{
  "version": "2026-08-18.1",
  "scenarios": [
    {
      "id": "SCENARIO-CHANGING-KNOWLEDGE-BASE",
      "title": "Changing support knowledge base",
      "situation": "Support answers must reflect maintained policies and link the user to the passages used.",
      "evidencePattern": ["fresh_facts=yes", "source_citations=yes", "authoritative_corpus=yes", "retrieval_operations=yes", "evaluation_set=yes"],
      "candidateOutcome": "RAG",
      "why": "The primary gap is a maintained knowledge path with traceability, not a new behavioral contract.",
      "firstExperiment": "Compare a prompt-only baseline with retrieval and score passage relevance, claim support, answer quality, and stale-index failures.",
      "sourceIds": ["SOURCE-RAG-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT"],
      "claimIds": ["CLAIM-RAG-FRESHNESS-FIT", "CLAIM-RAG-PROVENANCE"]
    },
    {
      "id": "SCENARIO-STABLE-STRUCTURED-TASK",
      "title": "Stable structured extraction task",
      "situation": "A stable input class must produce a consistent schema, and representative permitted examples already exist.",
      "evidencePattern": ["behavior_change=yes", "representative_examples=yes", "training_operations=yes", "stable_narrow_task=yes", "evaluation_set=yes"],
      "candidateOutcome": "FINE_TUNING",
      "why": "The target is repeatable task behavior rather than changing external facts.",
      "firstExperiment": "Compare prompt-only and tuned candidates on held-out schema validity, task errors, regressions, and operational rollback.",
      "sourceIds": ["SOURCE-OPENAI-MODEL-OPTIMIZATION", "SOURCE-LORA-PAPER"],
      "claimIds": ["CLAIM-FINE-TUNING-BEHAVIOR", "CLAIM-EXAMPLES-REQUIRED", "CLAIM-LORA-METHOD"]
    },
    {
      "id": "SCENARIO-DOMAIN-ASSISTANT",
      "title": "Domain assistant with specialized behavior",
      "situation": "The assistant must use a changing domain corpus and also follow a stable response or reasoning contract represented in examples.",
      "evidencePattern": ["fresh_facts=yes", "behavior_change=yes", "representative_examples=yes", "authoritative_corpus=yes", "retrieval_operations=yes", "training_operations=yes", "stable_narrow_task=yes", "evaluation_set=yes"],
      "candidateOutcome": "HYBRID",
      "why": "The knowledge and behavior gaps are independent and both lifecycle capabilities are present.",
      "firstExperiment": "Test retrieval and behavior adaptation separately; combine them only if each produces a measured improvement without unacceptable regressions.",
      "sourceIds": ["SOURCE-RAFT-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT"],
      "claimIds": ["CLAIM-HYBRID-PATTERN", "CLAIM-DUAL-OPERATIONS"]
    },
    {
      "id": "SCENARIO-NO-EVALUATION",
      "title": "Architecture request without an evaluation set",
      "situation": "Stakeholders request RAG or tuning, but no versioned tasks or acceptance criteria distinguish a real improvement from a demo.",
      "evidencePattern": ["evaluation_set=no"],
      "candidateOutcome": "INSUFFICIENT_EVIDENCE",
      "why": "The tool declines to choose when alternatives cannot be measured against the same workload contract.",
      "firstExperiment": "Create the evaluation set and baseline before implementing either architecture.",
      "sourceIds": ["SOURCE-OPENAI-MODEL-OPTIMIZATION"],
      "claimIds": ["CLAIM-EVAL-FIRST"]
    }
  ]
}