{
  "version": "2026-08-18.1",
  "name": "Evidence-bound RAG versus fine-tuning decision method",
  "publishedOn": "2026-08-18",
  "publisher": "Pharos Production",
  "purpose": "Classify the next architecture experiment without converting qualitative evidence into a probability or universal score.",
  "allowedAnswers": ["yes", "no", "unknown"],
  "priorityDirection": "ascending",
  "outcomes": [
    {
      "id": "RAG",
      "label": "RAG candidate",
      "meaning": "Test a retrieval-grounded baseline because changing or traceable knowledge is required and the corpus and retrieval operating path exist."
    },
    {
      "id": "FINE_TUNING",
      "label": "Fine-tuning candidate",
      "meaning": "Test parameter adaptation because the primary need is stable task behavior and representative examples and training operations exist."
    },
    {
      "id": "HYBRID",
      "label": "Hybrid candidate",
      "meaning": "Test retrieval plus behavior adaptation because both needs and both operational paths are independently evidenced."
    },
    {
      "id": "BASELINE_FIRST",
      "label": "Baseline first",
      "meaning": "Establish or improve a prompting/evaluation baseline before adding retrieval or training complexity."
    },
    {
      "id": "BLOCKED",
      "label": "Blocked",
      "meaning": "A required data or operating dependency is absent; a plausible architecture label cannot remove that blocker."
    },
    {
      "id": "INSUFFICIENT_EVIDENCE",
      "label": "Insufficient evidence",
      "meaning": "A core input or evaluation contract is unresolved, so the method declines to select a winner."
    }
  ],
  "decisionProcedure": [
    "Validate that every registered criterion has exactly one allowed answer and reject unknown criterion IDs.",
    "Evaluate decision rules in ascending unique priority and select the first matching outcome.",
    "Evaluate advisory rules independently and attach their tests, risks, and provenance without changing the outcome.",
    "Export the answers, winning rule, triggered advisories, blockers, unresolved inputs, source IDs, claim IDs, and registry versions."
  ],
  "evaluationContract": {
    "requiredBeforeWinner": true,
    "minimumJobs": [
      "Define representative tasks and failure classes.",
      "Record a prompt-only or current-system baseline.",
      "For retrieval, test retrieval relevance and claim-to-passage support separately from final-answer quality.",
      "For tuning, test target behavior and regression on capabilities that must be preserved.",
      "Measure latency, cost, privacy paths, and operational recovery in the intended topology."
    ]
  },
  "limitations": [
    "The outcome is a candidate for a bounded experiment, not an implementation guarantee.",
    "The method contains no universal dataset-size, accuracy, latency, document-count, or cost threshold.",
    "Provider features and data controls can change after the dated source verification.",
    "Independent human technical review of this microsite is not claimed."
  ],
  "disclosure": "Research synthesis and implementation were AI-assisted; claims are bounded to the listed sources and verification date."
}