{ "video": "https://modelfatigue.news/v/decisions-api/", "numbers": [ { "id": "n-test", "what": "Banking77 test messages every model answered", "value": 3080, "shown": "3,080", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "n-reasons", "what": "Reasons a message can be sorted into", "value": 77, "shown": "77", "whose": "Model Fatigue", "source": "Our frozen protocol for the bank-message test (FREEZE.md)", "url": null, "read": "2026-09-30", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "n-options", "what": "Options in each question: the 77 reasons plus none of these", "value": 78, "shown": "78", "whose": "Model Fatigue", "source": "Our results write-up (RESULTS.md, \"OpenAI's Decisions API (added 2026-10-07)\")", "url": null, "read": "2026-10-07", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "n-shots", "what": "Already-sorted example messages shown with each message", "value": 5, "shown": "5", "whose": "Model Fatigue", "source": "Our frozen protocol for the bank-message test (FREEZE.md)", "url": null, "read": "2026-09-30", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dev-n", "what": "Messages in the separate set every model's cut-off was set on (not test messages)", "value": 573, "shown": "573", "whose": "Model Fatigue", "source": "Our pass of the Decisions API over the separate set of 573 messages its cut-off was set on (not test messages)", "url": null, "read": "2026-10-07T01:09", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-acc", "what": "the Decisions API, share of the test messages sorted correctly", "value": 0.935064935064935, "shown": "93.5%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-acc-lo", "what": "the Decisions API, share right, 95% interval, low", "value": 0.9258090740348918, "shown": "92.6", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-acc-hi", "what": "the Decisions API, share right, 95% interval, high", "value": 0.9432368588042007, "shown": "94.3", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-none", "what": "the Decisions API, times it answered none of these", "value": 4, "shown": "4", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-p50", "what": "the Decisions API, median time per request from Berlin", "value": 0.23868, "shown": "239 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-p95", "what": "the Decisions API, 95th percentile time per request from Berlin", "value": 0.37005249999999973, "shown": "370 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-p99", "what": "the Decisions API, 99th percentile time per request from Berlin", "value": 0.5943527000000001, "shown": "594 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-per1k", "what": "the Decisions API, dollars per 1,000 decisions (price list × the tokens each response reports)", "value": 0.11360217532467533, "shown": "$0.114", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-thr", "what": "the Decisions API, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.99, "shown": "0.99", "whose": "Model Fatigue", "source": "Our pass of the Decisions API over the separate set of 573 messages its cut-off was set on (not test messages)", "url": null, "read": "2026-10-07T01:09", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-cov", "what": "the Decisions API, share of the test messages it handled without a person", "value": 0.8396103896103896, "shown": "84.0%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-risk", "what": "the Decisions API, wrong among the messages it handled without a person", "value": 0.021268368136117557, "shown": "2.13%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-ref", "what": "the Decisions API, messages per 1,000 sent to a person", "value": 160, "shown": "160", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-ece", "what": "the Decisions API, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.04079220779220774, "shown": "0.041", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-auroc", "what": "the Decisions API, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.8459574652777778, "shown": "0.85", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-acc", "what": "Jev, share of the test messages sorted correctly", "value": 0.938961038961039, "shown": "93.9%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-acc-lo", "what": "Jev, share right, 95% interval, low", "value": 0.9299469172345147, "shown": "93.0", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-acc-hi", "what": "Jev, share right, 95% interval, high", "value": 0.9468815164956743, "shown": "94.7", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-none", "what": "Jev, times it answered none of these", "value": 13, "shown": "13", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-p50", "what": "Jev, median time per request from Berlin", "value": 0.24234, "shown": "242 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-p95", "what": "Jev, 95th percentile time per request from Berlin", "value": 0.30670649999999994, "shown": "307 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-p99", "what": "Jev, 99th percentile time per request from Berlin", "value": 0.38633740000000005, "shown": "386 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-per1k", "what": "Jev, dollars per 1,000 decisions (price list × the tokens each response reports)", "value": 0.07753194545454546, "shown": "$0.078", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-thr", "what": "Jev, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.99, "shown": "0.99", "whose": "Model Fatigue", "source": "Our passes of Jev and GPT-6 Luna over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-09-30T21:28", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-cov", "what": "Jev, share of the test messages it handled without a person", "value": 0.8409090909090909, "shown": "84.1%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-risk", "what": "Jev, wrong among the messages it handled without a person", "value": 0.020463320463320462, "shown": "2.05%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-ref", "what": "Jev, messages per 1,000 sent to a person", "value": 159, "shown": "159", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-ece", "what": "Jev, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.03520129870129873, "shown": "0.035", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-auroc", "what": "Jev, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.8457354845354758, "shown": "0.85", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-acc", "what": "GPT-6 Luna through the Responses API, share of the test messages sorted correctly", "value": 0.9415584415584416, "shown": "94.2%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-acc-lo", "what": "GPT-6 Luna through the Responses API, share right, 95% interval, low", "value": 0.9327108134839683, "shown": "93.3", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-acc-hi", "what": "GPT-6 Luna through the Responses API, share right, 95% interval, high", "value": 0.9493059541736182, "shown": "94.9", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-none", "what": "GPT-6 Luna through the Responses API, times it answered none of these", "value": 5, "shown": "5", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-p50", "what": "GPT-6 Luna through the Responses API, median time per request from Berlin", "value": 1.04669, "shown": "1,047 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-p95", "what": "GPT-6 Luna through the Responses API, 95th percentile time per request from Berlin", "value": 1.6039169999999994, "shown": "1,604 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-p99", "what": "GPT-6 Luna through the Responses API, 99th percentile time per request from Berlin", "value": 2.372687800000001, "shown": "2,373 ms", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-p50s", "what": "GPT-6 Luna through the Responses API, median time per request from Berlin, in seconds", "value": 1.04669, "shown": "1.05 s", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-p95s", "what": "GPT-6 Luna through the Responses API, 95th percentile time per request from Berlin, in seconds", "value": 1.6039169999999994, "shown": "1.60 s", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-p99s", "what": "GPT-6 Luna through the Responses API, 99th percentile time per request from Berlin, in seconds", "value": 2.372687800000001, "shown": "2.37 s", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-per1k", "what": "GPT-6 Luna through the Responses API, dollars per 1,000 decisions (price list × the tokens each response reports)", "value": 0.14332879058441558, "shown": "$0.143", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-thr", "what": "GPT-6 Luna through the Responses API, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.9, "shown": "0.90", "whose": "Model Fatigue", "source": "Our passes of Jev and GPT-6 Luna over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-09-30T21:28", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-cov", "what": "GPT-6 Luna through the Responses API, share of the test messages it handled without a person", "value": 0.9087662337662338, "shown": "90.9%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-risk", "what": "GPT-6 Luna through the Responses API, wrong among the messages it handled without a person", "value": 0.02750982493747767, "shown": "2.75%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-ref", "what": "GPT-6 Luna through the Responses API, messages per 1,000 sent to a person", "value": 91, "shown": "91", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-ece", "what": "GPT-6 Luna through the Responses API, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.02408441558441566, "shown": "0.024", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-auroc", "what": "GPT-6 Luna through the Responses API, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.8631024904214559, "shown": "0.86", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-acc", "what": "Clef, share of the test messages sorted correctly", "value": 0.9506493506493506, "shown": "95.1%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-acc-lo", "what": "Clef, share right, 95% interval, low", "value": 0.9424225746200366, "shown": "94.2", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-acc-hi", "what": "Clef, share right, 95% interval, high", "value": 0.9577533617834414, "shown": "95.8", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-none", "what": "Clef, times it answered none of these", "value": 0, "shown": "0", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-p50", "what": "Clef, median time per request from Berlin", "value": 0.562495, "shown": "562 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-p95", "what": "Clef, 95th percentile time per request from Berlin", "value": 1.1474874999999964, "shown": "1,147 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-p99", "what": "Clef, 99th percentile time per request from Berlin", "value": 2.0007208000000003, "shown": "2,001 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-per1k", "what": "Clef, dollars per 1,000 decisions (price list × the tokens each response reports)", "value": 0.4797513506493506, "shown": "$0.480", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-thr", "what": "Clef, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.8034, "shown": "0.80", "whose": "Model Fatigue", "source": "Our passes of Clef, Clef-flash, Kev-27B and Kev-9B over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-10-02T15:38", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-cov", "what": "Clef, share of the test messages it handled without a person", "value": 0.9321428571428572, "shown": "93.2%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-risk", "what": "Clef, wrong among the messages it handled without a person", "value": 0.024033437826541274, "shown": "2.40%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-ref", "what": "Clef, messages per 1,000 sent to a person", "value": 68, "shown": "68", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-ece", "what": "Clef, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.029707012987013053, "shown": "0.030", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-auroc", "what": "Clef, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.9001181873741732, "shown": "0.90", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-acc", "what": "Clef-flash, share of the test messages sorted correctly", "value": 0.9516233766233766, "shown": "95.2%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-acc-lo", "what": "Clef-flash, share right, 95% interval, low", "value": 0.9434670440163816, "shown": "94.3", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-acc-hi", "what": "Clef-flash, share right, 95% interval, high", "value": 0.9586545176098705, "shown": "95.9", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-none", "what": "Clef-flash, times it answered none of these", "value": 0, "shown": "0", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-p50", "what": "Clef-flash, median time per request from Berlin", "value": 0.19639, "shown": "196 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-p95", "what": "Clef-flash, 95th percentile time per request from Berlin", "value": 1.013879999999998, "shown": "1,014 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-p99", "what": "Clef-flash, 99th percentile time per request from Berlin", "value": 2.1720025, "shown": "2,172 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-per1k", "what": "Clef-flash, dollars per 1,000 decisions (price list × the tokens each response reports)", "value": 0.17990675649350651, "shown": "$0.180", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-thr", "what": "Clef-flash, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.7323, "shown": "0.73", "whose": "Model Fatigue", "source": "Our passes of Clef, Clef-flash, Kev-27B and Kev-9B over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-10-02T15:38", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-cov", "what": "Clef-flash, share of the test messages it handled without a person", "value": 0.8168831168831169, "shown": "81.7%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-risk", "what": "Clef-flash, wrong among the messages it handled without a person", "value": 0.021462639109697933, "shown": "2.15%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-ref", "what": "Clef-flash, messages per 1,000 sent to a person", "value": 183, "shown": "183", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-ece", "what": "Clef-flash, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.14252301948051946, "shown": "0.143", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-auroc", "what": "Clef-flash, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.8190472592216047, "shown": "0.82", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-acc", "what": "Kev-27B, share of the test messages sorted correctly", "value": 0.9383116883116883, "shown": "93.8%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-acc-lo", "what": "Kev-27B, share right, 95% interval, low", "value": 0.9292566262741926, "shown": "92.9", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-acc-hi", "what": "Kev-27B, share right, 95% interval, high", "value": 0.9462747239741469, "shown": "94.6", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-none", "what": "Kev-27B, times it answered none of these", "value": 21, "shown": "21", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-p50", "what": "Kev-27B, median time per request from Berlin", "value": 0.528605, "shown": "529 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-p95", "what": "Kev-27B, 95th percentile time per request from Berlin", "value": 0.5508195, "shown": "551 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-p99", "what": "Kev-27B, 99th percentile time per request from Berlin", "value": 0.5976951000000001, "shown": "598 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-thr", "what": "Kev-27B, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.7928, "shown": "0.79", "whose": "Model Fatigue", "source": "Our passes of Clef, Clef-flash, Kev-27B and Kev-9B over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-10-02T15:38", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-cov", "what": "Kev-27B, share of the test messages it handled without a person", "value": 0.9103896103896104, "shown": "91.0%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-risk", "what": "Kev-27B, wrong among the messages it handled without a person", "value": 0.02282453637660485, "shown": "2.28%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-ref", "what": "Kev-27B, messages per 1,000 sent to a person", "value": 90, "shown": "90", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-ece", "what": "Kev-27B, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.014207987012986964, "shown": "0.014", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-auroc", "what": "Kev-27B, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.920481697322892, "shown": "0.92", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-acc", "what": "Kev-9B, share of the test messages sorted correctly", "value": 0.9133116883116883, "shown": "91.3%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-acc-lo", "what": "Kev-9B, share right, 95% interval, low", "value": 0.9028523242309728, "shown": "90.3", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-acc-hi", "what": "Kev-9B, share right, 95% interval, high", "value": 0.922741311966165, "shown": "92.3", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-none", "what": "Kev-9B, times it answered none of these", "value": 110, "shown": "110", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-p50", "what": "Kev-9B, median time per request from Berlin", "value": 0.436515, "shown": "437 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-p95", "what": "Kev-9B, 95th percentile time per request from Berlin", "value": 0.5482069999999997, "shown": "548 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-p99", "what": "Kev-9B, 99th percentile time per request from Berlin", "value": 0.6647247000000001, "shown": "665 ms", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-thr", "what": "Kev-9B, cut-off on its stated confidence, set on the separate set for a 2% error target", "value": 0.8127, "shown": "0.81", "whose": "Model Fatigue", "source": "Our passes of Clef, Clef-flash, Kev-27B and Kev-9B over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-10-02T15:38", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-cov", "what": "Kev-9B, share of the test messages it handled without a person", "value": 0.826948051948052, "shown": "82.7%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-risk", "what": "Kev-9B, wrong among the messages it handled without a person", "value": 0.024734982332155476, "shown": "2.47%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-ref", "what": "Kev-9B, messages per 1,000 sent to a person", "value": 173, "shown": "173", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-ece", "what": "Kev-9B, calibration error of its stated confidence (ECE; lower is closer)", "value": 0.03842435064935068, "shown": "0.038", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-auroc", "what": "Kev-9B, how well its confidence separates its right answers from its wrong ones (AUROC)", "value": 0.9057099794826321, "shown": "0.91", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-risk-lo", "what": "the Decisions API, wrong among handled, 95% interval, low", "value": 0.016376577961557774, "shown": "1.6%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-risk-hi", "what": "the Decisions API, wrong among handled, 95% interval, high", "value": 0.027580396086584788, "shown": "2.8%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-risk-lo", "what": "Jev, wrong among handled, 95% interval, low", "value": 0.015678870856910845, "shown": "1.6%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-risk-hi", "what": "Jev, wrong among handled, 95% interval, high", "value": 0.026668202302837993, "shown": "2.7%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-wrong-let-through", "what": "The Decisions API's wrong answers at 0.99 or above, which the 2% cut-off would let through without a person", "value": 55, "shown": "55", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-repeat-same", "what": "Messages sent to the Decisions API a second time that got the same answer", "value": 200, "shown": "200", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-repeat-n", "what": "Messages sent to the Decisions API a second time", "value": 200, "shown": "200", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dvj-jr", "what": "Messages Jev got right and the Decisions API got wrong", "value": 38, "shown": "38", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dvj-dr", "what": "Messages the Decisions API got right and Jev got wrong", "value": 26, "shown": "26", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "lvj-jr", "what": "Messages Jev got right and Luna got wrong", "value": 23, "shown": "23", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "lvj-lr", "what": "Messages Luna got right and Jev got wrong", "value": 31, "shown": "31", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "ph-lr", "what": "Messages Luna got right and the Decisions API got wrong (post hoc)", "value": 41, "shown": "41", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "ph-dr", "what": "Messages the Decisions API got right and Luna got wrong (post hoc)", "value": 21, "shown": "21", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "speedup", "what": "Luna's median time over the Decisions API's, same messages, same night (our arithmetic)", "value": 4.385327635327635, "shown": "4.4×", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "openai-150ms", "what": "OpenAI's time per request for the Decisions API (launch clip)", "value": 0.15, "shown": "150 ms", "whose": "OpenAI", "source": "OpenAI Developers' Decisions API launch clip, its on-screen text", "url": "https://x.com/OpenAIDevs/status/2105003318917697873", "read": "2026-09-29", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "openai-1-6s", "what": "OpenAI's time per request for the Responses API (launch clip)", "value": 1.6, "shown": "1.6 s", "whose": "OpenAI", "source": "OpenAI Developers' Decisions API launch clip, its on-screen text", "url": "https://x.com/OpenAIDevs/status/2105003318917697873", "read": "2026-09-29", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "openai-10x", "what": "OpenAI's speed-up, Decisions API over the Responses API (launch clip)", "value": 10, "shown": "10×", "whose": "OpenAI", "source": "OpenAI Developers' Decisions API launch clip, its on-screen text", "url": "https://x.com/OpenAIDevs/status/2105003318917697873", "read": "2026-09-29", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "openai-10k", "what": "Customer requests in OpenAI's launch clip", "value": 10000, "shown": "10,000", "whose": "OpenAI", "source": "OpenAI Developers' Decisions API launch clip, its on-screen text", "url": "https://x.com/OpenAIDevs/status/2105003318917697873", "read": "2026-09-29", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "docs-10x", "what": "OpenAI's speed-up in the Decisions guide", "value": 10, "shown": "10×", "whose": "OpenAI", "source": "OpenAI, the Decisions API guide", "url": "https://developers.openai.com/api/docs/guides/decisions", "read": "2026-10-07T01:05", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "under-150", "what": "Decisions calls that came back to Berlin in under 150 ms", "value": 0, "shown": "0", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "fastest", "what": "The fastest Decisions call from Berlin", "value": 0.16462, "shown": "165 ms", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "floor-start", "what": "The trip from Berlin to OpenAI's API and back with no model work, median, before the run", "value": 0.1661, "shown": "166 ms", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "floor-end", "what": "The same trip, median, after the run", "value": 0.1666, "shown": "167 ms", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-minus-floor", "what": "The Decisions API's median time minus the trip, before the run (our arithmetic)", "value": 0.07308, "shown": "73 ms", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "run-window", "what": "When the timed run went out (first network-floor check to the last)", "value": "23:13 to 00:39 UTC", "shown": "23:13 to 00:39 UTC", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "win-jev", "what": "When Jev's test pass ran", "value": "23:14 to 23:27 UTC", "shown": "23:14 to 23:27 UTC", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "win-decisions", "what": "When the Decisions API's test pass ran", "value": "23:27 to 23:40 UTC", "shown": "23:27 to 23:40 UTC", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "win-luna", "what": "When Luna's test pass ran", "value": "23:41 to 00:38 UTC", "shown": "23:41 to 00:38 UTC", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "rate-dec", "what": "The Decisions API's price per million input tokens (nothing else billed)", "value": 0.1, "shown": "$0.10", "whose": "OpenAI", "source": "OpenAI, the Decisions API guide", "url": "https://developers.openai.com/api/docs/guides/decisions", "read": "2026-10-07T01:05", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "rate-jev", "what": "Jev's price per million input tokens (TypeSafe's rate card)", "value": 0.042, "shown": "$0.042", "whose": "Model Fatigue", "source": "Our frozen protocol for the bank-message test (FREEZE.md)", "url": null, "read": "2026-09-30", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-tok", "what": "the Decisions API, input tokens per request as its service counted them, average", "value": 1136.0217532467532, "shown": "1,136", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-tok", "what": "Jev, input tokens per request as its service counted them, average", "value": 1845.9987012987012, "shown": "1,846", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "luna-tok", "what": "GPT-6 Luna through the Responses API, input tokens per request as its service counted them, average", "value": 1066.0217532467532, "shown": "1,066", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-tok", "what": "Clef, input tokens per request as its service counted them, average", "value": 1998.963961038961, "shown": "1,999", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-tok", "what": "Clef-flash, input tokens per request as its service counted them, average", "value": 1998.963961038961, "shown": "1,999", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "cost-ratio", "what": "The Decisions API's cost per decision over Jev's (our arithmetic)", "value": 1.4652305531437582, "shown": "1.5×", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-run-cost", "what": "The Decisions API test pass, all 3,080 messages, at the price list", "value": 0.3498947, "shown": "$0.35", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "n-99", "what": "Decisions answers stated at 0.99 or 1.00", "value": 2586, "shown": "2,586", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "n-100", "what": "Decisions answers stated at 1.00", "value": 2146, "shown": "2,146", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "bin90-n", "what": "Decisions answers stated at 0.90 or more", "value": 2902, "shown": "2,902", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "bin90-mean", "what": "The Decisions API's average stated confidence on those answers", "value": 0.9940902825637491, "shown": "0.994", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "bin90-right", "what": "Share of those answers that were right", "value": 0.9607167470709855, "shown": "96.1%", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "top-dec-dev", "what": "The Decisions API, separate set: wrong among its answers stated at 1.00", "value": 0.010582010582010581, "shown": "1.06%", "whose": "Model Fatigue", "source": "Our pass of the Decisions API over the separate set of 573 messages its cut-off was set on (not test messages)", "url": null, "read": "2026-10-07T01:09", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "top-jev-dev", "what": "Jev, separate set: wrong among its answers stated at 1.00", "value": 0.018561484918793503, "shown": "1.86%", "whose": "Model Fatigue", "source": "Our passes of Jev and GPT-6 Luna over the separate set of 573 messages their cut-offs were set on (not test messages)", "url": null, "read": "2026-09-30T21:28", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "top-dec-test", "what": "The Decisions API, test: wrong among its answers stated at 1.00", "value": 0.015843429636533086, "shown": "1.58%", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "top-jev-test", "what": "Jev, test: wrong among its answers stated at 1.00", "value": 0.01589242053789731, "shown": "1.59%", "whose": "Model Fatigue", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "target", "what": "The error target every cut-off was set for: wrong answers among those a model handles alone", "value": 0.02, "shown": "2%", "whose": "Model Fatigue", "source": "Our protocol addendum for the Decisions API, committed before the first test call (commit 23057db1)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "target-strict", "what": "The stricter target no setting could meet for the Decisions API or Jev", "value": 0.01, "shown": "1%", "whose": "Model Fatigue", "source": "Our results write-up (RESULTS.md, \"OpenAI's Decisions API (added 2026-10-07)\")", "url": null, "read": "2026-10-07", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-a-decisions", "what": "Message A: the Decisions API's probability for the answer it picked", "value": 0.36, "shown": "36%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-a-decisions-2", "what": "Message A: the Decisions API's probability for its second answer", "value": 0.19, "shown": "19%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-a-jev", "what": "Message A: Jev's probability for the answer it picked", "value": 0.6, "shown": "60%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-a-jev-2", "what": "Message A: Jev's probability for its second answer", "value": 0.22, "shown": "22%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-a-clef", "what": "Message A: Clef's probability for the answer it picked", "value": 0.5761, "shown": "58%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-a-clef-2", "what": "Message A: Clef's probability for its second answer", "value": 0.1003, "shown": "10%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-b-decisions", "what": "Message B: the Decisions API's probability for the answer it picked", "value": 0.84, "shown": "84%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-b-decisions-2", "what": "Message B: the Decisions API's probability for its second answer", "value": 0.06, "shown": "6%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-b-jev", "what": "Message B: Jev's probability for the answer it picked", "value": 0.48, "shown": "48%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-b-jev-2", "what": "Message B: Jev's probability for its second answer", "value": 0.36, "shown": "36%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-b-clef", "what": "Message B: Clef's probability for the answer it picked", "value": 0.34, "shown": "34%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-b-clef-2", "what": "Message B: Clef's probability for its second answer", "value": 0.2386, "shown": "24%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-c-decisions", "what": "Message C: the Decisions API's probability for the answer it picked", "value": 0.99, "shown": "99%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-c-decisions-2", "what": "Message C: the Decisions API's probability for its second answer", "value": 0.01, "shown": "1%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-c-jev", "what": "Message C: Jev's probability for the answer it picked", "value": 0.94, "shown": "94%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-c-jev-2", "what": "Message C: Jev's probability for its second answer", "value": 0.06, "shown": "6%", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-c-clef", "what": "Message C: Clef's probability for the answer it picked", "value": 0.7409, "shown": "74%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "sbs-c-clef-2", "what": "Message C: Clef's probability for its second answer", "value": 0.166, "shown": "17%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "all-right", "what": "Test messages the Decisions API, Jev and Clef all got right", "value": 2841, "shown": "2,841", "whose": "Model Fatigue, arithmetic on our runs", "source": "Our arithmetic across the two nights' passes: the Decisions API, Jev and Luna on 7 October, Clef, Clef-flash, the Kev models and every model's message-alone pass but the Decisions API's on 2 October", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "all-wrong", "what": "Test messages all three got wrong", "value": 125, "shown": "125", "whose": "Model Fatigue, arithmetic on our runs", "source": "Our arithmetic across the two nights' passes: the Decisions API, Jev and Luna on 7 October, Clef, Clef-flash, the Kev models and every model's message-alone pass but the Decisions API's on 2 October", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "opt-max", "what": "The most options the Decisions API accepted in one question", "value": 255, "shown": "255", "whose": "Model Fatigue", "source": "Our calls to the Decisions API with 1 to 1,000 options, to find its limit", "url": null, "read": "2026-10-07T01:08", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "opt-refused", "what": "Options in the question the API refused as too many", "value": 256, "shown": "256", "whose": "Model Fatigue", "source": "Our calls to the Decisions API with 1 to 1,000 options, to find its limit", "url": null, "read": "2026-10-07T01:08", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "opt-min", "what": "The fewest options the API accepted in one question", "value": 2, "shown": "2", "whose": "Model Fatigue", "source": "Our calls to the Decisions API with 1 to 1,000 options, to find its limit", "url": null, "read": "2026-10-07T01:08", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-di", "what": "the Decisions API, share right with the message alone (the Decision Index's question)", "value": 0.7769480519480519, "shown": "77.7%", "whose": "Model Fatigue", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-di-f1", "what": "the Decisions API, macro-F1 with the message alone", "value": 0.7694673649101087, "shown": "76.95", "whose": "Model Fatigue", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dec-gain", "what": "the Decisions API, share right on our question (five examples) minus on the message alone, points", "value": 15.81, "shown": "+15.8 points", "whose": "Model Fatigue, arithmetic on our runs", "source": "Our arithmetic across the two nights' passes: the Decisions API, Jev and Luna on 7 October, Clef, Clef-flash, the Kev models and every model's message-alone pass but the Decisions API's on 2 October", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-di", "what": "Jev, share right with the message alone (the Decision Index's question)", "value": 0.8071428571428572, "shown": "80.7%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-di-f1", "what": "Jev, macro-F1 with the message alone", "value": 0.8000046642892913, "shown": "80.00", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "jev-gain", "what": "Jev, share right on our question (five examples) minus on the message alone, points", "value": 13.18, "shown": "+13.2 points", "whose": "Model Fatigue, arithmetic on our runs", "source": "Our arithmetic across the two nights' passes: the Decisions API, Jev and Luna on 7 October, Clef, Clef-flash, the Kev models and every model's message-alone pass but the Decisions API's on 2 October", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-di", "what": "Clef, share right with the message alone (the Decision Index's question)", "value": 0.9422077922077922, "shown": "94.2%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-di-f1", "what": "Clef, macro-F1 with the message alone", "value": 0.9420316998167463, "shown": "94.20", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "clef-gain", "what": "Clef, share right on our question (five examples) minus on the message alone, points", "value": 0.84, "shown": "+0.8 points", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-di", "what": "Clef-flash, share right with the message alone (the Decision Index's question)", "value": 0.9087662337662338, "shown": "90.9%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-di-f1", "what": "Clef-flash, macro-F1 with the message alone", "value": 0.9084736076902837, "shown": "90.85", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-gain", "what": "Clef-flash, share right on our question (five examples) minus on the message alone, points", "value": 4.29, "shown": "+4.3 points", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-di", "what": "Kev-27B, share right with the message alone (the Decision Index's question)", "value": 0.8681818181818182, "shown": "86.8%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-di-f1", "what": "Kev-27B, macro-F1 with the message alone", "value": 0.8640871097889219, "shown": "86.41", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k27-gain", "what": "Kev-27B, share right on our question (five examples) minus on the message alone, points", "value": 7.01, "shown": "+7.0 points", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-di", "what": "Kev-9B, share right with the message alone (the Decision Index's question)", "value": 0.8331168831168831, "shown": "83.3%", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-di-f1", "what": "Kev-9B, macro-F1 with the message alone", "value": 0.8302597361284337, "shown": "83.03", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "k9-gain", "what": "Kev-9B, share right on our question (five examples) minus on the message alone, points", "value": 8.02, "shown": "+8.0 points", "whose": "Model Fatigue", "source": "Our test of Clef, Clef-flash, Kev-27B and Kev-9B on the same messages (the Clef vs Jev video's run)", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "di-options", "what": "Options in the Decision Index's question (the reasons, without none of these)", "value": 77, "shown": "77", "whose": "Model Fatigue", "source": "Our results write-up (RESULTS.md, \"OpenAI's Decisions API (added 2026-10-07)\")", "url": null, "read": "2026-10-07", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "cf-clef", "what": "Clef on BANKING77 in Cloudflare's table (macro-F1)", "value": 94.2, "shown": "94.20", "whose": "Cloudflare", "source": "Cloudflare, launch post for Clef and Clef-flash", "url": "https://blog.cloudflare.com/clef-decision-models/", "read": "2026-10-02T15:31", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "cf-jev", "what": "Jev on BANKING77 in Cloudflare's table (macro-F1)", "value": 79.74, "shown": "79.74", "whose": "Cloudflare", "source": "Cloudflare, launch post for Clef and Clef-flash", "url": "https://blog.cloudflare.com/clef-decision-models/", "read": "2026-10-02T15:31", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "mem-n", "what": "Messages from the public collection's training part, asked as written and reworded", "value": 200, "shown": "200", "whose": "Model Fatigue", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "mem-orig", "what": "The Decisions API, share right on the original messages (no examples)", "value": 0.695, "shown": "69.5%", "whose": "Model Fatigue", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "mem-para", "what": "The Decisions API, share right on the same messages reworded", "value": 0.69, "shown": "69.0%", "whose": "Model Fatigue", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dvj", "what": "The Decisions API minus Jev, share right, same messages, same night", "value": -0.38961038961038963, "shown": "−0.39 points", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dvj-lo", "what": "The Decisions API minus Jev, 95% interval, low", "value": -0.9090909090909091, "shown": "−0.91", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "dvj-hi", "what": "The Decisions API minus Jev, 95% interval, high", "value": 0.12987012987012986, "shown": "+0.13", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "lvj", "what": "GPT-6 Luna through the Responses API minus Jev, share right, same messages, same night", "value": 0.2597402597402597, "shown": "+0.26 points", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "lvj-lo", "what": "Luna minus Jev, 95% interval, low", "value": -0.22727272727272727, "shown": "−0.23", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "lvj-hi", "what": "Luna minus Jev, 95% interval, high", "value": 0.7467532467532467, "shown": "+0.75", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "ph", "what": "Luna through the Responses API minus the Decisions API, same messages (post hoc: compared after the run, not in the frozen plan)", "value": 0.6493506493506493, "shown": "+0.65 points", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "ph-lo", "what": "Luna minus the Decisions API, 95% interval, low (post hoc)", "value": 0.16233766233766234, "shown": "+0.16", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "ph-hi", "what": "Luna minus the Decisions API, 95% interval, high (post hoc)", "value": 1.1688311688311688, "shown": "+1.17", "whose": "Model Fatigue", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "ph-ratio", "what": "Luna-only right answers per Decisions-only right answer (post hoc, our arithmetic)", "value": 1.9523809523809523, "shown": "2.0", "whose": "Model Fatigue, arithmetic on our run", "source": "Our test, the night the Decisions API opened: Jev, the Decisions API and GPT-6 Luna through the Responses API, 3,080 Banking77 test messages, five sorted examples each, 78 options, one request at a time from a Mac mini in Berlin", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "server-p50", "what": "Time OpenAI's servers report spending on each Decisions call (openai-processing-ms), median", "value": 0.065, "shown": "65 ms", "whose": "OpenAI, read from each response", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "server-p95", "what": "OpenAI's reported server time per Decisions call, 95th percentile", "value": 0.144, "shown": "144 ms", "whose": "OpenAI, read from each response", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "server-p99", "what": "OpenAI's reported server time per Decisions call, 99th percentile", "value": 0.3670500000000002, "shown": "367 ms", "whose": "OpenAI, read from each response", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "server-rest", "what": "Our time per Decisions call minus OpenAI's reported server time, median (our arithmetic)", "value": 0.17044, "shown": "170 ms", "whose": "Model Fatigue, arithmetic on our run", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "rate-ratio", "what": "The Decisions API's price per token over Jev's (our arithmetic)", "value": 2.380952380952381, "shown": "2.4×", "whose": "Model Fatigue, arithmetic on the two price lists", "source": "Our arithmetic on two price lists: OpenAI's Decisions guide (read 7 Oct 2026, 01:05 CEST) and Jev's rate card in our frozen protocol (30 Sep 2026)", "url": null, "read": "2026-10-07T01:05", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "tok-ratio", "what": "Jev's token count per request over the Decisions API's (our arithmetic)", "value": 1.6249677402944371, "shown": "1.6×", "whose": "Model Fatigue, arithmetic on our run", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "flash-cost-ratio", "what": "Clef-flash's cost per decision over Jev's (our arithmetic)", "value": 2.320421026955659, "shown": "2.3×", "whose": "Model Fatigue, arithmetic on our run", "source": "Our arithmetic across the two nights' passes: the Decisions API, Jev and Luna on 7 October, Clef, Clef-flash, the Kev models and every model's message-alone pass but the Decisions API's on 2 October", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "share-99", "what": "Share of Decisions answers stated at 0.99 or 1.00 (our arithmetic)", "value": 0.8396103896103896, "shown": "84%", "whose": "Model Fatigue, arithmetic on our run", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "bin90-share", "what": "Share of Decisions answers stated at 0.90 or more (our arithmetic)", "value": 0.9422077922077922, "shown": "94%", "whose": "Model Fatigue, arithmetic on our run", "source": "Our arithmetic on the raw call records (derived.json, this page's own file)", "url": null, "read": "2026-10-07T01:13", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "di-gap", "what": "Jev minus the Decisions API with the message alone (our arithmetic)", "value": 3.019480519480522, "shown": "3.0 points", "whose": "Model Fatigue, arithmetic on our run", "source": "Our arithmetic across the two nights' passes: the Decisions API, Jev and Luna on 7 October, Clef, Clef-flash, the Kev models and every model's message-alone pass but the Decisions API's on 2 October", "url": null, "read": "2026-10-02T16:02", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "cf-gap", "what": "Clef minus Jev on BANKING77 in Cloudflare's table (our arithmetic)", "value": 14.460000000000008, "shown": "14.5 points", "whose": "Model Fatigue, arithmetic on Cloudflare's numbers", "source": "Cloudflare, launch post for Clef and Clef-flash", "url": "https://blog.cloudflare.com/clef-decision-models/", "read": "2026-10-02T15:31", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "mem-gap", "what": "Originals minus rewordings (our arithmetic)", "value": 0.5, "shown": "+0.5 points", "whose": "Model Fatigue, arithmetic on our run", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "mem-lo", "what": "Originals minus rewordings, 95% interval, low", "value": -3.5000000000000004, "shown": "−3.5", "whose": "Model Fatigue, arithmetic on our run", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null }, { "id": "mem-hi", "what": "Originals minus rewordings, 95% interval, high", "value": 5.0, "shown": "+5.0", "whose": "Model Fatigue, arithmetic on our run", "source": "Our passes of the Decisions API without examples, and the memorisation check (about six requests at a time, from another machine; no times quoted from them)", "url": null, "read": "2026-10-07T02:43", "moves": false, "caveat": null, "now": null, "now_shown": null, "now_read": null } ] }