← All Test Cases

NIH-003

needle haystack

high

Repetitions

Documents

Questions

Reasoning

CROSS_SOURCE

retrieval-precision cross-source-with-distractors needle-in-haystack

📖 In Plain English

What this category tests

Can the brain find one specific document among many distractors?

How the test works

One target document with a unique canary token is ingested alongside 10+ similar-sounding distractor documents. The test uses the canary to find only the target.

Why it matters

Precision matters at scale — a brain that returns 'mostly relevant' results is unusable in production.

⚙️ How a single rep runs

① Generate

Model creates 12 synthetic documents and 1 question with unique canary tokens

→ Fresh content per run prevents memorization and proves real retrieval

② Ingest (MCP)

Model calls brain_ingest to store the 12 documents

→ Tests the brain's storage and indexing pipeline

③ Query (MCP)

Model answers the question using brain retrieval tools (search, fetch, context_pack, etc.)

→ Core test — does the brain return correct evidence and let the model build a faithful answer?

④ Evaluate

Model judges the answer against ground truth (the document it generated in phase 1)

→ Produces a score 0–100 with detailed sub-scores (retrieval, fidelity, reasoning, etc.)

This rep is run 3 times per test run. A pass requires score ≥ 85 and no critical failures.

🔬 Technical Instructions (raw prompts sent to AI)

🔧 ① Setup Instructions 1382 chars

Generate TWO target documents that form a cross-source chain, plus 10 distractors.
The question requires joining the two targets while ignoring all distractors.

Requirements:
- Target A: "<Entity> has code <X>."
- Target B: "Code <X> maps to <Property>."
- 10 distractors: mention similar entities, codes, or properties but do NOT form valid chains with A+B.
- source_ids: KB-{{RUN}}-NIH-003-A-v1, KB-{{RUN}}-NIH-003-B-v1, KB-{{RUN}}-NIH-003-D01-v1 through D10-v1.
- Question: "What is the <Property> of <Entity>?" — requires joining A+B, ignoring distractors.

Output ONLY this JSON:
{
  "run_id": "{{RUN}}",
  "test_id": "NIH-003",
  "repetition": {{REP}},
  "documents": [
    {"source_id": "KB-{{RUN}}-NIH-003-A-v1", "content": "<Entity> has code <X>.", "title": "<title>", "version": 1},
    {"source_id": "KB-{{RUN}}-NIH-003-B-v1", "content": "Code <X> maps to <Property>.", "title": "<title>", "version": 1},
    ... (10 distractors)
  ],
  "questions": ["What is the <Property> of <Entity>?"],
  "expected_answers": [{
    "question_index": 0,
    "correct_answer": "<Property>",
    "required_source_ids": ["KB-{{RUN}}-NIH-003-A-v1", "KB-{{RUN}}-NIH-003-B-v1"],
    "required_quote": null,
    "required_label": "CROSS_SOURCE",
    "must_refuse": false,
    "notes": "Must join A+B while ignoring 10 distractors. Distractors form false chains that look plausible."
  }]
}

▶ ② Execution Instructions 543 chars

You have access to MCP Knowledge Brain tools.

Use brain_search with the exact canary token from the question.
The canary is unique — only ONE document contains it.
Find that document. Return immediately once found — do not search further.

⚠️ OUTPUT RULE: As soon as you find the target document, stop reasoning
and output the JSON. Do not re-examine other documents. Do not verify.
Just output the required JSON schema with the found document's source_id and quote.

Question: {{QUESTION}}

Return ONLY this JSON:
{{REQUIRED_ANSWER_SCHEMA}}

⚖ ③ Judge Instructions 331 chars

Ground truth:
{{SETUP_CONTEXT_JSON}}

Tested AI answer:
{{EXEC_ANSWER_JSON}}

Must use both target docs (A+B) and ignore all distractors. CROSS_SOURCE label required.
Critical failures: used distractor sources in the chain, missing required source_ids, wrong property, DIRECT label.

Return ONLY this JSON:
{{JUDGE_RESULT_SCHEMA}}

Critical Failure Conditions

used_distractor_in_chain
missing_required_source_id
wrong_final_answer
wrong_label_not_cross_source

Recent Run History

2 runs

When	Run ID	Pass Rate	Avg Score	Reps
2026-05-24 13:08	20260524T130808Z-kqze	0%	0	0/1	View →
2026-05-24 11:37	20260524T113756Z-kduj	0%	0	0/1	View →

📄 Raw YAML cases/needle_haystack/NIH-003.yaml

schema_version: "1.0"
test_id: "NIH-003"
category: "needle_haystack"
severity: "high"
repetitions: 3
reasoning_type: "CROSS_SOURCE"
num_documents: 12
num_questions: 1
tags: ["retrieval-precision", "cross-source-with-distractors", "needle-in-haystack"]

setup_instructions: |
  Generate TWO target documents that form a cross-source chain, plus 10 distractors.
  The question requires joining the two targets while ignoring all distractors.

  Requirements:
  - Target A: "<Entity> has code <X>."
  - Target B: "Code <X> maps to <Property>."
  - 10 distractors: mention similar entities, codes, or properties but do NOT form valid chains with A+B.
  - source_ids: KB-{{RUN}}-NIH-003-A-v1, KB-{{RUN}}-NIH-003-B-v1, KB-{{RUN}}-NIH-003-D01-v1 through D10-v1.
  - Question: "What is the <Property> of <Entity>?" — requires joining A+B, ignoring distractors.

  Output ONLY this JSON:
  {
    "run_id": "{{RUN}}",
    "test_id": "NIH-003",
    "repetition": {{REP}},
    "documents": [
      {"source_id": "KB-{{RUN}}-NIH-003-A-v1", "content": "<Entity> has code <X>.", "title": "<title>", "version": 1},
      {"source_id": "KB-{{RUN}}-NIH-003-B-v1", "content": "Code <X> maps to <Property>.", "title": "<title>", "version": 1},
      ... (10 distractors)
    ],
    "questions": ["What is the <Property> of <Entity>?"],
    "expected_answers": [{
      "question_index": 0,
      "correct_answer": "<Property>",
      "required_source_ids": ["KB-{{RUN}}-NIH-003-A-v1", "KB-{{RUN}}-NIH-003-B-v1"],
      "required_quote": null,
      "required_label": "CROSS_SOURCE",
      "must_refuse": false,
      "notes": "Must join A+B while ignoring 10 distractors. Distractors form false chains that look plausible."
    }]
  }

execution_instructions: |
  You have access to MCP Knowledge Brain tools.

  Use brain_search with the exact canary token from the question.
  The canary is unique — only ONE document contains it.
  Find that document. Return immediately once found — do not search further.

  ⚠️ OUTPUT RULE: As soon as you find the target document, stop reasoning
  and output the JSON. Do not re-examine other documents. Do not verify.
  Just output the required JSON schema with the found document's source_id and quote.

  Question: {{QUESTION}}

  Return ONLY this JSON:
  {{REQUIRED_ANSWER_SCHEMA}}

judge_instructions: |
  Ground truth:
  {{SETUP_CONTEXT_JSON}}

  Tested AI answer:
  {{EXEC_ANSWER_JSON}}

  Must use both target docs (A+B) and ignore all distractors. CROSS_SOURCE label required.
  Critical failures: used distractor sources in the chain, missing required source_ids, wrong property, DIRECT label.

  Return ONLY this JSON:
  {{JUDGE_RESULT_SCHEMA}}

critical_failures:
  - "used_distractor_in_chain"
  - "missing_required_source_id"
  - "wrong_final_answer"
  - "wrong_label_not_cross_source"