{
  "claim": "The five limits people attribute to the System One shape are not one limit. Two of them are forced by bounded output and hold for every family. Two more are the opposite prices of a decision the shape leaves open, so no model pays both. One is not a limit of the architecture at all, and a sixth group of failures that gets filed with them belongs to no architecture in particular.",
  "method": "Each row names the family it actually applies to and whether the limit follows from the architecture or merely co-occurs with it. Reported = a vendor or benchmark author says it. Measured = read or computed from a committed artifact. Derived = argued from the shape, with a falsifier in the article.",
  "captured": "2026-09-19",
  "note": "The 'applies to' column is the one that matters. A claim that 'decision models are order-sensitive' is false of every per-option scorer, and a claim that 'decision models cannot answer relational questions' is not forced for any model whose options share a context — it is just what Jev, and the two open models tested alongside it, happen to do.",
  "columns": [
    { "key": "limit", "label": "reported limit" },
    { "key": "family", "label": "applies to" },
    { "key": "status", "label": "status" },
    { "key": "evidence", "label": "evidence" }
  ],
  "rows": [
    {
      "limit": "cannot produce text that is not already an option",
      "family": "every family",
      "status": "Derived · forced by bounded output",
      "evidence": "TypeSafe: \"not trained to generate text\". Measured: 385 of 413 WebMCP decisions were handed to a generative model, and 379 of those filled at least one unconstrained string."
    },
    {
      "limit": "weak multi-step reasoning and planning",
      "family": "every family",
      "status": "Derived · partly. Bounded output forbids a scratchpad; some of the reported weakness is ordinary.",
      "evidence": "TypeSafe: \"multiple hops of reasoning costs accuracy\". json-render: planning is \"a documented weakness of Jev\". Reported: 1 of 120 ten-hop WikiRouter routes."
    },
    {
      "limit": "cannot answer a question spanning two options",
      "family": "per-option scorers only",
      "status": "Derived · from option isolation, not from bounded output",
      "evidence": "Reported: Jev 0/100 on relational choice; Laya 8/100; a prompted Qwen3.5-4B readout, whose options share a context, 53/100."
    },
    {
      "limit": "option order changes the answer",
      "family": "shared-context readouts only",
      "status": "Derived · from position carrying identity. Not expressible in a per-option scorer.",
      "evidence": "Measured: openjev flips 10 of 36 reversed questions, and slot A is worth +1.71 logits. Reported: Laya 0.150 and 0.230 instability at 20 options."
    },
    {
      "limit": "a ceiling on how many options you may send",
      "family": "no single family",
      "status": "Not one limit · five ceilings with five unrelated causes, one of which is no ceiling",
      "evidence": "Measured: 16, about 20, 255, a fixed K welded into a head, and none at all. See the next table."
    },
    {
      "limit": "literal reading, arithmetic, date ordering, context rot, prompt injection, contradictory criteria",
      "family": "no architecture in particular",
      "status": "Not derived · ordinary weaknesses, sitting on the same page as the others",
      "evidence": "Six of the eight entries on TypeSafe's jaggedness page. A prompted 400B model has all six; none of them follows from the option set being the output."
    }
  ]
}