| name | book-thesis |
| description | Metabook reasoning layer for the book-* skill suite. Owns the intent
substrate: a thesis tree, paragraph back-pointers to thesis nodes,
per-paragraph entailment checks, and an EDN/Cozo consistency pass over
claim facts for transitive cross-chapter contradictions.
Use when a book needs to enforce three properties simultaneously:
factual accuracy, cross-chapter consistency, and preservation of the
core point across revisions. Composes with book-knowledge (facts),
russellian-style (prose), book-compose (orchestrator), book-review
(qualitative), book-qa (mechanical gate).
Do NOT use for chapter drafting (use book-compose), claim ingestion
(use book-knowledge), or sentence-grain prose discipline (use
russellian-style).
|
| license | MIT |
book-thesis
The metabook reasoning layer. Owns the intent substrate of a book — the thesis statement, the sub-arguments that decompose it, and the requirement that every paragraph in the manuscript trace back to some node in that tree. Sibling to book-knowledge (which owns the fact substrate).
What it owns
- The thesis tree: a YAML/RDF graph rooted at
:Thesis, with sub-arguments and required-evidence slots.
- Paragraph back-pointers: every paragraph in a draft carries a
supports: field naming a tree node.
- The entailment loop: a per-paragraph LLM-critic dispatch that asks "does this paragraph actually advance its declared supports node?"
- The consistency pass: a recursive CozoScript rule set (compiled from booklogic EDN) that derives transitive contradictions across chapter-level claim facts.
- The exemplar pack: synthetic (supports-node, claim, paragraph) triples injected into the drafting prompt.
What it does NOT own
- Claim ingestion or claim-ledger maintenance. That is
book-knowledge.
- Sentence-grain prose linting (hedges, passive voice, modifier budget). That is
russellian-style.
- Chapter orchestration, bundling, and book release. That is
book-compose.
- Qualitative editorial review by personas. That is
book-review.
- Mechanical D1–D8 defect gating. That is
book-qa. (book-thesis contributes D9–D12.)
Architecture (4 layers)
book-thesis sits as Layer 2–4 on top of book-knowledge's Layer 1:
┌─────────────────────────────────────────────────────────────┐
│ Layer 4 EDN/Cozo consistency pass │
│ rules/consistency.cozo over claim facts │
└────────────────────────┬────────────────────────────────────┘
┌────────────────────────▼────────────────────────────────────┐
│ Layer 3 Verifier-generator entailment loop │
│ per-paragraph LLM critic, grounded against the KG │
└────────────────────────┬────────────────────────────────────┘
┌────────────────────────▼────────────────────────────────────┐
│ Layer 2 Thesis spine │
│ YAML / RDF tree + paragraph back-pointers │
└────────────────────────┬────────────────────────────────────┘
┌────────────────────────▼────────────────────────────────────┐
│ Layer 1 Claim ledger + EDN/Cozo store (book-knowledge) │
└─────────────────────────────────────────────────────────────┘
Components
scripts/compile_thesis.py — read thesis/<book-id>.yaml, emit RDF triples into the book-knowledge graph.scripts/lint_supports.py — walk every paragraph in the manuscript; flag orphans (no supports:), broken supports (node doesn't exist), and chains that don't transitively reach :Thesis.scripts/dispatch_entailment.py — prepare per-paragraph payloads for an LLM entailment critic. Returns entailed | weakly-entailed | unrelated | contradicts.scripts/consistency_cozo.py — project the thesis spine + claim facts into book-knowledge's Cozo store, run the recursive rules/consistency.cozo program, emit derived contradictions (D9–D11) and write the qa/datalog-defects.json gate artifact.scripts/synthesize_exemplars.py — Symbolic→LLM exemplar pack generator. Given the thesis tree + claim slice for a chapter, writes a few-shot exemplar bundle used by the drafting agent.rules/consistency.cozo — the recursive CozoScript consistency program (compiled from booklogic EDN) covering transitive contradiction, orphan paragraphs, thesis-reachability.
Composes with
book-knowledge — fact substrate. book-thesis projects claim facts into book-knowledge's Cozo store for the consistency pass, and compile_thesis emits the thesis spine as triples.book-compose — the chapter contract gains a thesis_advances: field naming sub-arguments. compile_thesis.py runs at the start of build_book.book-qa — defines four new defect classes (D9 orphan-paragraph, D10 transitive-contradiction, D11 failed-entailment, D12 unadvanced-sub-argument). lint_supports.py and consistency_cozo.py feed those gates.book-review — orthogonal. book-review's personas do qualitative review; book-thesis does formal-structural check.russellian-style — orthogonal. russellian-style polices sentence-grain; book-thesis polices argument-level.
Usage
python scripts/compile_thesis.py <workspace> <book-id>
python scripts/lint_supports.py <workspace> <release-version>
python scripts/consistency_cozo.py <workspace>
python scripts/synthesize_exemplars.py <workspace> <chapter-id>
python scripts/dispatch_entailment.py <workspace> <release-version>
Tests
Fixture-based pytest cases under tests/. Run with:
python -m pytest tests/ -v
Coverage:
- compile_thesis: round-trip YAML → triples → SPARQL query
- lint_supports: known-orphan paragraph fixture; known-good paragraph fixture
- consistency_cozo: contradicting-claim fixture (A→B in ch-1, B→¬A in ch-2), frozen against the C0.3 goldens
- dispatch_entailment: payload-shape contract
- synthesize_exemplars: exemplar pack shape contract
