Claude approved its own broken 3D screenshots — how a 480Hz verification gate fixed agent self-grading

youthanasia669 · reddit · 2026-10-24

A 30-year software engineer's retrospective on rebuilding the 3D pool game Sinuca with Claude: AI lowered the barrier to making things, but the barrier to knowing what's good still stands — and you can't let the agent grade its own homework.

Classic failures the agent had 'verified':

The table was just code boxes with a wood texture; after switching to a real 8-foot mesh (119,302 triangles), gltfpack -cc silently quantized coordinates above 1 meter and crushed half the felt into a dark strip — and the build still passed.
Wall clock hands orbited the wall due to wrong pivots (set to the shaft's geometric center, or the whole alarm clock's bounding box).
The cue stick had no taper and clipped through rails and balls instead of auto-elevating over cushions.

What changed:

A custom deterministic 480Hz physics engine in pure JS: 1/480s sub-steps, full 3-axis angular velocity, sliding/rolling friction, vertical spin decay, squirt deflection, massé curves, object-ball throw, velocity-dependent cushion restitution, and real pocket jaws.
Deliberately added "living world" details an LLM never proposes: Brazilian beer glasses with refraction 1.52 and rigid-body physics (knock them and they shatter with HRTF audio), 8 patrons tracking the cue ball, rainy windows, 16 original tracks.
A multi-layer verification gate (AI-DLC) replaced single-screenshot checks: 348 unit tests across 24 files, 33/33 mutants killed (100%), 50 canonical desktop screenshots (5 resolutions × 10 states), 66 mobile checks — all must pass before a goal closes.

Original post →

Claude approved its own broken 3D screenshots — how a 480Hz verification gate fixed agent self-grading

More from coding & agent