{
  "claim": "Every headline multiplier in this category, put next to the number somebody got when they ran both arms of the comparison. None of them is fabricated and none of them survives its own denominator. The largest surviving gap in the whole table is 16x, and it is a price-per-token ratio rather than a property of the architecture.",
  "method": "Each 'measured' figure is the one computed in this site's own teardown of that claim, from the artifact the claimant published: openroboto-ai/jev-robot-control's seed-0 pair for the speed and cost multipliers, nekuda-ai/WindTunnel's canonical results.csv for the interface/price factorisation, stas4000/jev-linkmap's out/ and runs/ for the link-map costs, mizorewww/laya-coreml's own benchmark JSON for the Apple silicon ratios, notnotsamuel/LFM2.5-350M-RLCD's committed cases for the speedup regression. Each row links back to the piece that did the arithmetic.",
  "source": "This site's own articles: /articles/three-tiers, /articles/webmcp-windtunnel, /articles/jev-linkmap, /articles/laya-mlx, /articles/any-model-can-be-jev, /articles/cua-s1-forms.",
  "captured": "2026-09-22",
  "note": "The pattern is not that anyone lied. In every case the claimed ratio is a real division of two real numbers; what moves is which two. Three failure shapes recur: a hosted round trip divided by a local forward pass (the Atlantic, not the model), a bounded question compared against a model writing a reasoning trace (a baseline the argument itself says to stop using), and a recurring cost quoted without the one-time cost that made it work.",
  "columns": [
    { "key": "claim", "label": "the claim as it travels" },
    { "key": "where", "label": "whose" },
    { "key": "measured", "label": "measured", "align": "right", "mono": true },
    { "key": "gap", "label": "what the difference is" }
  ],
  "rows": [
    {
      "claim": "193.6x faster",
      "where": "TypeSafe homepage; relayed as \"up to 200x\", recirculated this week as Jev Engineering",
      "measured": "4.57x",
      "gap": "per-call latency, both arms on one xArm7 task. 193.6x needs the LLM side to spend 136.9 s per call — a reasoning trace, not a classification."
    },
    {
      "claim": "444.6x cheaper",
      "where": "same",
      "measured": "336.0x",
      "gap": "per-call cost, same trial. Same order of magnitude; the gap is task mix. This is the one headline that broadly survives."
    },
    {
      "claim": "245x cheaper",
      "where": "WindTunnel board, vs GPT-6 Astra on screenshots",
      "measured": "15.25x × 16.07x",
      "gap": "Interface term times price term. Hold the interface at WebMCP and the price term is all that is left; against the best conventional WebMCP row it is 2.3x."
    },
    {
      "claim": "112x cheaper",
      "where": "same board, vs code execution",
      "measured": "6.94x × 16.07x",
      "gap": "Same 16.07x price term. The interface half is the transferable one, and it is available to any model that uses it."
    },
    {
      "claim": "240x cheaper than Opus",
      "where": "jev-linkmap announcement",
      "measured": "3.7x",
      "gap": "$17.91 all-in against $66.37, once the $15.51 rubric loop and the $2.07 editor pass are counted. Break-even on the tuning is 32,147 pages; the site has 566."
    },
    {
      "claim": "zero pages against 566",
      "where": "same, the race chart",
      "measured": "18x",
      "gap": "Per page, matched on the same 120 pages. A cost claim and a throughput claim drawn as one fact. At Jev's own concurrency Opus finishes the site in 3.3 minutes."
    },
    {
      "claim": "50x faster than Jev",
      "where": "attributed to laya-coreml, which never says it",
      "measured": "1.394x",
      "gap": "The only speed ratio either repo publishes about itself — Neural Engine against compiled MLX, same machine. The 50x divides a local forward pass into an HTTPS round trip."
    },
    {
      "claim": "63x faster",
      "where": "LFM2.5-350M-RLCD",
      "measured": "1.09x – 62.91x",
      "gap": "The ratio is (95.4 + 15.21·G) / (48.6 + 0.0392·B): it is how many tokens you made the baseline generate. The 62.91x case generated 226; the 1.09x case generated 13."
    },
    {
      "claim": "roughly 60 decisions per second",
      "where": "laya-coreml, via a third-party summary",
      "measured": "49.66/s",
      "gap": "60 is the row in the repo's own sweep that misses its deadline on 100% of ticks. A Snake decision is also three calls carrying 4, 2 and 2 options, not one."
    },
    {
      "claim": "at most 1 GB",
      "where": "same summary",
      "measured": "1,833 MiB",
      "gap": "943.6 MiB is peak MLX allocation for one short question. Ten full-context questions is 1.79 GiB. Neither is process RSS, and the weights are unquantised FP16."
    },
    {
      "claim": "99.7% against Jev's 83.6%",
      "where": "cua-s1-forms",
      "measured": "in-domain vs cold",
      "gap": "A 706K-parameter specialist measured on data from the generator it trained against, versus a generalist that has never seen the distribution. The release says so itself."
    },
    {
      "claim": "only a mild hit in accuracy",
      "where": "said of jeff, a frozen-encoder server",
      "measured": "37.8% vs 72.1%",
      "gap": "JevBench hard tier, using JevBench's own runner. 0 of 7 on ambiguous items, 1 of 19 on long policies. Latency and cost claims hold; the third one does not."
    },
    {
      "claim": "45% to 65% recall for $15.51",
      "where": "jev-linkmap's System 2 loop",
      "measured": "64.2%, for $0",
      "gap": "Moving two thresholds on data already collected captures 91% of the gain, at higher precision. The two rubric rewrites add the last 9%."
    }
  ]
}