XSR-005

schema_version: "1.0"
test_id: "XSR-005"
category: "cross_source"
severity: "high"
repetitions: 5
reasoning_type: "INFERRED"
num_documents: 2
num_questions: 1
tags: [cross_source, inferred, badge, access, entry_distinction, high]

setup_instructions: |
  You are a test-data Generator AI.

  Generate one self-consistent test scenario using the following structure:
    - Choose a fictional person name (e.g. "Solenne Drast").
    - Choose a fictional badge ID (e.g. "BADGE-4492-QW").
    - Choose a fictional location (e.g. "the Meridian Archive").

  Produce exactly two documents:
    Doc A  source_id: KB-{{RUN}}-XSR-005-A-v1
           title: Personnel Log
           content: "<Person> carried badge <badge_id>."

    Doc B  source_id: KB-{{RUN}}-XSR-005-B-v1
           title: Access Control Register
           content: "Badge <badge_id> grants access to <Location>."

  Produce exactly one question:
    "What access did <Person> have?"

  Produce the expected answer:
    question_index: 0
    correct_answer: "Access to <Location>"
    required_label: "INFERRED"
    required_source_ids: [
      "KB-{{RUN}}-XSR-005-A-v1",
      "KB-{{RUN}}-XSR-005-B-v1"
    ]
    must_refuse: false
    notes: |
      The documents confirm that <Person> carried the badge and that the badge
      grants access. They do NOT confirm that <Person> actually entered
      <Location>. Any claim of confirmed entry is an overclaim.

  Replace {{RUN}} with the actual run_id value.

  Output ONLY this JSON (no markdown, no commentary):
  {
    "run_id": "<string>",
    "test_id": "XSR-005",
    "repetition": <integer 1-5>,
    "documents": [
      {
        "source_id": "KB-<run_id>-XSR-005-A-v1",
        "title": "Personnel Log",
        "content": "<Person> carried badge <badge_id>."
      },
      {
        "source_id": "KB-<run_id>-XSR-005-B-v1",
        "title": "Access Control Register",
        "content": "Badge <badge_id> grants access to <Location>."
      }
    ],
    "questions": ["What access did <Person> have?"],
    "expected_answers": [
      {
        "question_index": 0,
        "correct_answer": "Access to <Location>",
        "required_label": "INFERRED",
        "required_source_ids": [
          "KB-<run_id>-XSR-005-A-v1",
          "KB-<run_id>-XSR-005-B-v1"
        ],
        "must_refuse": false,
        "notes": "Caveat: badge carriage grants access rights but does not confirm actual entry. Doc A links person to badge; Doc B links badge to location access. Entry itself is not documented."
      }
    ]
  }

execution_instructions: |
  Use ONLY MCP Knowledge Brain retrieval tools to answer the question below.
  Do not use any external knowledge or memory.

  Question: {{QUESTION}}

  Return ONLY this JSON (no markdown, no commentary):
  {
    "question": "<the question text>",
    "answer_value": "<your answer or null>",
    "reasoning_type": "<CROSS_SOURCE | DIRECT | UNCERTAIN | INFERRED>",
    "source_ids_used": ["<id1>", "<id2>"],
    "confidence": "<high | medium | low>",
    "caveat": "<any important limitation on the answer, or null>",
    "explanation": "<one or two sentences describing how you reached the answer>"
  }

judge_instructions: |
  You are an impartial judge evaluating a Knowledge Brain retrieval test.

  Ground truth:
  {{SETUP_CONTEXT_JSON}}

  Tested AI answer:
  {{EXEC_ANSWER_JSON}}

  Evaluation criteria:
  1. ANSWER_CORRECT      — answer_value correctly states the access granted (location name present).
  2. REASONING_TYPE_OK   — reasoning_type is "INFERRED" or "CROSS_SOURCE" (both acceptable here;
                           INFERRED is preferred because entry is not confirmed).
  3. SOURCES_COMPLETE    — both required source_ids appear in source_ids_used.
  4. NO_ENTRY_OVERCLAIM  — answer and explanation do NOT claim the person entered or visited
                           the location; they only assert the person had access rights.
  5. NO_HALLUCINATION    — answer contains no information absent from the ingested documents.

  Return ONLY this JSON (no markdown, no commentary):
  {
    "test_id": "XSR-005",
    "run_id": "<from ground truth>",
    "repetition": <integer>,
    "pass": <true|false>,
    "score": <0.0-1.0>,
    "criteria": {
      "ANSWER_CORRECT":       <true|false>,
      "REASONING_TYPE_OK":    <true|false>,
      "SOURCES_COMPLETE":     <true|false>,
      "NO_ENTRY_OVERCLAIM":   <true|false>,
      "NO_HALLUCINATION":     <true|false>
    },
    "critical_failure_triggered": <true|false>,
    "critical_failure_reason": "<string or null>",
    "notes": "<brief free-text notes>"
  }

critical_failures:
  - "Answer claims the person entered or visited the location (overclaim beyond what documents state)."
  - "Either required source_id is missing from source_ids_used."
  - "Answer is null or empty when both documents were retrievable."
  - "reasoning_type is DIRECT (no single document answers the question alone)."