or drag to plant luminous geometric flowers; each flower should bloom with motion and slight variation. Include a small visible bloom counter and a Reset control. Make it polished, responsive, and immediately understandable. Use only HTML, CSS, and vanilla JavaScript in a single file. No external assets, fonts, or libraries. It must work offline. Spend time on the experience, not explanations.
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-MSM51I66-51
Protocolproofrun-v0.6
Test lineagev1 · curated:prism-garden:v1
Identityrevealed
Evaluationlocked
Record hashsha256:555d37401b2a8769367afadffe6dcd7a127349f06bbc0dfb37374d7c12991970
AI assistance
Download the hashed publication record ↓Evidence analysis used anthropic/claude-fable-5 while identities were hidden; 7 findings were accepted and 1 rejected. The article draft used openai/gpt-5.6-sol, was revised, and was approved by the human editor. Recorded assistance cost: $0.7885.