{
  "updated": "2026-11-04",
  "purpose": "Technical review for Charles Ferguson / EWOR",
  "papers": [
    {
      "path": "/tech/papers/01-compiler.pdf",
      "sha256": "23c6b333052cd92cbfbf6a9191d897d192638b217270735914b7629e40034197",
      "status": "research draft; not peer reviewed"
    },
    {
      "path": "/tech/papers/02-authority.pdf",
      "sha256": "0aab34d4967cfc4f117f7e48bb935755c477b32f5fa3314bbbbe6e9717d76e3e",
      "status": "research draft; not peer reviewed"
    },
    {
      "path": "/tech/papers/04-sdk.pdf",
      "sha256": "2b274613657eb26a91cf7ea2b58dc890de6961a7429db4a28c0ac1a71b3c868a",
      "status": "research draft; not peer reviewed"
    },
    {
      "path": "/tech/papers/05-device.pdf",
      "sha256": "38ee1815e19aaf9f1d37d02ef970574ea8ea4cc765cd577ef7ded01f0095882b",
      "status": "research draft; not peer reviewed"
    }
  ],
  "claims": [
    {
      "study": "sdk-trained-first-oct25",
      "cohort": 32,
      "trained_correct": 32,
      "bare_correct": 16,
      "steered_teacher_correct": 32,
      "paired_wins": 16,
      "paired_losses": 0,
      "training_updates_in_confirmation": 0,
      "prior_training_updates": 48,
      "operational_calls_in_confirmation": 0,
      "scope": "Synthetic city/code/color profiles; Qwen2.5 family; one Mac; one trained checkpoint",
      "evaluation_sha256": "91677794f014268c7eb53af37f14eb48e580c9ebd2dd859c31f89229fa4b8627",
      "audit_sha256": "ed20dc868b07da0fb382d512dec34ca13f7010d6d26bfd0ddf7f65a5cda036bd",
      "task_definition": "One structured JSON reply per fictional profile: current city, favorite color and deleted code as UNKNOWN. Correct current facts are supplied by SDK before generation.",
      "pass_definition": "Normal completion, valid JSON, exactly three string fields, every value correct, no deleted-code exposure; no repair.",
      "failure_breakdown": "Tallied post hoc on the calibration split only: all 16 bare-student failures are one error, an unquoted UNKNOWN, on two of four query wordings. Held-out and new-cohort raw answers were not tallied, so for this cohort the reading is an inference from matching counts.",
      "interpretation": "The gain may be largely output formatting, and prompt wording is an unresolved alternative. An earlier, unmatched evaluation of this checkpoint under a fixed JSON grammar found the bare and trained students tied at 32/32."
    },
    {
      "study": "sdk-portability-v14-oct24",
      "independent_episodes": 8,
      "repeats": 2,
      "calls": 64,
      "soul_correct_each_size": 8,
      "history_correct_15b": 0,
      "history_correct_3b": 2,
      "median_prompt_reduction_percent": 55.1532,
      "scope": "State archive portability, not adapter transfer"
    },
    {
      "study": "sdk-portability-v15-oct24",
      "profiles": 32,
      "calls": 64,
      "correct_each_size": 32,
      "training": 0,
      "scope": "Fresh qualification; no history comparator in this cohort"
    },
    {
      "study": "sdk-checkpoint-routing-oct24",
      "native_calls": 4,
      "scope": "Separate narrow operational control-flow probe"
    },
    {
      "study": "integrated-v2-oct25",
      "heldout_profiles": 32,
      "correct": {
        "history_small": 6,
        "soul_small": 16,
        "distilled_small": 32,
        "combined_small": 27,
        "soul_large": 32,
        "combined_large": 32
      },
      "verdict": "FAIL",
      "verdict_note": "The primary gain threshold was arithmetically unreachable once a baseline reached ceiling, so the verdict is a property of these gates on a ceiling task.",
      "local_fraction": 0
    },
    {
      "study": "proxy-soul-002",
      "wins": 27,
      "losses": 7,
      "ties": 6,
      "scope": "One judge's preference, by the person who built the system, for a steered package over longer history. Steering is not isolated as the cause. Recognition after exposure and second-participant limits retained. Its source draft is not posted at present."
    }
  ],
  "proposed_targets": {
    "quality_relative_to_teacher": "noninferiority lower95%bound above\u22125percentagepoints",
    "local_share_min": 0.5,
    "cloud_token_reduction_min": 0.3,
    "planning_users": 30,
    "heldout_tasks": 1000
  },
  "limitations": [
    "No general frontier superiority",
    "No continuous-learning or total-cost-saving claim",
    "No new model family or cross-family adapter portability demonstrated",
    "Draft paper count is not independent replication count",
    "Independent execution checks are not external peer review"
  ],
  "narrative_updated": "2026-10-25: original 20-slide story and graphics retained; latest results visible beside diagrams, measured blocks relabeled, direct paper links on every slide and closing slide.",
  "demonstration": {
    "study": "sdk-desktop-chat-factory-v2-oct25",
    "pairs": 4,
    "native_calls": 8,
    "observed": "Both comparison arms followed Lisbon, Porto and deletion to UNKNOWN. Both missed the separate lunch prompt two-sentence request.",
    "scope": "Live interaction check; no quality superiority claim; training, steering and teacher fallback off.",
    "audit_sha256": "99f77aca9636ae3dc7c3ae7f1c3915eb1be8b0ca0bc38babe9cb92b8b5ed8a85"
  }
}