Source Verification — Tracing Claims Back to Exact Passages

Source Verification

Every claim in the brain is traceable back to the exact passage in the original text. This enables fact-checking, auditing, and trust.

How it works

When a source is added with /dist:add, the pipeline:

1. Saves chunks permanently in brain/sources/{slug}/chunks/ — verbatim text from the original PDF/EPUB
2. Extracts claims with passage references — each claim points to the exact chunk and line range it came from
3. Captures evidence with backing — 9 universal evidence categories with warrants explaining WHY the evidence supports the claim

# Every claim has this in its frontmatter:
passages:
  - chunk: "chunk_01.txt"
    lines: [115, 118]
    snippet: "Multi-factor authentication for remote access..."
  - chunk: "chunk_02.txt"
    lines: [39, 40]
    snippet: "Multi-factor authentication for users' access..."
confidence: exact
backing:
  - category: textual
    subtype: "regulatory requirement"
    ref: "ECC 2-2-3-2"
    snippet: "Organizations shall implement multi-factor..."
    strength: definitive
    warrant: "Direct regulatory mandate requiring MFA"

Fact-checking a claim

To verify any claim:

1. Read the claim’s passages: field
2. Open the referenced chunk file at the line range
3. Compare the claim’s proposition to the full surrounding context

Or use the verify command to automate this:

/dist:verify brain/sources/kiyosaki-rich-dad/
/dist:verify all

The verify agent reads each claim, opens its referenced chunks, and reports:

• VERIFIED — passage fully supports the claim
• PARTIALLY VERIFIED — passage supports but claim oversimplifies
• UNVERIFIED — passage does not support the claim

Confidence levels

Level	Meaning
exact	One passage directly states what the claim asserts
synthesized	Claim combines multiple passages (all support it, none states it alone)
inferred	Claim is derived through reasoning, not directly stated

Research agent integration (Method K)

The deep research agent uses source verification as its final quality check:

/dist:research what does the brain say about integrity?

When the research agent cites a claim as evidence, it checks the claim’s passages: entries against the chunks. If a passage doesn’t support the claim, the agent flags it and adjusts confidence.

Doctor integration

The doctor agent flags quality issues:

/dist:doctor

It catches:

• Claims missing the passages: field
• Claims where confidence: exact but multiple passages suggest synthesized
• Claims that need re-extraction

Publishing and copyright

Chunks are always stored locally — they cost nothing and enable full verification on your machine.

The copyright question only matters when publishing:

# In _source.md:
chunks_available: true    # always true
publishable: true         # false for copyrighted sources

When you run /dist:publish:

• Sources with publishable: true → chunks included on the website (full remote verification)
• Sources with publishable: false → chunks excluded (claims and snippets still travel, but full text stays local)

What verification does NOT do

• Does not verify the source itself is correct (e.g., whether a hadith is actually sahih)
• Does not fix OCR errors in the original PDF extraction
• Does not replace human judgment — it surfaces the evidence for humans to evaluate

It makes the brain auditable: every claim has a paper trail back to the exact words in the original source.