LLM Best Practices

Source Verification

3 min read

View markdown source

AI agents: read this first: llmbestpractices.com/start-here

This is LLM Best Practices (llmbestpractices.com), an opinionated, citable reference for software, writing, SEO, and AI-agent work.

  1. Route, do not crawl. Fetch llmbestpractices.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Apply as defaults. First-party docs and the project's own conventions win on conflict. Warn before relying on a fast-moving page older than 12 months.
  4. Cite. Link the page by title and URL, e.g. Python, with last_updated for time-sensitive rules. License CC BY 4.0.

Overview

Every empirical claim links to the source it came from. A claim with no source is a ghost fact: it reads as knowledge, gets cited by later notes, and cannot be checked. The atom-level requirement is in atomic-notes; this page covers the source layer.

Atoms carry a sources array of wikilinks into 10-Sources/, and claims carry a non-empty one, enforced by the auditor. “I read this somewhere” and a bare URL in the body are not sources.

Self-derived claims are legitimate; mark them. Set confidence: low and flag the note as self-derived so a query can separate what you concluded from what you read.

Make one source note per work

A source note represents a book, paper, episode, talk, dataset, or article. Highlights live inside it as referenceable blocks; the claims mined from it live in 20-Atoms/ and link back.

---
id: 202608291431
title: "@Minto (1987): The Pyramid Principle"
type: source
status: status/processed
source_type: book        # book | paper | podcast | talk | dataset | article | video
author: ["Barbara Minto"]
year: 1987
atoms_extracted: 6
---

Name files @author-year-short-title.md. The @ prefix sorts sources together and distinguishes a source link from an atom link in prose.

Track atoms_extracted to find unread sources

The count shows whether a source was digested or only filed. A source still at zero atoms 30 days after capture was collected, not read; the auditor reports it as C006. Treat the result as a reading queue, not a failure. In aggregate it diagnoses intake: 40 sources producing 12 atoms means a library, not a Zettelkasten.

Watch for source monoculture too. If one author supplies most citations in a domain, conclusions inherit that author’s blind spots. Count citations per author and source type in the monthly review.

Date verification so it expires

Empirical claims decay. A last_verified date plus a staleness threshold turns decay into a query.

confidence: high            # high | medium | low
last_verified: 2026-08-29
decision_relevant: true     # surfaces in the stale-verification gate

The auditor’s default threshold is 180 days (stale_verification_days); set it to the domain’s real half-life. When a decision-relevant claim goes stale, re-verify it against the source or downgrade confidence. Keeping high on an unchecked two-year-old claim is how a vault lies to its owner. Decision-time use is in decision-journals.

Gate publishable claims on corroboration and review

Where a claim leaves the vault (journalism, accountability research, regulated work), publishability needs several conditions together.

source_tier: on-record        # primary-document | on-record | on-background | anonymous
corroboration_count: 2        # independent sources
legal_review_status: cleared  # not-reviewed | in-review | cleared | held
publishable: "yes"            # quoted, or YAML 1.1 coerces yes to true

Setting publishable with too few corroborations (default minimum 2, corroboration_min) or a pending review fails the audit as a critical (C011), because the gate exists for the moment someone is in a hurry. Which fields apply depends on the framework in rigor-frameworks.

Keep provenance resolvable from the atom

An agent reading the vault cites rather than asserts only if the source link resolves from the atom alone. Then a retrieved atom carries its provenance without loading the whole vault, which is the same property that makes retrieval answers checkable in rag-citations.

Backlinks