harness-probe

Three single-file Three.js tasks, one protocol, several surfaces. Claude Code column below; other surfaces are added to runs/<surface>/<task>/. Protocol: HANDOFF · RESULTS.md · FAILURE-MODES.md · post-hoc comparison with the Google demos (side-by-side)

Claude Code runs (open in a real browser; the judge screenshots below are headless at ~5 fps)

Task A · seaplane, golden hour →Water addon + own sky shader · Space pauses · iterations

Task B · brougham from a reference image →click doors · R resets · reference · iterations

Task C · marble run, hand-written physics →button releases · R clears · iterations

RESULTS.csv

Rendered live from RESULTS.csv (written by tools/sweep.sh). Blank cells are human-judged columns, or surfaces not yet run. Token figures for Claude Code are estimates (±30%).