{
  "_comment": [
    "Canonical numeric claims for the site. Single source of truth: index.html renders",
    "the receipts table from this file, so a number is corrected in exactly one place.",
    "",
    "`superseded` lists values this claim used to assert. scripts/check_claims.py fails",
    "the build if any of them reappears anywhere in the site, so a corrected number",
    "cannot quietly come back — which is exactly what happened to three of these: they",
    "survived a full restructure of index.html because the prose and the curated meta",
    "were edited independently."
  ],
  "generated_note": "Verified 2026-07-25 by running the suites and reading the code.",
  "claims": [
    {
      "claim": "MCP servers on the official registry",
      "detail": "published and installable from PyPI",
      "value": "8 servers",
      "source": "Xylem",
      "url": "https://github.com/jarmstrong158/Xylem",
      "superseded": [
        "7 servers",
        "seven servers"
      ]
    },
    {
      "claim": "agentsync under contention",
      "detail": "1,000 simulated concurrent races",
      "value": "0 lost / 0 double-granted",
      "source": "agentsync",
      "url": "https://github.com/jarmstrong158/agentsync"
    },
    {
      "claim": "context-keeper retrieval",
      "detail": "versus dumping the whole store",
      "value": "~92% less context; hit@5 70% (embedding) / 50% (lexical) on 59 problem-authored cases, frozen corpus",
      "source": "context-keeper",
      "url": "https://jarmstrong158.github.io/context-keeper/",
      "superseded": [
        "76% (embedding)",
        "Hit@5 13/15",
        "hit@5 76%",
        "hit@5 80%",
        "hit@5 93%"
      ],
      "note": "Measured 2026-08-05 by evals/run_retrieval_eval.py against the FROZEN 7-store corpus committed at evals/fixtures/corpus. An earlier published figure of 76% was measured against live stores before the corpus was frozen and no longer reproduces -- that is why the corpus is frozen. Lower than the retired synthetic figures on purpose: those queries were reworded from the entries they targeted."
    },
    {
      "claim": "Clark generalization",
      "detail": "held-out simulation, 5 to 50 workers",
      "value": "~93% completion",
      "source": "Clark",
      "url": "https://github.com/jarmstrong158/Clark"
    },
    {
      "claim": "Clark fine-tune cost",
      "detail": "~50 episodes on a consumer GPU",
      "value": "~3.3 h",
      "source": "Clark",
      "url": "https://github.com/jarmstrong158/Clark",
      "superseded": [
        "~30 min"
      ]
    },
    {
      "claim": "Jack single-facility agent",
      "detail": "across ~9.4 simulated years",
      "value": "98.2% completion",
      "source": "Jack",
      "url": "https://github.com/jarmstrong158/Jack"
    },
    {
      "claim": "skein test suite",
      "detail": "ruff + mypy clean in CI",
      "value": "130/130 passing",
      "source": "skein",
      "url": "https://github.com/jarmstrong158/skein",
      "superseded": [
        "71/71"
      ]
    },
    {
      "claim": "Conductor tool surface",
      "detail": "exposed to Claude over MCP",
      "value": "24 MCP tools",
      "source": "Conductor",
      "url": "https://github.com/jarmstrong158/Conductor",
      "superseded": [
        "22 MCP tools"
      ]
    },
    {
      "claim": "waveform-MCP tool surface",
      "detail": "counted from the server's own tool list",
      "value": "161 tools",
      "source": "waveform-MCP",
      "url": "https://github.com/jarmstrong158/waveform-MCP"
    },
    {
      "claim": "Measured automation savings",
      "detail": "warehouse reporting, proof of concept",
      "value": "~20 h / week",
      "source": "shipexec-metrics-tracker",
      "url": "https://github.com/jarmstrong158/shipexec-metrics-tracker"
    }
  ]
}