# Capsule Brief: Ray Marching 3D Scene ## 1. Why this capsule exists Ray marching with SDFs is one of the highest-signal coding benchmarks available for LLMs because it sits at the intersection of three things a model can't fake with pattern-matching alone: - **3D math correctness** (normal estimation via gradient, camera ray generation, reflection/lighting equations) - **Numerical stability** (step count, epsilon thresholds, banding artifacts if done wrong) - **Performance awareness** (naive per-pixel loops in JS Canvas can be catastrophically slow if the model doesn't reason about ray-march step limits or resolution) Unlike most "impressive-looking" AI-generated demos (particle systems, simple CSS art), a broken ray marcher is immediately, visually obvious even to someone with zero coding background: the scene looks flat, shapes are wrong, shadows don't move, or the browser tab freezes. This makes it an excellent "Benchy-style" capsule — verifiable ground truth without domain expertise. ## 2. Expected spread between models - **Weak/small models:** usually produce a flat-shaded scene with a single sphere, no shadows, no blending, sometimes syntactically broken GLSL/JS that fails silently (blank canvas). - **Mid-tier models:** get the SDF math right for basic shapes and lighting, but skip soft shadows or AO, or produce visible banding/artifacts. - **Frontier models:** correctly implement smooth union blending, soft shadows, AO, and keep frame rate usable — this is where the gap becomes dramatic and immediately visible on the Timeline comparison view. ## 3. Known implementation risks to watch for when grading - Some models fake "3D" with a pre-rendered CSS/SVG trick rather than actual ray marching — check the source code, not just the visual output. - Some models hardcode a static image behind a CSS animation instead of an actual per-frame SDF evaluation — again, verifiable by reading the code. - WebGL context creation can silently fail on some sandboxed viewers; make sure the Timeline run environment actually supports a WebGL canvas before penalizing a model for a blank screen that isn't its fault. --- ## 4. Ground Truth ("Vérité Terrain") A correct implementation must satisfy ALL of the following, checkable by anyone just by looking at the rendered output — no code reading required for the pass/fail call (code reading is only needed for the "did it cheat" check): | # | Ground truth check | How to verify visually | |---|---|---| | GT1 | At least 3 shapes are visually merging into each other with rounded blend seams (not just touching/overlapping as separate hard-edged objects) | Look at the junction between shapes — should look organic/rounded, like melted wax, not a sharp intersection line | | GT2 | A ground plane is visible and shapes appear to sit on it | There's a flat surface below the shapes with correct perspective | | 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 | | GT4 | Some part of the scene is animating (camera orbiting, light moving, or shape morphing) | Just watch it for 3 seconds — something is moving | | GT5 | No blank canvas / no browser console errors on load | Open dev tools console — should be error-free | | GT6 | Runs at usable frame rate (not a slideshow) on a normal laptop | Subjective but obvious — is it smooth-ish or a slideshow? | | GT7 | Ambient occlusion is present in some form | Look at concave junctions between merged shapes — should be slightly darker/shaded, not flat-lit like the rest of the surface | **Pass/fail severity tiers (for the Winners metric):** - **Full pass:** GT1–GT7 all satisfied → top tier - **Partial pass:** GT1, GT2, GT4, GT5 satisfied but shadows are hard-edged (GT3 fail) or no AO (GT7 fail) → mid tier - **Fail:** blank canvas, console errors, or a single flat-shaded non-blended primitive → bottom tier --- ## 5. Scoring Checklist (for the multi-model analysis report) ``` [ ] GT1 - Smooth blending between ≥3 primitives (pass/fail) [ ] GT2 - Ground plane present with correct perspective (pass/fail) [ ] GT3 - Soft shadows with visible penumbra gradient (pass/fail) [ ] GT4 - Scene is animated (camera/light/shape motion) (pass/fail) [ ] GT5 - No console errors, canvas renders on first load (pass/fail) [ ] GT6 - Frame rate subjectively usable, not a slideshow (pass/fail) [ ] GT7 - Ambient occlusion visible at concave junctions (pass/fail) Code integrity check (not part of visual score, but flags a "cheat"): [ ] Confirmed actual per-pixel SDF ray marching in source (not faked with CSS/SVG/pre-rendered image + animation) Final tier: [ Full pass / Partial pass / Fail ] Notes: _______________________________________________ ``` --- ## 6. Reference Asset No reference image needed for this capsule (unlike the perception capsules) — the ground truth is procedural/behavioral, not a fixed answer key. The "reference" is this checklist itself, applied consistently across every model run. Optional: once the first 3–5 runs come in, consider saving a screenshot of the **best-performing run** as a visual reference point in the capsule's metadata, purely for calibration when grading future runs (not as a "correct answer" — outputs will legitimately vary in style).