{
  "schemaVersion": "kingy-document-report-narrative/1.0.0",
  "studyId": "document-comparison-v24-20261018.1",
  "createdAt": "2026-10-18T18:01:50.441741+00:00",
  "analysisCompletionSha256": "aaadd02fde4486dc2720f43034bc95913a2f69e9fcfc8eedc3609cb549fb8e7d",
  "finalAuditSha256": "f9d6f762dd963af631c2146a59457d7231554828520e8b1b89de3decde6a1b00",
  "sections": {
    "scoringRepair": {
      "title": "Reading the frozen results",
      "paragraphs": [
        "The independent source review found the requested factual values and roles correct for both workflows on all thirty answerable tasks. Both also cleanly declined all ten absent requests. The supported-answer score additionally requires the claimed source passages to support the requested facts; it is not a factual-answer-accuracy score alone.",
        "Kingy provided supporting passages on Q11 and Q29 where native LM Studio supplied correct values but claimed chunks that omitted the requested facts. This is a demonstrated source-support advantage on two unique cross-page tasks, repeated three times. Native gave shorter complete answers on several other tasks. No overall semantic winner follows.",
        "Native Q23 is complete and supported after semantic review, but its event-based deadline paraphrase remains unresolved under the frozen finite matcher. The original scores were not changed: each native repetition has 27 confirmed supported answers, one unresolved and two citation-support failures. All paired case-resampling envelopes include zero; forty unique tasks repeated three times are not 120 independent questions."
      ],
      "evidenceIds": [
        "scorer-validation",
        "corpus-review",
        "final-audit",
        "paired-results"
      ]
    },
    "engineeringRepair": {
      "title": "Repairs completed before the fresh corpus",
      "paragraphs": [
        "Candidate 1.5.6 binds each requested field to its full question context, keeps sibling facts separate, and checks that a selected complete-token source passage supplies that field. The candidate and scientific scorer were fixed before a separate agent authored this fresh five-PDF, thirty-page corpus.",
        "Kingy retrieves local PDF pages and uses the local model to select and check source passages, with up to four model requests per question. The observed result tests that whole configured workflow. Development replays and public smoke tests are documented separately from the 240 fresh study originals.",
        "A new Python environment and an exact fresh extraction of the final ZIP passed all four public installation selftests. The unauthenticated model API returned 401, permitted loopback returned 200, and the declared external IPv4, external IPv6 and wrong-loopback-port probes were denied. This is a same-Mac installation check with existing prerequisites, not novice-user or second-machine validation."
      ],
      "evidenceIds": [
        "candidate-changes",
        "scorer-validation",
        "freeze",
        "installation",
        "selftest-raw",
        "network-checks"
      ]
    },
    "control": {
      "title": "Prospective native controls and separate timing boundaries",
      "paragraphs": [
        "The actual study retained all 240 planned originals in the frozen A1, B1, B2, A2, A3, B3 order. Each native original used one journaled external System Events AXPress, same-conversation acceptance, the exact five PDFs, matching settings, and original completion events. Previous incomplete studies and repairs remain preserved.",
        "Preparation, document import, cold load, the single document-free warmup and inter-original evidence-review pauses are separate from per-query timing. Native results retain both the original dispatch watchdog and submission-to-final-event interval. No clock was reset to make a late observation pass. Latency does not measure automation throughput, novice setup time or a general speed advantage.",
        "These results cover one shared M4 Pro Mac, one synthetic English selectable-text corpus, one Qwen artifact and the declared runtime versions. LM Studio also documents local offline document chat. This study does not prove an exclusive privacy feature, cloud-cost saving, GPU-upgrade benefit or a market-wide product ranking.",
        "All thirty answerable v24 questions require facts from one PDF each; five span adjacent pages within that PDF. None requires combining facts across different PDFs. Cross-document answer synthesis is unmeasured. The generic multi-document corpus description in the immutable tested ZIP README does not describe v24; this study-specific correction takes precedence. The installed package, original answers and frozen scores are unchanged."
      ],
      "evidenceIds": [
        "control-provenance",
        "completion-eligibility",
        "protocol",
        "freeze",
        "session-provenance",
        "final-audit"
      ]
    }
  }
}