[{"id":"agent-loop-cells","title":"Every single-call and agent-loop cell","columns":{"$k":["key","label","unit"],"$r":[["config","Configuration","text"],["origin","Cell","text"],["strict","Strict passes","text"],["ci","95% interval","text"],["lenient","Answer correct (lenient)","text"],["formatMisses","Format misses","count"],["wrong","Wrong answers","count"],["noAnswer","No answer (error, time-out, turn limit)","count"],["medianTotal","Median total (s)","seconds"],["rangeTotal","Fastest to slowest (s)","text"],["medianTools","Median tool calls","count"],["usedTools","Attempts that ran a tool","text"],["medianTokens","Median total tokens","tokens"],["perPass","USD per strict pass (calculation)","usd"],["contaminated","Contaminated (left out)","count"]]},"rows":{"$k":["config","origin","strict","ci","lenient","formatMisses","wrong","noAnswer","medianTotal","rangeTotal","medianTools","usedTools","medianTokens","perPass","contaminated"],"$r":[["Claude Haiku 4.5 (single call) · Claude Code","reference (hard head-to-head)","11/24","28% to 65%","16/24",5,8,0,39.01,"15.3 to 75.1",0,"tools off",9038,0.0672,0],["Claude Haiku 4.5 (agent loop) · Claude Code","new run","13/24","35% to 72%","20/24",7,4,0,56.77,"24.5 to 223.7",3,"24/24",78432,0.14225,0],["Claude Sonnet 5.5 (single call) · Claude Code","reference (hard head-to-head)","24/24","86% to 100%","24/24",0,0,0,7.75,"2.3 to 34.8",0,"tools off",3380,0.01435,0],["Claude Sonnet 5.5 (agent loop) · Claude Code","new run","16/16","81% to 100%","16/16",0,0,0,7.41,"2.7 to 24.2",0,"3/16",10483,0.02746,0],["GPT-6 Luna (single call) · Codex CLI","new run","10/16","39% to 82%","11/16",1,5,0,5.16,"3.6 to 11.3",0,"tools off",11954,0.00116,0],["GPT-6 Luna (agent loop) · Codex CLI","new run","12/14","60% to 96%","12/14",0,2,0,9.32,"3.8 to 15.9",0,"2/14",16058,0.00099,2]]}},{"id":"agent-loop-audit","title":"Sandbox audit of every agent-loop attempt","columns":{"$k":["key","label","unit"],"$r":[["config","Configuration","text"],["attempts","Attempts","count"],["outsideEdits","Edits outside the work folder that ran","count"],["refused","Outside edit or read attempts refused before they ran","count"],["contaminated","Contaminated attempts (read outside the work folder)","count"],["which","Which (task, repetition)","text"],["timedOut","Time-outs (10 min)","count"],["maxTurns","Stopped at 30 turns","count"]]},"rows":{"$k":["config","attempts","outsideEdits","refused","contaminated","which","timedOut","maxTurns"],"$r":[["Claude Haiku 4.5 (agent loop) · Claude Code",24,0,2,0,"none",0,0],["Claude Sonnet 5.5 (agent loop) · Claude Code",16,0,0,0,"none",0,0],["GPT-6 Luna (agent loop) · Codex CLI",16,0,0,2,"DST day-length fix r1, DST day-length fix r2",0,0]]}}]