AI LABS
Trial Wing · Controlled Comparison
TRIAL EXP-025 · n=3 EACH ARM · PRE-REGISTERED JUDGE
REPLAY
same model · same 9 tasks · same kernel judge — the only variable is the AGIOS loop
CELL A · VERIFY LOOP ON harness + oracle + repair + memory 9 / 9 CRYSTALLIZED 100% · 3 CONSECUTIVE RUNS · ZERO VARIANCE · $0.165/RUN CELL B · RAW MODEL one shot · no tools · no oracle · no memory 81.5% MEAN · 2 SHATTERED RANGE 77.8–88.9 · 11PP VARIANCE ACROSS RUNS · $0.064/RUN
Δ +18.5
PERCENTAGE POINTS
Trial Record
every pedestal is one benchmark task · a crystal = kernel-verified · shards = the judge said no and nobody argued · full protocol + repro in the Findings Portal (EXP-024 / EXP-025)