Methodology Disclosure

This book is built from a registry that combines AI-assisted extraction with curator review. Honesty about where each operates is a condition of the registry's claim to be trustworthy. This section is the canonical statement.

Where the AI is allowed to act unattended

The AI-assisted pipeline can:

• Discover candidates — run literature watchlists, propose new citations matching curator-set criteria, fetch CrossRef + multi-model metadata.
• Verify metadata — confirm DOI accuracy, author lists, journal/volume/issue/pages, publication year. Routes through a multi-vendor consensus (OpenAI + Anthropic + Google) to flag disagreement.
• Synthesize priors — once an EffectSize row exists in the registry, the synthesis engine pools contributing rows under a (from_construct, predicate, to_construct) tuple using random-effects meta-analysis (DerSimonian-Laird τ²) with quality-weighting (A=1.0, B=0.7, C=0.4, D=0.1). The synthesis is deterministic from the contributing rows.
• File research tasks — the gap detector identifies thin-prior tuples, unverified citations, and stale watchlists; tasks are filed for curator attention.
• Propose but never promote effect sizes. Agent runners (and the Consensus REST adapter) write proposals to a gitignored directory; the curator promotes via the promote-effect-size CLI.

Where the curator (and only the curator) acts

The AI-assisted pipeline cannot:

• Write an EffectSize row to the store. Only the promote-effect-size CLI (PRN-030), invoked by the curator, writes effect sizes. This is locked under docs/specification/loop/curator-policy.md Decision D4.
• Set extraction_status: "verified" on a citation without CrossRef or multi-model concurrence. Unattended watchlists write needs_review and wait.
• Mark a citation as a survey_candidate. The Decision D2 hold prevents verified rows from becoming load-bearing until the curator deliberately approves. Phase B unlocks per-monitor auto-promotion on trusted watchlists only — and even then never bypasses novelty verification.
• Hard-delete any citation row. Rejected citations carry extraction_status: "rejected" plus a rejection_reason enum and a curator_review provenance stamp. The audit trail survives.
• Synthesize a CanonicalPrior without contributing EffectSize rows. The engine throws on empty input.

How the disagreement queue works

Two sources can disagree on the same field:

• Extractor vs verifier disagreement — flagged on the Evidence or EffectSize row; routes to the curator's daily digest under "Disputed claims."
• Multi-vendor disagreement — the multi-model citation validator records each vendor's answer; verdict: "disagreement" routes the row to needs_review regardless of CrossRef status.
• Replication failure — when a published replication contradicts an earlier finding, both rows persist; the synthesized CanonicalPrior for the tuple reflects the combined evidence and the replication_status flag downgrades accordingly.

The curator resolves disagreements deliberately. The registry never silently picks a winner.

Inter-rater reliability

Where the AI extracts a structured field that the curator subsequently reviews, the extractor's hit rate is logged. Reliability targets are documented per extraction type in docs/research/methodology.md. Below-target categories are flagged for either improved prompt engineering, additional verification passes, or routing to deep research rather than AI extraction.

Snapshot and reproducibility

Every chapter in this book carries a frozen_at timestamp and a registry snapshot id. The build pipeline records the exact JSON content of every row referenced. A reader who wants to audit a claim can:

1. Read the cited source.
2. Fetch the registry row at the snapshot id (the build artifact preserves the JSON).
3. Reconcile the cited source with the row.

Edits to the registry after a snapshot do not retroactively modify the book. New editions render new snapshots.

The novelty-verification commitment

Principia's most important methodological commitment is the one that's hardest to maintain at scale: every load-bearing claim is novelty-verified before it ships. AI-assisted extraction can hallucinate plausibly-sourced numbers; the only defense is a second pass that confirms the number actually appears in the cited source. The PRN-032 novelty-verification queue enforces this for synthesized priors and for any EffectSize row that the curator promotes via the standard CLI.

A reader who finds a claim in this book that does not appear in the cited source should report it. The reporting URL is on the colophon page.

---

Canonical source: `docs/research/methodology.md`. Curator policy locked at `docs/specification/loop/curator-policy.md` (2026-05-20). This disclosure is mirrored to PA-site's `docs/magazine/MASTHEAD.md` for the editorial surface.