[{"id":"structured-output-cells","title":"Every cell: configuration by prompt","columns":{"$k":["key","label","unit"],"$r":[["config","Configuration","text"],["prompt","Prompt","text"],["strict","Strict passes","text"],["ci","95% interval","text"],["formatMisses","Format misses","count"],["wrong","Wrong values","count"],["errors","Errors","count"],["fenced","Replies in a code fence","count"],["completedN","Completed calls (n for time and tokens)","count"],["medianTotal","Median time (s)","seconds"],["timeRange","Time range (s, not an interval)","text"],["medianOutput","Median output tokens","tokens"],["outputRange","Output token range","text"],["medianReasoning","Median reasoning tokens","tokens"],["reasoningRange","Reasoning token range","text"]]},"rows":{"$k":["config","prompt","strict","ci","formatMisses","wrong","errors","fenced","medianTotal","timeRange","medianOutput","outputRange","medianReasoning","reasoningRange","completedN"],"$r":[["Claude Haiku 4.5 (instructions) · Claude Code","Order to JSON","0/8","0% to 32%",8,0,0,8,7.01,"5.67 to 9.8",897,"698 to 1134",758.5,"559 to 996",8],["Claude Haiku 4.5 (instructions) · Claude Code","Invoice to JSON","0/8","0% to 32%",8,0,0,8,9.92,"7.19 to 13.03",1280,"1088 to 1636",1047,"856 to 1403",8],["Claude Haiku 4.5 (instructions) · Claude Code","Meeting notes to action items","0/8","0% to 32%",1,7,0,8,10.34,"8.2 to 17",1235,"991 to 2059",1030,"761 to 1830",8],["Claude Haiku 4.5 (JSON schema) · Claude Code","Order to JSON","8/8","68% to 100%",0,0,0,0,7.19,"6.45 to 10.5",791.5,"716 to 1222",586,"522 to 1010",8],["Claude Haiku 4.5 (JSON schema) · Claude Code","Invoice to JSON","8/8","68% to 100%",0,0,0,0,9.39,"7.7 to 9.91",1120.5,"909 to 1226",788,"576 to 893",8],["Claude Haiku 4.5 (JSON schema) · Claude Code","Meeting notes to action items","2/8","7% to 59%",0,6,0,0,9.27,"5.9 to 12.23",1164.5,"806 to 1462",890.5,"532 to 1187",8],["Claude Sonnet 5.5 (instructions) · Claude Code","Order to JSON","4/4","51% to 100%",0,0,0,0,3.27,"2.67 to 3.8",166.5,"154 to 175",66.5,"54 to 75",4],["Claude Sonnet 5.5 (instructions) · Claude Code","Invoice to JSON","4/4","51% to 100%",0,0,0,0,3.65,"3.32 to 4.12",367.5,"362 to 373",181.5,"176 to 187",4],["Claude Sonnet 5.5 (instructions) · Claude Code","Meeting notes to action items","4/4","51% to 100%",0,0,0,0,3.7,"3.49 to 3.93",417.5,"392 to 456",239.5,"214 to 278",4],["Claude Sonnet 5.5 (JSON schema) · Claude Code","Order to JSON","4/4","51% to 100%",0,0,0,0,3.22,"2.95 to 6.14",273.5,"273 to 275",64.5,"64 to 66",4],["Claude Sonnet 5.5 (JSON schema) · Claude Code","Invoice to JSON","4/4","51% to 100%",0,0,0,0,4.37,"4.2 to 4.54",503.5,"495 to 541",76,"0 to 154",4],["Claude Sonnet 5.5 (JSON schema) · Claude Code","Meeting notes to action items","4/4","51% to 100%",0,0,0,0,4.25,"3.9 to 5.35",423.5,"408 to 498",185.5,"170 to 260",4],["GPT-6.1 Sol (low, instructions) · Codex CLI","Order to JSON","4/4","51% to 100%",0,0,0,0,6.21,"4.2 to 8.24",92,"69 to 93",21,"0 to 22",4],["GPT-6.1 Sol (low, instructions) · Codex CLI","Invoice to JSON","4/4","51% to 100%",0,0,0,0,10.19,"6.75 to 12.27",248,"235 to 259",49,"36 to 60",4],["GPT-6.1 Sol (low, instructions) · Codex CLI","Meeting notes to action items","4/4","51% to 100%",0,0,0,0,5.01,"4.71 to 5.81",117,"117 to 117",0,"0 to 0",4],["GPT-6.1 Sol (low, JSON schema) · Codex CLI","Order to JSON","4/4","51% to 100%",0,0,0,0,5.73,"4.62 to 20.97",99.5,"98 to 102",22.5,"21 to 25",4],["GPT-6.1 Sol (low, JSON schema) · Codex CLI","Invoice to JSON","4/4","51% to 100%",0,0,0,0,6.13,"5.77 to 6.37",172.5,"166 to 181",40.5,"34 to 49",4],["GPT-6.1 Sol (low, JSON schema) · Codex CLI","Meeting notes to action items","4/4","51% to 100%",0,0,0,0,5.76,"4.87 to 6.51",123,"123 to 123",0,"0 to 0",4]]}},{"id":"structured-output-pairs","title":"Paired view: the same prompt and repetition, asked both ways","columns":{"$k":["key","label","unit"],"$r":[["model","Model and route","text"],["pairs","Pairs","count"],["both","Both passed","count"],["onlyInstructions","Only instructions passed","count"],["onlySchema","Only schema passed","count"],["neither","Neither passed","count"],["mcnemarP","Exact two-sided McNemar p (calculation)","score"]]},"rows":{"$k":["model","pairs","both","onlyInstructions","onlySchema","neither","mcnemarP"],"$r":[["Claude Haiku 4.5 · Claude Code",24,0,0,18,6,0.00000762939453125],["Claude Sonnet 5.5 · Claude Code",12,12,0,0,0,1],["GPT-6.1 Sol · Codex CLI",12,12,0,0,0,1]]}}]