Every research note on this site passes through a staged pipeline that closes the path from scouted research to a published claim: an adversarial verification pass that checks every cited paper against its source, a pre-commit gate that keeps a complete local mirror of every cited arXiv PDF, and a founder review of everything — draft to citations — before publish. The deliberate exception is the build notes — first-party write-ups of features shipped on this site. They cite no dossier; their claims are backed by the measured numbers they publish and the code running behind this page, and they carry the same founder sign-off before publish.
A research agent scans the tag's literature — arXiv, standards bodies, code repos — and surfaces 4–6 scouted sources per topic: papers, specs, and official docs, each with an exact title, ID, and a one-paragraph relevance case. No claim enters the dossier without a resolvable source.
Every scouted source is fetched and read. Key claims, numbers, and benchmark results are extracted with exact supporting quotes. The dossier is the write-side constraint — a post may cite only what the dossier already contains.
The draft is written from the dossier, not from memory. Every factual claim carries an inline citation to a specific dossier source — most of them arXiv papers. The draft targets 2,000–2,800 words, claim-form H2s, and an honest-limits section that names what the cited work does not cover.
A second pass, adversarial to the first: every citation is re-checked against its source. Every arXiv ID is resolved. Every number is traced back to the paper section it came from. A claim its source can't support as written is cut, or reduced to exactly what the source does support — never left standing on a hedge.
Nothing publishes on autopilot. The founder reads every post end to end — the draft, each claim, and every citation — and signs off before it enters the publish queue. Human-in-the-loop is the standing rule, not the exception path.
The pipeline's anti-hallucination gate has two automated layers, and both must pass before a post ships; the founder's sign-off (stage 05) sits above both.
Layer one — adversarial verify. After the draft is complete, a second agent pass re-checks every citation against its source. Every arXiv ID is resolved against arxiv.org/abs/<id>. Every number is traced back to the paper section it came from. A claim its source can't support as written is cut, or reduced to exactly what the source does support. The post then passes through validate-post.mjs, a deterministic build-safety gate that checks frontmatter, tag enums, MDX hazards, dead links, forward-dated links, and word-band compliance.
Layer two — pre-commit PDF mirror. A commit that stages any post triggers the repo's .husky/pre-commit hook, which rescans the whole collection and mirrors every cited paper into a local PDF cache, keyed by arXiv ID. The gate is idempotent: papers already in the cache are skipped, so its steady state is a fast scan with zero downloads. Only newly cited papers are fetched from arxiv.org/pdf/<id> — verified to actually be PDFs, so a rate-limit page is never cached as a paper — and any cited paper that cannot be fetched, whether a dead ID or exhausted retries, blocks the commit. It is the last automated gate: a paper that cannot be fetched cannot be cited.
Across the corpus, 180 out of 180 arXiv IDs have been resolved and verified. Zero hallucinated citations.
The pipeline guarantees that every post on this site is sourced, verified, and auditable. That corpus is now live and searchable — with the same engineering discipline behind the query path. Open the Search all Field Notes bar at the bottom of any page (or press CTRL+K) to try it.
Ask the field notes — a hybrid retrieval (BM25 + dense, RRF fusion) search over all 73 live posts, running on Cloudflare Workers AI and Vectorize. Type a question and get a streamed, citation-grounded answer with links to the source posts.
The same anti-hallucination discipline applies to the search path: BM25 catches exact-token queries (arXiv IDs, error strings) that dense embeddings miss; dense embeddings catch paraphrases that keyword search misses; RRF fusion weights both. Every result links back to a post that already passed the verification gate.
The first call is a free 30 minutes. You'll leave with the first cut of the build/no-build path, the riskiest assumption, and the eval we'd use to test it.