[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"capsule:06f5113d-6737-4c3b-bb89-40340f0e8dbf":3},["Reactive",4],{"id":5,"user_id":6,"title":7,"prompt_content":8,"prompt_type":9,"is_locked":10,"is_public":10,"fork_of":11,"created_at":12,"updated_at":13,"category":14,"brief":15,"attachments":16,"expected_markers":17,"user":19},"06f5113d-6737-4c3b-bb89-40340f0e8dbf","695c8e6c-9654-4858-b2d3-925adefacc35","Ray Marching 3D Scene","Create a single self-contained HTML file that renders a 3D scene using ray marching with signed distance functions (SDF) — no Three.js, no WebGL libraries, no external dependencies. You may use raw Canvas 2D with a per-pixel JS ray marcher, or a raw WebGL context with a hand-written GLSL fragment shader. Either approach is valid.\n\nRequirements:\n1. The scene must contain at least 3 distinct SDF primitives (e.g. spheres, boxes, toruses) blended together with a smooth minimum (smooth union), so they merge like metaballs rather than just overlapping.\n2. Include a ground plane.\n3. Implement at least one light source with soft shadows (shadows with penumbra, not just a hard binary shadow ray).\n4. Implement basic ambient occlusion (even a cheap approximation based on the SDF is fine).\n5. Animate the scene: either the camera orbits around the scene, or the light source moves, or the primitives move\u002Fdeform over time. Pick at least one.\n6. The whole thing must run smoothly in a single HTML file when opened directly in a browser, no build step, no external assets.\n\nOutput only the final HTML file, fully self-contained.","text",true,null,"2026-07-05T13:05:55.144405+00:00","2026-07-05T14:33:20.733+00:00","coding","# Capsule Brief: Ray Marching 3D Scene\n## 1. Why this capsule exists\n\nRay marching with SDFs is one of the highest-signal coding benchmarks available for\nLLMs because it sits at the intersection of three things a model can't fake with\npattern-matching alone:\n\n- **3D math correctness** (normal estimation via gradient, camera ray generation,\n  reflection\u002Flighting equations)\n- **Numerical stability** (step count, epsilon thresholds, banding artifacts if done\n  wrong)\n- **Performance awareness** (naive per-pixel loops in JS Canvas can be catastrophically\n  slow if the model doesn't reason about ray-march step limits or resolution)\n\nUnlike most \"impressive-looking\" AI-generated demos (particle systems, simple CSS art),\na broken ray marcher is immediately, visually obvious even to someone with zero coding\nbackground: the scene looks flat, shapes are wrong, shadows don't move, or the browser\ntab freezes. This makes it an excellent \"Benchy-style\" capsule — verifiable ground\ntruth without domain expertise.\n\n## 2. Expected spread between models\n\n- **Weak\u002Fsmall models:** usually produce a flat-shaded scene with a single sphere, no\n  shadows, no blending, sometimes syntactically broken GLSL\u002FJS that fails silently\n  (blank canvas).\n- **Mid-tier models:** get the SDF math right for basic shapes and lighting, but skip\n  soft shadows or AO, or produce visible banding\u002Fartifacts.\n- **Frontier models:** correctly implement smooth union blending, soft shadows, AO, and\n  keep frame rate usable — this is where the gap becomes dramatic and immediately\n  visible on the Timeline comparison view.\n\n## 3. Known implementation risks to watch for when grading\n\n- Some models fake \"3D\" with a pre-rendered CSS\u002FSVG trick rather than actual ray\n  marching — check the source code, not just the visual output.\n- Some models hardcode a static image behind a CSS animation instead of an actual\n  per-frame SDF evaluation — again, verifiable by reading the code.\n- WebGL context creation can silently fail on some sandboxed viewers; make sure the\n  Timeline run environment actually supports a WebGL canvas before penalizing a\n  model for a blank screen that isn't its fault.\n\n---\n\n## 4. Ground Truth (\"Vérité Terrain\")\n\nA correct implementation must satisfy ALL of the following, checkable by anyone just\nby looking at the rendered output — no code reading required for the pass\u002Ffail call\n(code reading is only needed for the \"did it cheat\" check):\n\n| # | Ground truth check | How to verify visually |\n|---|---|---|\n| GT1 | At least 3 shapes are visually merging into each other with rounded blend seams (not just touching\u002Foverlapping as separate hard-edged objects) | Look at the junction between shapes — should look organic\u002Frounded, like melted wax, not a sharp intersection line |\n| GT2 | A ground plane is visible and shapes appear to sit on it | There's a flat surface below the shapes with correct perspective |\n| GT3 | Shadows exist and have a soft edge (gradient from dark to light), not a hard binary edge | Zoom into the shadow boundary — should fade, not have a jagged 1px hard line |\n| GT4 | Some part of the scene is animating (camera orbiting, light moving, or shape morphing) | Just watch it for 3 seconds — something is moving |\n| GT5 | No blank canvas \u002F no browser console errors on load | Open dev tools console — should be error-free |\n| GT6 | Runs at usable frame rate (not a slideshow) on a normal laptop | Subjective but obvious — is it smooth-ish or a slideshow? |\n| GT7 | Ambient occlusion is present in some form | Look at concave junctions between merged shapes — should be slightly darker\u002Fshaded, not flat-lit like the rest of the surface |\n\n**Pass\u002Ffail severity tiers (for the Winners metric):**\n- **Full pass:** GT1–GT7 all satisfied → top tier\n- **Partial pass:** GT1, GT2, GT4, GT5 satisfied but shadows are hard-edged (GT3 fail) or no AO (GT7 fail) → mid tier\n- **Fail:** blank canvas, console errors, or a single flat-shaded non-blended primitive → bottom tier\n\n---\n\n## 5. Scoring Checklist (for the multi-model analysis report)\n\n```\n[ ] GT1 - Smooth blending between ≥3 primitives (pass\u002Ffail)\n[ ] GT2 - Ground plane present with correct perspective (pass\u002Ffail)\n[ ] GT3 - Soft shadows with visible penumbra gradient (pass\u002Ffail)\n[ ] GT4 - Scene is animated (camera\u002Flight\u002Fshape motion) (pass\u002Ffail)\n[ ] GT5 - No console errors, canvas renders on first load (pass\u002Ffail)\n[ ] GT6 - Frame rate subjectively usable, not a slideshow (pass\u002Ffail)\n[ ] GT7 - Ambient occlusion visible at concave junctions (pass\u002Ffail)\n\nCode integrity check (not part of visual score, but flags a \"cheat\"):\n[ ] Confirmed actual per-pixel SDF ray marching in source (not faked with\n    CSS\u002FSVG\u002Fpre-rendered image + animation)\n\nFinal tier: [ Full pass \u002F Partial pass \u002F Fail ]\nNotes: _______________________________________________\n```\n\n---\n\n## 6. Reference Asset\n\nNo reference image needed for this capsule (unlike the perception capsules) — the\nground truth is procedural\u002Fbehavioral, not a fixed answer key. The \"reference\" is\nthis checklist itself, applied consistently across every model run.\n\nOptional: once the first 3–5 runs come in, consider saving a screenshot of the\n**best-performing run** as a visual reference point in the capsule's metadata, purely\nfor calibration when grading future runs (not as a \"correct answer\" — outputs will\nlegitimately vary in style).",[],{"items":18,"matchAll":10},[],{"username":20,"avatar_url":21},"Ezarwebmaster","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F42354978?v=4"]