How Distillary works

The insight

Every source (book, video, article, regulatory framework) has an argument structure: a central thesis supported by a hierarchy of claims backed by evidence. Distillary makes that structure explicit, navigable, and verifiable.

The pipeline (11 steps)

Run the full pipeline with one command:

/dist:add books/my-book.pdf

What happens under the hood:

Source text (book, PDF, video transcript, article)
  ↓ split into ~20KB chunks
  ↓ save chunks permanently to brain/sources/{slug}/chunks/
  ↓ parallel haiku agents → atomic claims with backing + passages
  ↓ haiku agent → deduplicate (merge passages lists)
  ↓ haiku agent → extract entities (people, concepts, frameworks)
  ↓ haiku agent → add [[wikilinks]] to claim bodies (preserve passages in frontmatter)
  ↓ opus agents → group claims into argumentative clusters (layer 1)
  ↓ opus agent → build clusters into structure claims (layer 2) + root thesis (layer 3)
  ↓ haiku agents → find tensions, patterns, evidence between claims
  ↓ haiku agent → verify sample of claims against source chunks (optional)
  ↓ Python → assemble vault, fix links, build entity hubs, doctor
  ↓ opus agent → map concepts to other sources in the brain
  ↓ haiku agent → build bridge entities for same-concept pairs
Brain vault

Chunks are always saved locally — no question asked. They cost nothing and enable fact-checking every claim. Copyright only matters at publish time: sources marked publishable: false in _source.md have their chunks excluded from the public site.

The pyramid

Layer 3: Root thesis (1 note)
  "The Lean Startup engineers sustainable businesses
   through validated learning"
    ↓
Layer 2: Structure (~2-3 notes)
  "Pivots and growth engines drive course correction"
  "Innovation accounting replaces vanity metrics"
    ↓
Layer 1: Clusters (~5-8 notes)
  "Small batches accelerate the feedback loop"
  "Customer validation precedes scaling"
    ↓
Layer 0: Atoms (~50-300 notes)
  "Zappos tested demand by posting shoe photos" [backed by experiential/case_study]
  "Cohort analysis reveals true retention" [backed by empirical/methodology, strength: strong]

Each layer summarizes the layer below. Drill down for detail + evidence. Drill up for context.

The note format (v4.0)

---
tags:
  - type/claim/atom
  - priority/core
  - certainty/argued
  - stance/endorsed
  - domain/methodology
  - role/argument
  - source/ries-lean-startup
  - backing/empirical
  - strength/strong
kind: claim
layer: 0
proposition: "validated learning → measures progress → better than vanity metrics"
source_ref: "Chapter 7: Measure"
published: 2011
extracted_by: claude-haiku-4.5
prompt_version: v4.0
backing:
  - category: empirical
    subtype: case_study
    ref: "IMVU pivot story"
    snippet: "We had been spending our time improving a product that nobody wanted"
    strength: strong
    warrant: "The IMVU case demonstrates that traditional metrics masked the fundamental product-market fit failure"
passages:
  - chunk: "chunk_04.txt"
    lines: [42, 48]
    snippet: "We had been spending our time improving a product that nobody wanted"
confidence: exact
---
 
Validated learning is empirically demonstrated progress through experiments
that test specific hypotheses about the business.
 
Parent: [[Innovation accounting replaces vanity metrics]]

What each field does

Field	Purpose
`proposition`	The claim in canonical form (subject → relationship → object)
`source_ref`	Where in the source (chapter, section, control ID)
`backing`	The evidence: what type, how strong, why it supports the claim
`passages`	Pointers to exact source text in chunk files (for fact-checking)
`confidence`	How directly the source text states the claim (exact / synthesized / inferred)

The 9 backing categories

Every piece of evidence fits one of 9 universal categories:

Category	What it means	Examples
textual	Direct citation from authoritative text	Quranic verse, statute, primary source
transmitted	Report through chain of people	Hadith, witness testimony, reported data
consensus	Collective expert agreement	Scholarly ijma, scientific consensus, legal precedent
analogical	Extension from known to unknown	Qiyas, legal analogy, comparative study
empirical	Direct observation/measurement	Experiment, RCT, statistic
rational	Logical deduction/induction	Proof, syllogism, cost-benefit analysis
experiential	First-hand lived experience	Case study, anecdote
authority	Recognized expert statement	Scholar opinion, expert testimony
silence	Absence of evidence IS evidence	No text → ijtihad permitted

These work across any domain — Islamic jurisprudence, cybersecurity, academic research, business, philosophy, law.

Tag dimensions

Tag	Question it answers
`priority/`	How central is this to the argument?
`certainty/`	How solid is the evidence?
`stance/`	Does the author endorse, criticize, or neutrally present this?
`domain/`	What field is this about?
`role/`	Is this a fact, argument, definition, requirement, rebuttal, or methodology?
`source/`	Which source does this come from?
`backing/`	What type of evidence backs this? (textual, transmitted, etc.)
`strength/`	How strong is the evidence? (definitive, strong, moderate, weak)

Cross-source bridges

When two sources discuss the same concept under different names, a bridge entity unifies them:

Lean Startup: "Vanity Metrics"  ←→  Mom Test: "Compliments"
                     ↓
          Bridge: "False Signals"
          (aliases both names, backlinks from both sources)

The bridge page shows both perspectives and every claim from both sources. One read = complete cross-source answer.

Source verification

Run /dist:verify to fact-check claims against source text:

Claim: "MFA for remote access"
  → passages: chunk_01.txt, lines 117-118
  → open chunk_01.txt → read line 117
  → "Multi-factor authentication for remote access.2-2-3-2"
  → VERIFIED — exact match

Chunks are always stored locally, so this always works on your machine. For published brains, verification depends on whether the publisher included chunks (copyrighted sources typically don’t).

Entity pages are question-answering hubs. Their backlinks ARE the answer to “what does this brain know about X?”

/dist:lookup Customer Validation
  → find entity: "Customer Validation"
  → read its Referenced-by section
  → Lean Startup: test with MVP before building
  → Mom Test: look for commitment, not compliments
  → Zero to One: find a secret others have missed
  = 1 entity page, multi-source answer with citations

Deep research

For complex questions, use /dist:research:

/dist:research how do the sources disagree about debt?

The research agent iterates through multiple passes:

Scope the question → break into sub-questions
Search claims → check backing quality
Follow backlinks → entity → claims → entity → claims (3+ hops)
Walk the pyramid → root to atom for context
Cross-reference sources → check bridges for agreement/conflict
Advanced methods → warrant mining, analogical transfer, rebuttal reconstruction
Source verification → read chunks to confirm claims (Method K)
Report with evidence, confidence rating, and gaps

See Deep research agent for the full methodology.

Distillary

Explorer

How It Works