{
 "video": "https://modelfatigue.news/v/clef-vs-jev/",
 "numbers": [
  {
   "id": "n-test",
   "what": "BANKING77 test messages every model answered",
   "value": 3080,
   "shown": "3,080",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "n-intents",
   "what": "Reasons (intents) a message can be sorted into",
   "value": 77,
   "shown": "77",
   "whose": "Model Fatigue",
   "source": "Our dev-slice and example-pool counts (dev-slice/stats.json)",
   "url": null,
   "read": "2026-09-30",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "n-dev",
   "what": "Training messages held out to set each model's cut-off before the test",
   "value": 573,
   "shown": "573",
   "whose": "Model Fatigue",
   "source": "Our dev-slice and example-pool counts (dev-slice/stats.json)",
   "url": null,
   "read": "2026-09-30",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "n-pool",
   "what": "Labelled training messages the five examples are drawn from",
   "value": 9405,
   "shown": "9,405",
   "whose": "Model Fatigue",
   "source": "Our dev-slice and example-pool counts (dev-slice/stats.json)",
   "url": null,
   "read": "2026-09-30",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "n-options",
   "what": "Options in our request: the reasons plus none-of-these",
   "value": 78,
   "shown": "78",
   "whose": "Model Fatigue",
   "source": "Our protocol (FREEZE.md), fixed before any test message was sent",
   "url": null,
   "read": "2026-09-30",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "n-shots",
   "what": "Labelled examples in each of our requests",
   "value": 5,
   "shown": "5",
   "whose": "Model Fatigue",
   "source": "Our protocol (FREEZE.md), fixed before any test message was sent",
   "url": null,
   "read": "2026-09-30",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "budget",
   "what": "Wrong answers allowed among those a model routes on its own, the aim each cut-off was set for",
   "value": 0.02,
   "shown": "2%",
   "whose": "Model Fatigue",
   "source": "Our protocol (FREEZE.md), fixed before any test message was sent",
   "url": null,
   "read": "2026-09-30",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-version",
   "what": "Jev, the version that answered",
   "value": "jev-1.13.0",
   "shown": "jev-1.13.0",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jdi-edition",
   "what": "Edition of the Decision Index Cloudflare ran",
   "value": "0.2.1",
   "shown": "0.2.1",
   "whose": "Cloudflare",
   "source": "Clef's model card on Hugging Face (the full Decision Index table)",
   "url": "https://huggingface.co/Cloudflare/clef",
   "read": "2026-10-02T15:31",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "price-clef",
   "what": "Clef, price per million input tokens on Workers AI",
   "value": 0.24,
   "shown": "$0.240",
   "whose": "Cloudflare",
   "source": "Cloudflare Workers AI pricing",
   "url": "https://developers.cloudflare.com/workers-ai/platform/pricing/",
   "read": "2026-10-02T20:52",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "price-flash",
   "what": "Clef-flash, price per million input tokens on Workers AI",
   "value": 0.09,
   "shown": "$0.090",
   "whose": "Cloudflare",
   "source": "Cloudflare Workers AI pricing",
   "url": "https://developers.cloudflare.com/workers-ai/platform/pricing/",
   "read": "2026-10-02T20:52",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "price-jev",
   "what": "Jev, price per million input tokens",
   "value": 0.042,
   "shown": "$0.042",
   "whose": "TypeSafe",
   "source": "TypeSafe's models page (Jev's price and inputs)",
   "url": "https://docs.typesafe.ai/models",
   "read": "2026-10-02T20:52",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-acc",
   "what": "Jev, share right with five examples",
   "value": 0.9392857142857143,
   "shown": "93.9%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-acc-lo",
   "what": "Jev, share right, 95% interval, low",
   "value": 0.9302921633913691,
   "shown": "93.0",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-acc-hi",
   "what": "Jev, share right, 95% interval, high",
   "value": 0.9471848120797445,
   "shown": "94.7",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-correct",
   "what": "Jev, messages sorted correctly with five examples",
   "value": 2893,
   "shown": "2,893",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-none",
   "what": "Jev, times it answered none-of-these (always wrong on this test)",
   "value": 14,
   "shown": "14",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-p50",
   "what": "Jev, median time per request from Berlin",
   "value": 0.235395,
   "shown": "235 ms",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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": 0.33507649999999994,
   "shown": "335 ms",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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": 0.44170570000000026,
   "shown": "442 ms",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-per1k",
   "what": "Jev, dollars per 1,000 messages (price list × input tokens counted)",
   "value": 0.07753194545454546,
   "shown": "$0.0775",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-di-f1",
   "what": "Jev, macro-F1 on the Decision Index's request",
   "value": 0.8000046642892913,
   "shown": "80.00",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-di-f1-lo",
   "what": "Jev, macro-F1 on their request, 95% interval, low",
   "value": 0.7860837415265793,
   "shown": "78.61",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-di-f1-hi",
   "what": "Jev, macro-F1 on their request, 95% interval, high",
   "value": 0.8108486777677364,
   "shown": "81.08",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-di-acc",
   "what": "Jev, share right on the Decision Index's request",
   "value": 0.8071428571428572,
   "shown": "80.7%",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-ece",
   "what": "Jev, calibration error of its probabilities (ECE; lower is closer)",
   "value": 0.03540259740259737,
   "shown": "0.035",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-auroc",
   "what": "Jev, how well its probabilities separate its right answers from its wrong ones (AUROC)",
   "value": 0.8396526005053688,
   "shown": "0.84",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-thr",
   "what": "Jev, cut-off on its top probability, set on the held-out training messages",
   "value": 0.99,
   "shown": "0.99",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-cov",
   "what": "Jev, share of messages routed without a person",
   "value": 0.8438311688311688,
   "shown": "84.4%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-risk",
   "what": "Jev, wrong among the messages it routed itself",
   "value": 0.020392458637937667,
   "shown": "2.04%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-risk-lo",
   "what": "Jev, wrong among routed, 95% interval, low",
   "value": 0.015624462633927156,
   "shown": "1.56",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-risk-hi",
   "what": "Jev, wrong among routed, 95% interval, high",
   "value": 0.02657618453568931,
   "shown": "2.66",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-ref",
   "what": "Jev, messages sent to a person",
   "value": 481,
   "shown": "481",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-errs",
   "what": "Jev, wrong answers routed without a person",
   "value": 53,
   "shown": "53",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-acc",
   "what": "Clef, share right with five examples",
   "value": 0.9506493506493506,
   "shown": "95.1%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-correct",
   "what": "Clef, messages sorted correctly with five examples",
   "value": 2928,
   "shown": "2,928",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-none",
   "what": "Clef, times it answered none-of-these (always wrong on this test)",
   "value": 0,
   "shown": "0",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-per1k",
   "what": "Clef, dollars per 1,000 messages (price list × input tokens counted)",
   "value": 0.4797513506493506,
   "shown": "$0.4798",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-di-f1",
   "what": "Clef, macro-F1 on the Decision Index's request",
   "value": 0.9420316998167463,
   "shown": "94.20",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-di-f1-lo",
   "what": "Clef, macro-F1 on their request, 95% interval, low",
   "value": 0.9329378000053707,
   "shown": "93.29",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-di-f1-hi",
   "what": "Clef, macro-F1 on their request, 95% interval, high",
   "value": 0.9494195074154366,
   "shown": "94.94",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-di-acc",
   "what": "Clef, share right on the Decision Index's request",
   "value": 0.9422077922077922,
   "shown": "94.2%",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-ece",
   "what": "Clef, calibration error of its probabilities (ECE; lower is closer)",
   "value": 0.029707012987013053,
   "shown": "0.030",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-auroc",
   "what": "Clef, how well its probabilities separate its right answers from its wrong ones (AUROC)",
   "value": 0.9001181873741732,
   "shown": "0.90",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-thr",
   "what": "Clef, cut-off on its top probability, set on the held-out training messages",
   "value": 0.8034,
   "shown": "0.80",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-cov",
   "what": "Clef, share of messages routed without a person",
   "value": 0.9321428571428572,
   "shown": "93.2%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-risk",
   "what": "Clef, wrong among the messages it routed itself",
   "value": 0.024033437826541274,
   "shown": "2.40%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-risk-lo",
   "what": "Clef, wrong among routed, 95% interval, low",
   "value": 0.019034914578768047,
   "shown": "1.90",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-risk-hi",
   "what": "Clef, wrong among routed, 95% interval, high",
   "value": 0.030304012477247837,
   "shown": "3.03",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-ref",
   "what": "Clef, messages sent to a person",
   "value": 209,
   "shown": "209",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-errs",
   "what": "Clef, wrong answers routed without a person",
   "value": 69,
   "shown": "69",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-acc",
   "what": "Clef-flash, share right with five examples",
   "value": 0.9516233766233766,
   "shown": "95.2%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-correct",
   "what": "Clef-flash, messages sorted correctly with five examples",
   "value": 2931,
   "shown": "2,931",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-none",
   "what": "Clef-flash, times it answered none-of-these (always wrong on this test)",
   "value": 0,
   "shown": "0",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "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 run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-per1k",
   "what": "Clef-flash, dollars per 1,000 messages (price list × input tokens counted)",
   "value": 0.17990675649350651,
   "shown": "$0.1799",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-di-f1",
   "what": "Clef-flash, macro-F1 on the Decision Index's request",
   "value": 0.9084736076902837,
   "shown": "90.85",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-di-f1-lo",
   "what": "Clef-flash, macro-F1 on their request, 95% interval, low",
   "value": 0.8977984284255756,
   "shown": "89.78",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-di-f1-hi",
   "what": "Clef-flash, macro-F1 on their request, 95% interval, high",
   "value": 0.9173864837464847,
   "shown": "91.74",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-di-acc",
   "what": "Clef-flash, share right on the Decision Index's request",
   "value": 0.9087662337662338,
   "shown": "90.9%",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-ece",
   "what": "Clef-flash, calibration error of its probabilities (ECE; lower is closer)",
   "value": 0.14252301948051946,
   "shown": "0.143",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-auroc",
   "what": "Clef-flash, how well its probabilities separate its right answers from its wrong ones (AUROC)",
   "value": 0.8190472592216047,
   "shown": "0.82",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-thr",
   "what": "Clef-flash, cut-off on its top probability, set on the held-out training messages",
   "value": 0.7323,
   "shown": "0.73",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-cov",
   "what": "Clef-flash, share of messages routed without a person",
   "value": 0.8168831168831169,
   "shown": "81.7%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-risk",
   "what": "Clef-flash, wrong among the messages it routed itself",
   "value": 0.021462639109697933,
   "shown": "2.15%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-risk-lo",
   "what": "Clef-flash, wrong among routed, 95% interval, low",
   "value": 0.01648687278154381,
   "shown": "1.65",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-risk-hi",
   "what": "Clef-flash, wrong among routed, 95% interval, high",
   "value": 0.027897504395180316,
   "shown": "2.79",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-ref",
   "what": "Clef-flash, messages sent to a person",
   "value": 564,
   "shown": "564",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-errs",
   "what": "Clef-flash, wrong answers routed without a person",
   "value": 54,
   "shown": "54",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-acc",
   "what": "Kev-27B, share right with five examples",
   "value": 0.9383116883116883,
   "shown": "93.8%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-correct",
   "what": "Kev-27B, messages sorted correctly with five examples",
   "value": 2890,
   "shown": "2,890",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-none",
   "what": "Kev-27B, times it answered none-of-these (always wrong on this test)",
   "value": 21,
   "shown": "21",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-per1k",
   "what": "Kev-27B, dollars per 1,000 messages (our GPU time as Modal billed it, one call at a time)",
   "value": 0.7295102727351169,
   "shown": "$0.7295",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-di-f1",
   "what": "Kev-27B, macro-F1 on the Decision Index's request",
   "value": 0.8640871097889219,
   "shown": "86.41",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-di-f1-lo",
   "what": "Kev-27B, macro-F1 on their request, 95% interval, low",
   "value": 0.8514143596341035,
   "shown": "85.14",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-di-f1-hi",
   "what": "Kev-27B, macro-F1 on their request, 95% interval, high",
   "value": 0.8740878848855513,
   "shown": "87.41",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-di-acc",
   "what": "Kev-27B, share right on the Decision Index's request",
   "value": 0.8681818181818182,
   "shown": "86.8%",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-ece",
   "what": "Kev-27B, calibration error of its probabilities (ECE; lower is closer)",
   "value": 0.014207987012986964,
   "shown": "0.014",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-auroc",
   "what": "Kev-27B, how well its probabilities separate its right answers from its wrong ones (AUROC)",
   "value": 0.920481697322892,
   "shown": "0.92",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-thr",
   "what": "Kev-27B, cut-off on its top probability, set on the held-out training messages",
   "value": 0.7928,
   "shown": "0.79",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-cov",
   "what": "Kev-27B, share of messages routed without a person",
   "value": 0.9103896103896104,
   "shown": "91.0%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-risk",
   "what": "Kev-27B, wrong among the messages it routed itself",
   "value": 0.02282453637660485,
   "shown": "2.28%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-risk-lo",
   "what": "Kev-27B, wrong among routed, 95% interval, low",
   "value": 0.017914905043793377,
   "shown": "1.79",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-risk-hi",
   "what": "Kev-27B, wrong among routed, 95% interval, high",
   "value": 0.0290398804398322,
   "shown": "2.90",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-ref",
   "what": "Kev-27B, messages sent to a person",
   "value": 276,
   "shown": "276",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-errs",
   "what": "Kev-27B, wrong answers routed without a person",
   "value": 64,
   "shown": "64",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-acc",
   "what": "Kev-9B, share right with five examples",
   "value": 0.9133116883116883,
   "shown": "91.3%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-correct",
   "what": "Kev-9B, messages sorted correctly with five examples",
   "value": 2813,
   "shown": "2,813",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-none",
   "what": "Kev-9B, times it answered none-of-these (always wrong on this test)",
   "value": 110,
   "shown": "110",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "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 run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-per1k",
   "what": "Kev-9B, dollars per 1,000 messages (our GPU time as Modal billed it, one call at a time)",
   "value": 0.6211176265005325,
   "shown": "$0.6211",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-di-f1",
   "what": "Kev-9B, macro-F1 on the Decision Index's request",
   "value": 0.8302597361284337,
   "shown": "83.03",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-di-f1-lo",
   "what": "Kev-9B, macro-F1 on their request, 95% interval, low",
   "value": 0.8157288132697301,
   "shown": "81.57",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-di-f1-hi",
   "what": "Kev-9B, macro-F1 on their request, 95% interval, high",
   "value": 0.8409507773641868,
   "shown": "84.10",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-di-acc",
   "what": "Kev-9B, share right on the Decision Index's request",
   "value": 0.8331168831168831,
   "shown": "83.3%",
   "whose": "Model Fatigue",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-ece",
   "what": "Kev-9B, calibration error of its probabilities (ECE; lower is closer)",
   "value": 0.03842435064935068,
   "shown": "0.038",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-auroc",
   "what": "Kev-9B, how well its probabilities separate its right answers from its wrong ones (AUROC)",
   "value": 0.9057099794826321,
   "shown": "0.91",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-thr",
   "what": "Kev-9B, cut-off on its top probability, set on the held-out training messages",
   "value": 0.8127,
   "shown": "0.81",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-cov",
   "what": "Kev-9B, share of messages routed without a person",
   "value": 0.826948051948052,
   "shown": "82.7%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-risk",
   "what": "Kev-9B, wrong among the messages it routed itself",
   "value": 0.024734982332155476,
   "shown": "2.47%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-risk-lo",
   "what": "Kev-9B, wrong among routed, 95% interval, low",
   "value": 0.019380968427526332,
   "shown": "1.94",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-risk-hi",
   "what": "Kev-9B, wrong among routed, 95% interval, high",
   "value": 0.03152050659938243,
   "shown": "3.15",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-ref",
   "what": "Kev-9B, messages sent to a person",
   "value": 533,
   "shown": "533",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-errs",
   "what": "Kev-9B, wrong answers routed without a person",
   "value": 63,
   "shown": "63",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel2-n",
   "what": "Jev, answers given 0.2 to 0.3: how many",
   "value": 1,
   "shown": "1",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel2-pred",
   "what": "Jev, answers given 0.2 to 0.3: average probability given",
   "value": 0.28,
   "shown": "0.28",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel2-obs",
   "what": "Jev, answers given 0.2 to 0.3: share right",
   "value": 0.0,
   "shown": "0.00",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel3-n",
   "what": "Jev, answers given 0.3 to 0.4: how many",
   "value": 3,
   "shown": "3",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel3-pred",
   "what": "Jev, answers given 0.3 to 0.4: average probability given",
   "value": 0.3466666666666667,
   "shown": "0.35",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel3-obs",
   "what": "Jev, answers given 0.3 to 0.4: share right",
   "value": 0.6666666666666666,
   "shown": "0.67",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel4-n",
   "what": "Jev, answers given 0.4 to 0.5: how many",
   "value": 17,
   "shown": "17",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel4-pred",
   "what": "Jev, answers given 0.4 to 0.5: average probability given",
   "value": 0.44588235294117645,
   "shown": "0.45",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel4-obs",
   "what": "Jev, answers given 0.4 to 0.5: share right",
   "value": 0.35294117647058826,
   "shown": "0.35",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel5-n",
   "what": "Jev, answers given 0.5 to 0.6: how many",
   "value": 37,
   "shown": "37",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel5-pred",
   "what": "Jev, answers given 0.5 to 0.6: average probability given",
   "value": 0.5454054054054055,
   "shown": "0.55",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel5-obs",
   "what": "Jev, answers given 0.5 to 0.6: share right",
   "value": 0.4864864864864865,
   "shown": "0.49",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel6-n",
   "what": "Jev, answers given 0.6 to 0.7: how many",
   "value": 39,
   "shown": "39",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel6-pred",
   "what": "Jev, answers given 0.6 to 0.7: average probability given",
   "value": 0.6448717948717948,
   "shown": "0.64",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel6-obs",
   "what": "Jev, answers given 0.6 to 0.7: share right",
   "value": 0.4358974358974359,
   "shown": "0.44",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel7-n",
   "what": "Jev, answers given 0.7 to 0.8: how many",
   "value": 46,
   "shown": "46",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel7-pred",
   "what": "Jev, answers given 0.7 to 0.8: average probability given",
   "value": 0.751304347826087,
   "shown": "0.75",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel7-obs",
   "what": "Jev, answers given 0.7 to 0.8: share right",
   "value": 0.6304347826086957,
   "shown": "0.63",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel8-n",
   "what": "Jev, answers given 0.8 to 0.9: how many",
   "value": 81,
   "shown": "81",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel8-pred",
   "what": "Jev, answers given 0.8 to 0.9: average probability given",
   "value": 0.8523456790123456,
   "shown": "0.85",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel8-obs",
   "what": "Jev, answers given 0.8 to 0.9: share right",
   "value": 0.7160493827160493,
   "shown": "0.72",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel9-n",
   "what": "Jev, answers given 0.9 to 1.0: how many",
   "value": 2856,
   "shown": "2,856",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel9-pred",
   "what": "Jev, answers given 0.9 to 1.0: average probability given",
   "value": 0.9951995798319327,
   "shown": "1.00",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-rel9-obs",
   "what": "Jev, answers given 0.9 to 1.0: share right",
   "value": 0.967436974789916,
   "shown": "0.97",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel2-n",
   "what": "Clef, answers given 0.2 to 0.3: how many",
   "value": 5,
   "shown": "5",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel2-pred",
   "what": "Clef, answers given 0.2 to 0.3: average probability given",
   "value": 0.2811,
   "shown": "0.28",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel2-obs",
   "what": "Clef, answers given 0.2 to 0.3: share right",
   "value": 0.4,
   "shown": "0.40",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel3-n",
   "what": "Clef, answers given 0.3 to 0.4: how many",
   "value": 13,
   "shown": "13",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel3-pred",
   "what": "Clef, answers given 0.3 to 0.4: average probability given",
   "value": 0.35826153846153846,
   "shown": "0.36",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel3-obs",
   "what": "Clef, answers given 0.3 to 0.4: share right",
   "value": 0.6923076923076923,
   "shown": "0.69",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel4-n",
   "what": "Clef, answers given 0.4 to 0.5: how many",
   "value": 23,
   "shown": "23",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel4-pred",
   "what": "Clef, answers given 0.4 to 0.5: average probability given",
   "value": 0.4494304347826087,
   "shown": "0.45",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel4-obs",
   "what": "Clef, answers given 0.4 to 0.5: share right",
   "value": 0.5652173913043478,
   "shown": "0.57",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel5-n",
   "what": "Clef, answers given 0.5 to 0.6: how many",
   "value": 38,
   "shown": "38",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel5-pred",
   "what": "Clef, answers given 0.5 to 0.6: average probability given",
   "value": 0.549507894736842,
   "shown": "0.55",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel5-obs",
   "what": "Clef, answers given 0.5 to 0.6: share right",
   "value": 0.5,
   "shown": "0.50",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel6-n",
   "what": "Clef, answers given 0.6 to 0.7: how many",
   "value": 55,
   "shown": "55",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel6-pred",
   "what": "Clef, answers given 0.6 to 0.7: average probability given",
   "value": 0.6482090909090911,
   "shown": "0.65",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel6-obs",
   "what": "Clef, answers given 0.6 to 0.7: share right",
   "value": 0.5636363636363636,
   "shown": "0.56",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel7-n",
   "what": "Clef, answers given 0.7 to 0.8: how many",
   "value": 69,
   "shown": "69",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel7-pred",
   "what": "Clef, answers given 0.7 to 0.8: average probability given",
   "value": 0.7555376811594203,
   "shown": "0.76",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel7-obs",
   "what": "Clef, answers given 0.7 to 0.8: share right",
   "value": 0.6956521739130435,
   "shown": "0.70",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel8-n",
   "what": "Clef, answers given 0.8 to 0.9: how many",
   "value": 187,
   "shown": "187",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel8-pred",
   "what": "Clef, answers given 0.8 to 0.9: average probability given",
   "value": 0.8563213903743316,
   "shown": "0.86",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel8-obs",
   "what": "Clef, answers given 0.8 to 0.9: share right",
   "value": 0.8342245989304813,
   "shown": "0.83",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel9-n",
   "what": "Clef, answers given 0.9 to 1.0: how many",
   "value": 2690,
   "shown": "2,690",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel9-pred",
   "what": "Clef, answers given 0.9 to 1.0: average probability given",
   "value": 0.9594422304832713,
   "shown": "0.96",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-rel9-obs",
   "what": "Clef, answers given 0.9 to 1.0: share right",
   "value": 0.9851301115241635,
   "shown": "0.99",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel2-n",
   "what": "Clef-flash, answers given 0.2 to 0.3: how many",
   "value": 7,
   "shown": "7",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel2-pred",
   "what": "Clef-flash, answers given 0.2 to 0.3: average probability given",
   "value": 0.2601,
   "shown": "0.26",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel2-obs",
   "what": "Clef-flash, answers given 0.2 to 0.3: share right",
   "value": 0.14285714285714285,
   "shown": "0.14",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel3-n",
   "what": "Clef-flash, answers given 0.3 to 0.4: how many",
   "value": 12,
   "shown": "12",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel3-pred",
   "what": "Clef-flash, answers given 0.3 to 0.4: average probability given",
   "value": 0.34730833333333333,
   "shown": "0.35",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel3-obs",
   "what": "Clef-flash, answers given 0.3 to 0.4: share right",
   "value": 0.3333333333333333,
   "shown": "0.33",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel4-n",
   "what": "Clef-flash, answers given 0.4 to 0.5: how many",
   "value": 59,
   "shown": "59",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel4-pred",
   "what": "Clef-flash, answers given 0.4 to 0.5: average probability given",
   "value": 0.4583101694915255,
   "shown": "0.46",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel4-obs",
   "what": "Clef-flash, answers given 0.4 to 0.5: share right",
   "value": 0.6779661016949152,
   "shown": "0.68",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel5-n",
   "what": "Clef-flash, answers given 0.5 to 0.6: how many",
   "value": 87,
   "shown": "87",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel5-pred",
   "what": "Clef-flash, answers given 0.5 to 0.6: average probability given",
   "value": 0.5560540229885057,
   "shown": "0.56",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel5-obs",
   "what": "Clef-flash, answers given 0.5 to 0.6: share right",
   "value": 0.7816091954022989,
   "shown": "0.78",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel6-n",
   "what": "Clef-flash, answers given 0.6 to 0.7: how many",
   "value": 241,
   "shown": "241",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel6-pred",
   "what": "Clef-flash, answers given 0.6 to 0.7: average probability given",
   "value": 0.6557560165975104,
   "shown": "0.66",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel6-obs",
   "what": "Clef-flash, answers given 0.6 to 0.7: share right",
   "value": 0.8713692946058091,
   "shown": "0.87",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel7-n",
   "what": "Clef-flash, answers given 0.7 to 0.8: how many",
   "value": 736,
   "shown": "736",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel7-pred",
   "what": "Clef-flash, answers given 0.7 to 0.8: average probability given",
   "value": 0.75900625,
   "shown": "0.76",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel7-obs",
   "what": "Clef-flash, answers given 0.7 to 0.8: share right",
   "value": 0.9510869565217391,
   "shown": "0.95",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel8-n",
   "what": "Clef-flash, answers given 0.8 to 0.9: how many",
   "value": 1339,
   "shown": "1,339",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel8-pred",
   "what": "Clef-flash, answers given 0.8 to 0.9: average probability given",
   "value": 0.8537438386855862,
   "shown": "0.85",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel8-obs",
   "what": "Clef-flash, answers given 0.8 to 0.9: share right",
   "value": 0.9783420463032113,
   "shown": "0.98",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel9-n",
   "what": "Clef-flash, answers given 0.9 to 1.0: how many",
   "value": 599,
   "shown": "599",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel9-pred",
   "what": "Clef-flash, answers given 0.9 to 1.0: average probability given",
   "value": 0.9228242070116862,
   "shown": "0.92",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-rel9-obs",
   "what": "Clef-flash, answers given 0.9 to 1.0: share right",
   "value": 0.998330550918197,
   "shown": "1.00",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel1-n",
   "what": "Kev-27B, answers given 0.1 to 0.2: how many",
   "value": 1,
   "shown": "1",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel1-pred",
   "what": "Kev-27B, answers given 0.1 to 0.2: average probability given",
   "value": 0.1888,
   "shown": "0.19",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel1-obs",
   "what": "Kev-27B, answers given 0.1 to 0.2: share right",
   "value": 1.0,
   "shown": "1.00",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel2-n",
   "what": "Kev-27B, answers given 0.2 to 0.3: how many",
   "value": 9,
   "shown": "9",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel2-pred",
   "what": "Kev-27B, answers given 0.2 to 0.3: average probability given",
   "value": 0.2532888888888889,
   "shown": "0.25",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel2-obs",
   "what": "Kev-27B, answers given 0.2 to 0.3: share right",
   "value": 0.2222222222222222,
   "shown": "0.22",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel3-n",
   "what": "Kev-27B, answers given 0.3 to 0.4: how many",
   "value": 25,
   "shown": "25",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel3-pred",
   "what": "Kev-27B, answers given 0.3 to 0.4: average probability given",
   "value": 0.3658760000000001,
   "shown": "0.37",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel3-obs",
   "what": "Kev-27B, answers given 0.3 to 0.4: share right",
   "value": 0.36,
   "shown": "0.36",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel4-n",
   "what": "Kev-27B, answers given 0.4 to 0.5: how many",
   "value": 51,
   "shown": "51",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel4-pred",
   "what": "Kev-27B, answers given 0.4 to 0.5: average probability given",
   "value": 0.4537529411764705,
   "shown": "0.45",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel4-obs",
   "what": "Kev-27B, answers given 0.4 to 0.5: share right",
   "value": 0.43137254901960786,
   "shown": "0.43",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel5-n",
   "what": "Kev-27B, answers given 0.5 to 0.6: how many",
   "value": 55,
   "shown": "55",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel5-pred",
   "what": "Kev-27B, answers given 0.5 to 0.6: average probability given",
   "value": 0.5475854545454546,
   "shown": "0.55",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel5-obs",
   "what": "Kev-27B, answers given 0.5 to 0.6: share right",
   "value": 0.45454545454545453,
   "shown": "0.45",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel6-n",
   "what": "Kev-27B, answers given 0.6 to 0.7: how many",
   "value": 62,
   "shown": "62",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel6-pred",
   "what": "Kev-27B, answers given 0.6 to 0.7: average probability given",
   "value": 0.6482032258064516,
   "shown": "0.65",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel6-obs",
   "what": "Kev-27B, answers given 0.6 to 0.7: share right",
   "value": 0.5645161290322581,
   "shown": "0.56",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel7-n",
   "what": "Kev-27B, answers given 0.7 to 0.8: how many",
   "value": 75,
   "shown": "75",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel7-pred",
   "what": "Kev-27B, answers given 0.7 to 0.8: average probability given",
   "value": 0.7496440000000001,
   "shown": "0.75",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel7-obs",
   "what": "Kev-27B, answers given 0.7 to 0.8: share right",
   "value": 0.7733333333333333,
   "shown": "0.77",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel8-n",
   "what": "Kev-27B, answers given 0.8 to 0.9: how many",
   "value": 204,
   "shown": "204",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel8-pred",
   "what": "Kev-27B, answers given 0.8 to 0.9: average probability given",
   "value": 0.8589098039215687,
   "shown": "0.86",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel8-obs",
   "what": "Kev-27B, answers given 0.8 to 0.9: share right",
   "value": 0.8578431372549019,
   "shown": "0.86",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel9-n",
   "what": "Kev-27B, answers given 0.9 to 1.0: how many",
   "value": 2598,
   "shown": "2,598",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel9-pred",
   "what": "Kev-27B, answers given 0.9 to 1.0: average probability given",
   "value": 0.9753343341031563,
   "shown": "0.98",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-rel9-obs",
   "what": "Kev-27B, answers given 0.9 to 1.0: share right",
   "value": 0.9865280985373364,
   "shown": "0.99",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel2-n",
   "what": "Kev-9B, answers given 0.2 to 0.3: how many",
   "value": 2,
   "shown": "2",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel2-pred",
   "what": "Kev-9B, answers given 0.2 to 0.3: average probability given",
   "value": 0.2871,
   "shown": "0.29",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel2-obs",
   "what": "Kev-9B, answers given 0.2 to 0.3: share right",
   "value": 0.5,
   "shown": "0.50",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel3-n",
   "what": "Kev-9B, answers given 0.3 to 0.4: how many",
   "value": 21,
   "shown": "21",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel3-pred",
   "what": "Kev-9B, answers given 0.3 to 0.4: average probability given",
   "value": 0.36151428571428573,
   "shown": "0.36",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel3-obs",
   "what": "Kev-9B, answers given 0.3 to 0.4: share right",
   "value": 0.23809523809523808,
   "shown": "0.24",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel4-n",
   "what": "Kev-9B, answers given 0.4 to 0.5: how many",
   "value": 88,
   "shown": "88",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel4-pred",
   "what": "Kev-9B, answers given 0.4 to 0.5: average probability given",
   "value": 0.4609920454545455,
   "shown": "0.46",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel4-obs",
   "what": "Kev-9B, answers given 0.4 to 0.5: share right",
   "value": 0.375,
   "shown": "0.38",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel5-n",
   "what": "Kev-9B, answers given 0.5 to 0.6: how many",
   "value": 102,
   "shown": "102",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel5-pred",
   "what": "Kev-9B, answers given 0.5 to 0.6: average probability given",
   "value": 0.5521333333333334,
   "shown": "0.55",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel5-obs",
   "what": "Kev-9B, answers given 0.5 to 0.6: share right",
   "value": 0.47058823529411764,
   "shown": "0.47",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel6-n",
   "what": "Kev-9B, answers given 0.6 to 0.7: how many",
   "value": 115,
   "shown": "115",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel6-pred",
   "what": "Kev-9B, answers given 0.6 to 0.7: average probability given",
   "value": 0.6540373913043479,
   "shown": "0.65",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel6-obs",
   "what": "Kev-9B, answers given 0.6 to 0.7: share right",
   "value": 0.6434782608695652,
   "shown": "0.64",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel7-n",
   "what": "Kev-9B, answers given 0.7 to 0.8: how many",
   "value": 180,
   "shown": "180",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel7-pred",
   "what": "Kev-9B, answers given 0.7 to 0.8: average probability given",
   "value": 0.7548805555555556,
   "shown": "0.75",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel7-obs",
   "what": "Kev-9B, answers given 0.7 to 0.8: share right",
   "value": 0.8055555555555556,
   "shown": "0.81",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel8-n",
   "what": "Kev-9B, answers given 0.8 to 0.9: how many",
   "value": 440,
   "shown": "440",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel8-pred",
   "what": "Kev-9B, answers given 0.8 to 0.9: average probability given",
   "value": 0.8604661363636363,
   "shown": "0.86",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel8-obs",
   "what": "Kev-9B, answers given 0.8 to 0.9: share right",
   "value": 0.9204545454545454,
   "shown": "0.92",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel9-n",
   "what": "Kev-9B, answers given 0.9 to 1.0: how many",
   "value": 2132,
   "shown": "2,132",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel9-pred",
   "what": "Kev-9B, answers given 0.9 to 1.0: average probability given",
   "value": 0.9565132270168856,
   "shown": "0.96",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-rel9-obs",
   "what": "Kev-9B, answers given 0.9 to 1.0: share right",
   "value": 0.9859287054409006,
   "shown": "0.99",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-vj-ar",
   "what": "Messages Clef got right and Jev got wrong",
   "value": 54,
   "shown": "54",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-vj-jr",
   "what": "Messages Jev got right and Clef got wrong",
   "value": 19,
   "shown": "19",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-vj-ar",
   "what": "Messages Clef-flash got right and Jev got wrong",
   "value": 60,
   "shown": "60",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-vj-jr",
   "what": "Messages Jev got right and Clef-flash got wrong",
   "value": 22,
   "shown": "22",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vj-ar",
   "what": "Messages Kev-27B got right and Jev got wrong",
   "value": 24,
   "shown": "24",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vj-jr",
   "what": "Messages Jev got right and Kev-27B got wrong",
   "value": 27,
   "shown": "27",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-vj-ar",
   "what": "Messages Kev-9B got right and Jev got wrong",
   "value": 17,
   "shown": "17",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-vj-jr",
   "what": "Messages Jev got right and Kev-9B got wrong",
   "value": 97,
   "shown": "97",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-nonone",
   "what": "Kev-27B, share right with none-of-these set aside (its top real reason)",
   "value": 0.9409090909090909,
   "shown": "94.1%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-nonone",
   "what": "Kev-9B, share right with none-of-these set aside (its top real reason)",
   "value": 0.9363636363636364,
   "shown": "93.6%",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "hourly",
   "what": "A person's time, our assumption for the break-even",
   "value": 15,
   "shown": "$15",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "only-clef",
   "what": "Messages Clef routed on its own that Jev sent to a person",
   "value": 318,
   "shown": "318",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "only-clef-wrong",
   "what": "Of those, how many Clef got wrong",
   "value": 31,
   "shown": "31",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "disagree",
   "what": "Messages where one of Clef and Jev was right and the other wrong",
   "value": 73,
   "shown": "73",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "both-below",
   "what": "Of those, messages where both answers were below their cut-offs",
   "value": 47,
   "shown": "47",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "right-to-person",
   "what": "Of those, messages where the model that was right would have sent it to a person",
   "value": 53,
   "shown": "53",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-lo-n",
   "what": "Jev's answers below 0.99",
   "value": 481,
   "shown": "481",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-lo-said",
   "what": "Jev's answers below 0.99: average stated probability",
   "value": 0.836985446985447,
   "shown": "0.84",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-lo-right",
   "what": "Jev's answers below 0.99: share right",
   "value": 0.7214137214137214,
   "shown": "72%",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-b7-lo",
   "what": "Clef-flash, that group's lowest probability",
   "value": 0.7,
   "shown": "0.7",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-b7-hi",
   "what": "Clef-flash, that group's highest probability",
   "value": 0.8,
   "shown": "0.8",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-b7-n",
   "what": "Clef-flash, answers it gave a probability of 0.7 to 0.8",
   "value": 736,
   "shown": "736",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-b7-said",
   "what": "Those answers: average stated probability",
   "value": 0.75900625,
   "shown": "0.76",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-b7-right",
   "what": "Those answers: share right",
   "value": 0.9510869565217391,
   "shown": "95%",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flips-fixed",
   "what": "Jev's answers wrong on their request and right with five examples",
   "value": 437,
   "shown": "437",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flips-broken",
   "what": "Jev's answers right on their request and wrong with five examples",
   "value": 30,
   "shown": "30",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-ctl",
   "what": "Jev on our request with the examples taken out (jev-2, 21 September)",
   "value": 0.7834415584415585,
   "shown": "78.3%",
   "whose": "Model Fatigue (M37's jev-2 run)",
   "source": "M37's jev-2 run of Jev on the same protocol and the same messages, without examples",
   "url": "https://www.youtube.com/watch?v=MwGATggBw-c",
   "read": "2026-09-21",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-ctl-none",
   "what": "Jev in that control, times it answered none-of-these",
   "value": 131,
   "shown": "131",
   "whose": "Model Fatigue (M37's jev-2 run)",
   "source": "M37's jev-2 run of Jev on the same protocol and the same messages, without examples",
   "url": "https://www.youtube.com/watch?v=MwGATggBw-c",
   "read": "2026-09-21",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sba-jev-zs",
   "what": "Message A, Jev, probability of its answer on their request",
   "value": 0.63,
   "shown": "0.63",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sba-jev-ex",
   "what": "Message A, Jev, probability of its answer on with five examples",
   "value": 1.0,
   "shown": "1.00",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sba-clef-zs",
   "what": "Message A, Clef, probability of its answer on their request",
   "value": 0.8974,
   "shown": "0.90",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sba-clef-ex",
   "what": "Message A, Clef, probability of its answer on with five examples",
   "value": 0.9571,
   "shown": "0.96",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sba-flash-zs",
   "what": "Message A, Clef-flash, probability of its answer on their request",
   "value": 0.7699,
   "shown": "0.77",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sba-flash-ex",
   "what": "Message A, Clef-flash, probability of its answer on with five examples",
   "value": 0.8909,
   "shown": "0.89",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbb-jev-zs",
   "what": "Message B, Jev, probability of its answer on their request",
   "value": 0.48,
   "shown": "0.48",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbb-jev-ex",
   "what": "Message B, Jev, probability of its answer on with five examples",
   "value": 0.81,
   "shown": "0.81",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbb-clef-zs",
   "what": "Message B, Clef, probability of its answer on their request",
   "value": 0.283,
   "shown": "0.28",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbb-clef-ex",
   "what": "Message B, Clef, probability of its answer on with five examples",
   "value": 0.2686,
   "shown": "0.27",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbb-flash-zs",
   "what": "Message B, Clef-flash, probability of its answer on their request",
   "value": 0.0964,
   "shown": "0.10",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbb-flash-ex",
   "what": "Message B, Clef-flash, probability of its answer on with five examples",
   "value": 0.4753,
   "shown": "0.48",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbc-jev-zs",
   "what": "Message C, Jev, probability of its answer on their request",
   "value": 0.64,
   "shown": "0.64",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbc-jev-ex",
   "what": "Message C, Jev, probability of its answer on with five examples",
   "value": 0.44,
   "shown": "0.44",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbc-clef-zs",
   "what": "Message C, Clef, probability of its answer on their request",
   "value": 0.6723,
   "shown": "0.67",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbc-clef-ex",
   "what": "Message C, Clef, probability of its answer on with five examples",
   "value": 0.6029,
   "shown": "0.60",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbc-flash-zs",
   "what": "Message C, Clef-flash, probability of its answer on their request",
   "value": 0.4004,
   "shown": "0.40",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "sbc-flash-ex",
   "what": "Message C, Clef-flash, probability of its answer on with five examples",
   "value": 0.5868,
   "shown": "0.59",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-slow",
   "what": "Clef-flash, calls that took over a second",
   "value": 159,
   "shown": "159",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-slow-q1",
   "what": "Of those, calls in the first quarter of its pass",
   "value": 139,
   "shown": "139",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "floor-ts",
   "what": "A bare request to TypeSafe's host from the Mac mini, warm, median of 20, start of the run",
   "value": 0.1708,
   "shown": "171 ms",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "floor-cf",
   "what": "A bare request to Cloudflare's API host, warm, median of 20, start of the run (not the inference path)",
   "value": 0.2037,
   "shown": "204 ms",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-server",
   "what": "Kev-27B, median time the model took by its own server's count",
   "value": 0.1105,
   "shown": "110 ms",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-server",
   "what": "Kev-9B, median time the model took by its own server's count",
   "value": 0.0354,
   "shown": "35 ms",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-hour",
   "what": "Kev-27B, Modal's bill per container-hour (our arithmetic on the bill)",
   "value": 4.916910767787354,
   "shown": "$4.92",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-c6",
   "what": "Kev-27B, dollars per 1,000 at six requests at a time, projected from measured throughput",
   "value": 0.1572897050265309,
   "shown": "$0.157",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-billed",
   "what": "Kev-27B, everything Modal billed for this model",
   "value": 4.76201352,
   "shown": "$4.76",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-overhead",
   "what": "Kev-27B, start-up and idle time on the bill",
   "value": 1.6879882241502782,
   "shown": "$1.69",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-hour",
   "what": "Kev-9B, Modal's bill per container-hour (our arithmetic on the bill)",
   "value": 4.908732888551607,
   "shown": "$4.91",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-c6",
   "what": "Kev-9B, dollars per 1,000 at six requests at a time, projected from measured throughput",
   "value": 0.11191627653552004,
   "shown": "$0.112",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-billed",
   "what": "Kev-9B, everything Modal billed for this model",
   "value": 4.670063519999999,
   "shown": "$4.67",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-overhead",
   "what": "Kev-9B, start-up and idle time on the bill",
   "value": 2.1389299474949026,
   "shown": "$2.14",
   "whose": "Model Fatigue",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-di-tokcall",
   "what": "Kev-27B, input tokens per call on the Decision Index's request (its own count)",
   "value": 852.5896103896104,
   "shown": "853",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-tokcall",
   "what": "Kev-27B, input tokens per call on our test (its own count)",
   "value": 974.9639610389611,
   "shown": "975",
   "whose": "Model Fatigue",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "kev-breadth-27",
   "what": "Kev-27B on breadth-v1, datasets no Kev trained on (release notes)",
   "value": 52.3,
   "shown": "52.3",
   "whose": "Jared Palmer",
   "source": "Kev 1.0 release notes (GitHub, jaredpalmer/kev)",
   "url": "https://github.com/jaredpalmer/kev/releases/tag/kev-1.0",
   "read": "2026-10-02T20:31",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "kev-breadth-jev",
   "what": "Jev on breadth-v1, as Kev's release notes report it",
   "value": 54.0,
   "shown": "54.0",
   "whose": "Jared Palmer",
   "source": "Kev 1.0 release notes (GitHub, jaredpalmer/kev)",
   "url": "https://github.com/jaredpalmer/kev/releases/tag/kev-1.0",
   "read": "2026-10-02T20:31",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "cf0-clef",
   "what": "Cloudflare's table, BFCL · case exact: Clef",
   "value": 98.47,
   "shown": "98.47",
   "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": "cf0-flash",
   "what": "Cloudflare's table, BFCL · case exact: Clef-flash",
   "value": 98.76,
   "shown": "98.76",
   "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": "cf0-jev",
   "what": "Cloudflare's table, BFCL · case exact: Jev",
   "value": 95.75,
   "shown": "95.75",
   "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": "cf1-clef",
   "what": "Cloudflare's table, ToolRet · nDCG@10: Clef",
   "value": 69.19,
   "shown": "69.19",
   "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": "cf1-flash",
   "what": "Cloudflare's table, ToolRet · nDCG@10: Clef-flash",
   "value": 66.43,
   "shown": "66.43",
   "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": "cf1-jev",
   "what": "Cloudflare's table, ToolRet · nDCG@10: Jev",
   "value": 65.28,
   "shown": "65.28",
   "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": "cf2-clef",
   "what": "Cloudflare's table, API-Bank · accuracy: Clef",
   "value": 91.93,
   "shown": "91.93",
   "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": "cf2-flash",
   "what": "Cloudflare's table, API-Bank · accuracy: Clef-flash",
   "value": 93.11,
   "shown": "93.11",
   "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": "cf2-jev",
   "what": "Cloudflare's table, API-Bank · accuracy: Jev",
   "value": 88.19,
   "shown": "88.19",
   "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": "cf3-clef",
   "what": "Cloudflare's table, Home appliances · case exact: Clef",
   "value": 82.95,
   "shown": "82.95",
   "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": "cf3-flash",
   "what": "Cloudflare's table, Home appliances · case exact: Clef-flash",
   "value": 97.73,
   "shown": "97.73",
   "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": "cf3-jev",
   "what": "Cloudflare's table, Home appliances · case exact: Jev",
   "value": 52.27,
   "shown": "52.27",
   "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": "cf4-clef",
   "what": "Cloudflare's table, When2Call · accuracy: Clef",
   "value": 72.37,
   "shown": "72.37",
   "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": "cf4-flash",
   "what": "Cloudflare's table, When2Call · accuracy: Clef-flash",
   "value": 65.58,
   "shown": "65.58",
   "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": "cf4-jev",
   "what": "Cloudflare's table, When2Call · accuracy: Jev",
   "value": 80.97,
   "shown": "80.97",
   "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": "cf5-clef",
   "what": "Cloudflare's table, BANKING77 · macro-F1: Clef",
   "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": "cf5-flash",
   "what": "Cloudflare's table, BANKING77 · macro-F1: Clef-flash",
   "value": 90.93,
   "shown": "90.93",
   "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": "cf5-jev",
   "what": "Cloudflare's table, BANKING77 · macro-F1: Jev",
   "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": "cf6-clef",
   "what": "Cloudflare's table, CLINC150+OOS · macro-F1: Clef",
   "value": 97.43,
   "shown": "97.43",
   "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": "cf6-flash",
   "what": "Cloudflare's table, CLINC150+OOS · macro-F1: Clef-flash",
   "value": 66.77,
   "shown": "66.77",
   "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": "cf6-jev",
   "what": "Cloudflare's table, CLINC150+OOS · macro-F1: Jev",
   "value": 89.27,
   "shown": "89.27",
   "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": "cf7-clef",
   "what": "Cloudflare's table, BRIGHT · nDCG@10: Clef",
   "value": 45.91,
   "shown": "45.91",
   "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": "cf7-flash",
   "what": "Cloudflare's table, BRIGHT · nDCG@10: Clef-flash",
   "value": 39.26,
   "shown": "39.26",
   "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": "cf7-jev",
   "what": "Cloudflare's table, BRIGHT · nDCG@10: Jev",
   "value": 47.52,
   "shown": "47.52",
   "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": "cf8-clef",
   "what": "Cloudflare's table, Amazon ESCI · macro-F1: Clef",
   "value": 57.48,
   "shown": "57.48",
   "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": "cf8-flash",
   "what": "Cloudflare's table, Amazon ESCI · macro-F1: Clef-flash",
   "value": 57.39,
   "shown": "57.39",
   "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": "cf8-jev",
   "what": "Cloudflare's table, Amazon ESCI · macro-F1: Jev",
   "value": 55.21,
   "shown": "55.21",
   "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": "cf9-clef",
   "what": "Cloudflare's table, PhishNChips · accuracy: Clef",
   "value": 79.6,
   "shown": "79.60",
   "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": "cf9-flash",
   "what": "Cloudflare's table, PhishNChips · accuracy: Clef-flash",
   "value": 75.05,
   "shown": "75.05",
   "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": "cf9-jev",
   "what": "Cloudflare's table, PhishNChips · accuracy: Jev",
   "value": 62.55,
   "shown": "62.55",
   "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": "wf0-clef",
   "what": "Cloudflare's workflow table, Invoice processing: Clef",
   "value": 64.7,
   "shown": "64.7",
   "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": "wf0-flash",
   "what": "Cloudflare's workflow table, Invoice processing: Clef-flash",
   "value": 57.1,
   "shown": "57.1",
   "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": "wf0-jev",
   "what": "Cloudflare's workflow table, Invoice processing: Jev",
   "value": 61.8,
   "shown": "61.8",
   "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": "wf1-clef",
   "what": "Cloudflare's workflow table, Customer service: Clef",
   "value": 76.3,
   "shown": "76.3",
   "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": "wf1-flash",
   "what": "Cloudflare's workflow table, Customer service: Clef-flash",
   "value": 77.0,
   "shown": "77.0",
   "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": "wf1-jev",
   "what": "Cloudflare's workflow table, Customer service: Jev",
   "value": 76.0,
   "shown": "76.0",
   "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": "wf2-clef",
   "what": "Cloudflare's workflow table, Security incidents: Clef",
   "value": 62.9,
   "shown": "62.9",
   "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": "wf2-flash",
   "what": "Cloudflare's workflow table, Security incidents: Clef-flash",
   "value": 61.7,
   "shown": "61.7",
   "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": "wf2-jev",
   "what": "Cloudflare's workflow table, Security incidents: Jev",
   "value": 61.7,
   "shown": "61.7",
   "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": "wf3-clef",
   "what": "Cloudflare's workflow table, Agent trace observability: Clef",
   "value": 68.5,
   "shown": "68.5",
   "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": "wf3-flash",
   "what": "Cloudflare's workflow table, Agent trace observability: Clef-flash",
   "value": 69.8,
   "shown": "69.8",
   "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": "wf3-jev",
   "what": "Cloudflare's workflow table, Agent trace observability: Jev",
   "value": 71.6,
   "shown": "71.6",
   "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-lat-med-clef",
   "what": "Cloudflare's latency table, median ms: Clef",
   "value": 209.3,
   "shown": "209.3",
   "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-lat-med-flash",
   "what": "Cloudflare's latency table, median ms: Clef-flash",
   "value": 38.8,
   "shown": "38.8",
   "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-lat-med-jev",
   "what": "Cloudflare's latency table, median ms: Jev",
   "value": 524.1,
   "shown": "524.1",
   "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-lat-p95-clef",
   "what": "Cloudflare's latency table, p95 ms: Clef",
   "value": 238.6,
   "shown": "238.6",
   "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-lat-p95-flash",
   "what": "Cloudflare's latency table, p95 ms: Clef-flash",
   "value": 122.4,
   "shown": "122.4",
   "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-lat-p95-jev",
   "what": "Cloudflare's latency table, p95 ms: Jev",
   "value": 536.0,
   "shown": "536.0",
   "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-k9",
   "what": "Kev 9B (the board's older checkpoint) on BANKING77 in Cloudflare's table",
   "value": 84.83,
   "shown": "84.83",
   "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": "hf-tests",
   "what": "Tests in the Decision Index table on Clef's model card (our count)",
   "value": 41,
   "shown": "41",
   "whose": "Model Fatigue, count of Cloudflare's table",
   "source": "Clef's model card on Hugging Face (the full Decision Index table)",
   "url": "https://huggingface.co/Cloudflare/clef",
   "read": "2026-10-02T15:31",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "hf-clef-ahead",
   "what": "Of those, tests where Clef is ahead of Jev (our count)",
   "value": 26,
   "shown": "26",
   "whose": "Model Fatigue, count of Cloudflare's table",
   "source": "Clef's model card on Hugging Face (the full Decision Index table)",
   "url": "https://huggingface.co/Cloudflare/clef",
   "read": "2026-10-02T15:31",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "hf-jev-ahead",
   "what": "Of those, tests where Jev is ahead of Clef (our count)",
   "value": 15,
   "shown": "15",
   "whose": "Model Fatigue, count of Cloudflare's table",
   "source": "Clef's model card on Hugging Face (the full Decision Index table)",
   "url": "https://huggingface.co/Cloudflare/clef",
   "read": "2026-10-02T15:31",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "cf-evals",
   "what": "Benchmarks Cloudflare's post says it ran",
   "value": 43,
   "shown": "43",
   "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": "clef-mem-orig",
   "what": "Clef, share right on training messages",
   "value": 0.945,
   "shown": "94.5%",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-mem-para",
   "what": "Clef, share right on reworded copies of them",
   "value": 0.87,
   "shown": "87.0%",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-mem-orig",
   "what": "Clef-flash, share right on training messages",
   "value": 0.955,
   "shown": "95.5%",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-mem-para",
   "what": "Clef-flash, share right on reworded copies of them",
   "value": 0.835,
   "shown": "83.5%",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "mem-n",
   "what": "Training messages in the check (and as many reworded copies)",
   "value": 200,
   "shown": "200",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "knn-gap",
   "what": "A vote of the five nearest labelled examples, which can't remember anything: originals minus reworded copies",
   "value": 15.0,
   "shown": "+15.0 points",
   "whose": "Model Fatigue (M37's jev-2 run)",
   "source": "Our results write-up (RESULTS.md), jev-2's memorisation yardstick",
   "url": null,
   "read": "2026-09-21",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-same",
   "what": "Clef, same answer on a second pass",
   "value": 200,
   "shown": "200",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-same",
   "what": "Clef-flash, same answer on a second pass",
   "value": 200,
   "shown": "200",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "repeat-n",
   "what": "Messages in the second pass",
   "value": 200,
   "shown": "200",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-tok-clef",
   "what": "Clef's price per token against Jev's (our arithmetic)",
   "value": 5.7142857142857135,
   "shown": "5.7×",
   "whose": "Model Fatigue, arithmetic on the price lists",
   "source": "Cloudflare Workers AI pricing",
   "url": "https://developers.cloudflare.com/workers-ai/platform/pricing/",
   "read": "2026-10-02T20:52",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-tok-flash",
   "what": "Clef-flash's price per token against Jev's (our arithmetic)",
   "value": 2.142857142857143,
   "shown": "2.1×",
   "whose": "Model Fatigue, arithmetic on the price lists",
   "source": "Cloudflare Workers AI pricing",
   "url": "https://developers.cloudflare.com/workers-ai/platform/pricing/",
   "read": "2026-10-02T20:52",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-per1m",
   "what": "Jev, dollars per million messages (our arithmetic)",
   "value": 77.53194545454546,
   "shown": "$77.53",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-tok",
   "what": "Jev, input tokens per request (our arithmetic)",
   "value": 1845.9987012987012,
   "shown": "1,846",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-gain",
   "what": "Jev, share right with five examples minus on their request (our arithmetic)",
   "value": 13.214285714285712,
   "shown": "+13.2 points",
   "whose": "Model Fatigue, arithmetic on our runs",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-ref1k",
   "what": "Jev, sent to a person per 1,000 messages (our arithmetic)",
   "value": 156.16883116883116,
   "shown": "156",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-wrong1k",
   "what": "Jev, wrong answers routed without a person per 1,000 messages (our arithmetic)",
   "value": 17.207792207792206,
   "shown": "17.2",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-per1m",
   "what": "Clef, dollars per million messages (our arithmetic)",
   "value": 479.75135064935057,
   "shown": "$479.75",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-tok",
   "what": "Clef, input tokens per request (our arithmetic)",
   "value": 1998.963961038961,
   "shown": "1,999",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-gain",
   "what": "Clef, share right with five examples minus on their request (our arithmetic)",
   "value": 0.8441558441558472,
   "shown": "+0.8 points",
   "whose": "Model Fatigue, arithmetic on our runs",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-ref1k",
   "what": "Clef, sent to a person per 1,000 messages (our arithmetic)",
   "value": 67.85714285714286,
   "shown": "68",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-wrong1k",
   "what": "Clef, wrong answers routed without a person per 1,000 messages (our arithmetic)",
   "value": 22.4025974025974,
   "shown": "22.4",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-per1m",
   "what": "Clef-flash, dollars per million messages (our arithmetic)",
   "value": 179.90675649350652,
   "shown": "$179.91",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-tok",
   "what": "Clef-flash, input tokens per request (our arithmetic)",
   "value": 1998.963961038961,
   "shown": "1,999",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-gain",
   "what": "Clef-flash, share right with five examples minus on their request (our arithmetic)",
   "value": 4.285714285714281,
   "shown": "+4.3 points",
   "whose": "Model Fatigue, arithmetic on our runs",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-ref1k",
   "what": "Clef-flash, sent to a person per 1,000 messages (our arithmetic)",
   "value": 183.11688311688312,
   "shown": "183",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-wrong1k",
   "what": "Clef-flash, wrong answers routed without a person per 1,000 messages (our arithmetic)",
   "value": 17.532467532467532,
   "shown": "17.5",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-per1m",
   "what": "Kev-27B, dollars per million messages (our arithmetic)",
   "value": 729.5102727351169,
   "shown": "$729.51",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-tok",
   "what": "Kev-27B, input tokens per request (our arithmetic)",
   "value": 974.9639610389611,
   "shown": "975",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-gain",
   "what": "Kev-27B, share right with five examples minus on their request (our arithmetic)",
   "value": 7.012987012987015,
   "shown": "+7.0 points",
   "whose": "Model Fatigue, arithmetic on our runs",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-ref1k",
   "what": "Kev-27B, sent to a person per 1,000 messages (our arithmetic)",
   "value": 89.6103896103896,
   "shown": "90",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-wrong1k",
   "what": "Kev-27B, wrong answers routed without a person per 1,000 messages (our arithmetic)",
   "value": 20.77922077922078,
   "shown": "20.8",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-per1m",
   "what": "Kev-9B, dollars per million messages (our arithmetic)",
   "value": 621.1176265005325,
   "shown": "$621.12",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-tok",
   "what": "Kev-9B, input tokens per request (our arithmetic)",
   "value": 974.9639610389611,
   "shown": "975",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-gain",
   "what": "Kev-9B, share right with five examples minus on their request (our arithmetic)",
   "value": 8.019480519480515,
   "shown": "+8.0 points",
   "whose": "Model Fatigue, arithmetic on our runs",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-ref1k",
   "what": "Kev-9B, sent to a person per 1,000 messages (our arithmetic)",
   "value": 173.05194805194805,
   "shown": "173",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-wrong1k",
   "what": "Kev-9B, wrong answers routed without a person per 1,000 messages (our arithmetic)",
   "value": 20.454545454545453,
   "shown": "20.5",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "jev-di-per1k",
   "what": "Jev, dollars per 1,000 messages on the Decision Index's request (price list × tokens; our arithmetic)",
   "value": 0.07197231818181819,
   "shown": "$0.072",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-di-per1k",
   "what": "Clef, dollars per 1,000 messages on the Decision Index's request (price list × tokens; our arithmetic)",
   "value": 0.4477415064935064,
   "shown": "$0.448",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-di-per1k",
   "what": "Clef-flash, dollars per 1,000 messages on the Decision Index's request (price list × tokens; our arithmetic)",
   "value": 0.16790306493506493,
   "shown": "$0.168",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-vj",
   "what": "Clef minus Jev, share right with five examples, same messages",
   "value": 1.1363636363636365,
   "shown": "+1.14 points",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-vj-lo",
   "what": "Clef minus Jev, 95% interval, low",
   "value": 0.6168831168831169,
   "shown": "+0.62",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "clef-vj-hi",
   "what": "Clef minus Jev, 95% interval, high",
   "value": 1.6883116883116882,
   "shown": "+1.69",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-vj",
   "what": "Clef-flash minus Jev, share right with five examples, same messages",
   "value": 1.2337662337662338,
   "shown": "+1.23 points",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-vj-lo",
   "what": "Clef-flash minus Jev, 95% interval, low",
   "value": 0.6818181818181818,
   "shown": "+0.68",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-vj-hi",
   "what": "Clef-flash minus Jev, 95% interval, high",
   "value": 1.8181818181818181,
   "shown": "+1.82",
   "whose": "Model Fatigue",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vj",
   "what": "Kev-27B minus Jev, share right with five examples, same messages",
   "value": -0.09740259740259741,
   "shown": "−0.10 points",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vj-lo",
   "what": "Kev-27B minus Jev, 95% interval, low",
   "value": -0.551948051948052,
   "shown": "−0.55",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vj-hi",
   "what": "Kev-27B minus Jev, 95% interval, high",
   "value": 0.35714285714285715,
   "shown": "+0.36",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-vj",
   "what": "Kev-9B minus Jev, share right with five examples, same messages",
   "value": -2.5974025974025974,
   "shown": "−2.60 points",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-vj-lo",
   "what": "Kev-9B minus Jev, 95% interval, low",
   "value": -3.279220779220779,
   "shown": "−3.28",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-vj-hi",
   "what": "Kev-9B minus Jev, 95% interval, high",
   "value": -1.9155844155844155,
   "shown": "−1.92",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vc",
   "what": "Kev-27B minus Clef, share right with five examples, same messages",
   "value": -1.2337662337662338,
   "shown": "−1.23 points",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vc-lo",
   "what": "Kev-27B minus Clef, 95% interval, low",
   "value": -1.753246753246753,
   "shown": "−1.75",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-vc-hi",
   "what": "Kev-27B minus Clef, 95% interval, high",
   "value": -0.7142857142857143,
   "shown": "−0.71",
   "whose": "Model Fatigue",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "gap-di-f1",
   "what": "Clef minus Jev, macro-F1 on their request (our arithmetic)",
   "value": 14.202703552745504,
   "shown": "+14.2 points",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "gap-di-flash-f1",
   "what": "Clef-flash minus Jev, macro-F1 on their request (our arithmetic)",
   "value": 10.84689434009924,
   "shown": "+10.8 points",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "gap-di-acc",
   "what": "Clef minus Jev, share right on their request (our arithmetic)",
   "value": 13.5064935064935,
   "shown": "+13.5 points",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our replication of the Decision Index's BANKING77 request, all five models",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-clef",
   "what": "Clef's price per message against Jev's, with five examples (our arithmetic)",
   "value": 6.1877894052150895,
   "shown": "6.2×",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-flash",
   "what": "Clef-flash's price per message against Jev's, with five examples (our arithmetic)",
   "value": 2.320421026955659,
   "shown": "2.3×",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-flash-clef",
   "what": "Clef-flash's price per message as a share of Clef's (our arithmetic)",
   "value": 0.3750000000000001,
   "shown": "38%",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-tok-count",
   "what": "How many more input tokens Cloudflare's models count than Jev for the same request (our arithmetic)",
   "value": 8.286314591264098,
   "shown": "8%",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "r-k27",
   "what": "Kev-27B's cost per message as run against Jev's (our arithmetic)",
   "value": 9.40915732809524,
   "shown": "9.4×",
   "whose": "Model Fatigue, arithmetic on our runs",
   "source": "Our run of Kev-27B and Kev-9B on H100s we rented on Modal (Europe)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "saved1k",
   "what": "Fewer messages per 1,000 sent to a person by Clef than by Jev (our arithmetic)",
   "value": 88.3116883116883,
   "shown": "88",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "saved1k-exact",
   "what": "The same, unrounded (our arithmetic)",
   "value": 88.31168831168831,
   "shown": "88.3",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "extra1k",
   "what": "Clef's extra API cost per 1,000 messages over Jev's (our arithmetic)",
   "value": 0.40221940519480515,
   "shown": "$0.40",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "be-usd",
   "what": "Break-even: Clef's extra cost per person's check it saves (our arithmetic)",
   "value": 0.004554543264705882,
   "shown": "$0.0046",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "be-sec",
   "what": "Break-even as seconds of a person's time at that rate (our arithmetic)",
   "value": 1.0930903835294115,
   "shown": "1.1 s",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our arithmetic on the raw call records and Cloudflare's table (derived.json, this page's own file)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "morewrong1k",
   "what": "More wrong answers per 1,000 routed without a person by Clef than by Jev (our arithmetic)",
   "value": 5.194805194805195,
   "shown": "5.2",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "gap-acc-ex",
   "what": "Clef minus Jev, share right with five examples (our arithmetic)",
   "value": 1.1363636363636354,
   "shown": "+1.1 points",
   "whose": "Model Fatigue, arithmetic on our run",
   "source": "Our run with five examples: Jev, Clef and Clef-flash on the same day",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k27-overhead-share",
   "what": "Kev-27B, start-up and idle as a share of its bill (our arithmetic)",
   "value": 0.3544694312733237,
   "shown": "35%",
   "whose": "Model Fatigue, arithmetic on the bill",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "k9-overhead-share",
   "what": "Kev-9B, start-up and idle as a share of its bill (our arithmetic)",
   "value": 0.45800874834672545,
   "shown": "46%",
   "whose": "Model Fatigue, arithmetic on the bill",
   "source": "Modal's bill for our Kev deployment, read back per pass (kev_gpu_cost.json)",
   "url": null,
   "read": "2026-10-02",
   "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": "cf-x-clef",
   "what": "Jev's median time over Clef's in Cloudflare's table (our arithmetic)",
   "value": 2.5040611562350694,
   "shown": "2.5×",
   "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": "cf-x-flash",
   "what": "Jev's median time over Clef-flash's in Cloudflare's table (our arithmetic)",
   "value": 13.507731958762887,
   "shown": "13.5×",
   "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": "clef-mem-gap",
   "what": "Clef, originals minus reworded copies",
   "value": 7.5,
   "shown": "+7.5 points",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  },
  {
   "id": "flash-mem-gap",
   "what": "Clef-flash, originals minus reworded copies",
   "value": 12.0,
   "shown": "+12.0 points",
   "whose": "Model Fatigue",
   "source": "Our memorisation check: 200 training messages and reworded copies",
   "url": null,
   "read": "2026-10-02",
   "moves": false,
   "caveat": null,
   "now": null,
   "now_shown": null,
   "now_read": null
  }
 ]
}