GPT-5.6 Sol — effort matrix
Same task (build the ThinkRun story page), same model, isolated sandboxes. Big number = quality score 0-100. No-ship runs (built nothing): medium-r1, high-r2, high-r3, high-r8.
The 20-run experiment (16 shipped) — best to worst
80medium-r10medium · $1.13 · 21min · 26KB77high-r10high · $2.87 · 57min · 45KB61medium-r7medium · $0.65 · 24min · 21KB60medium-r6medium · $1.05 · 26min · 18KB60medium-r3medium · $0.52 · 51min · 21KB60medium-r9medium · $0.46 · 4min · 23KB60medium-r8medium · $1.19 · 60min · 25KB60high-r4high · $1.33 · 13min · 34KB60high-r5high · $1.35 · 53min · 34KB60high-r7high · $0.66 · 7min · 25KB58high-r1high · $1.17 · 53min · 46KB52medium-r4medium · $0.75 · 5min · 29KB52medium-r5medium · $0.95 · 53min · 20KB52high-r6high · $1.40 · 52min · 46KB52high-r9high · $1.21 · 9min · 37KB · unverified29medium-r2medium · $1.16 · 12min · 28KB
Earlier story runs (the context experiments, pre-matrix)
→A · staged sources · medium"323 Days to Make the Web Runnable"→B · staged sources · high"The machine that learned to run"→C · full repo · medium"The browser learned to remember"→D · full repo · high"The browser remembers"