The standard answer to "make the AI know my sources" is RAG: embed the pile, retrieve at query time, hope the top-k was right. Chancery is the other bet: read once, judge explicitly, distill into small verified notes — and make those the thing your project retrieves from.

You feed it URLs as you research. Each source is quarantined and judged against your written bar — by your AI agent, with the judgment kept on file — and, if it clears, distilled into concept notes with hashed evidence, typed links, and flashcards. The corpus is plain markdown: you read it in Obsidian, agents load it through kb query and kb context, and CI runs kb verify so it can't silently rot. The result is locally verified research: claims you can cite, because the tool can prove what the evidence said when you cited it.

The loop, once

$ kb ingest https://example.com/attention-paper      → staging/attention-paper.md
$ kb assess staging/attention-paper.md               → prints a judgment form
    (your agent fills it in → verdict.json)
$ kb assess … --verdict verdict.json                 → PROMOTE — clears the bar
$ kb promote attention-paper                         → prints a drafting form
    (your agent writes the note → draft.json)
$ kb promote … --draft draft.json --apply            → concepts/attention.md ✓
$ kb verify                                          → PASS — 12 files, 0 errors

That's the whole tool. kb never calls a model: everything checkable by rule, it checks; everything needing judgment becomes a self-contained form that any supplier — your coding agent, a panel, you — fills in, and the CLI validates and files. Every note carries the paperwork that admitted it. Retrieval then happens over what you've vetted: kb query answers from the curated layer and rejects any answer citing a concept it didn't retrieve.

What keeps it trustworthy

• Admission is judged, and kept. Sources are scored against a rubric with knockout rules — a fatal flaw can't be averaged away. Borderline calls go to a queue for you, reasons on file. Nothing enters unjudged; nothing is auto-deleted.
• Citations are falsifiable. Every source is content-hashed into an append-only evidence store. kb revalidate reports which claims' evidence drifted, not just which links died; kb support checks claim-by-claim against the exact snapshot judged.
• Both audiences are first-class. Humans get Obsidian-compatible markdown and spaced-repetition cards; agents get grounded query, deterministic context bundles (kb context), and an MCP server. Same corpus, same gate.

Proof, not promises

The corpus in this repo was generated by the tool itself: 275 research sources through the judged loop → 212 concepts, 212 decks, 1,112 cards — with a 44% admission rate, every rejection reasoned, weak sources upgraded to primary ones by a budgeted search pass, and the queue holding the honest borderline calls. Method, numbers, and confounds: the parallel run. We also ran it on real work first — routing research where kb support caught our own draft overclaiming its source (launch note).

Start

npm install -g @amiable-dev/chancery
mkdir my-kb && cd my-kb && kb init

Then the first-corpus tutorial: bring 3–5 sources you're actually researching; in ~20 minutes you'll have a verified corpus and kb query answering from it — your agent judging, via a copy-paste prompt. kb holds no model credential, ever.

Why "Chancery"?

A medieval chancery authenticated documents under seal and kept the rolls — nothing entered the record without passing the office, and the record could prove it had. Same job here: judgment arrives from outside; the office makes it canon through examination, a sealed envelope, and an enrolment it can verify (the longer story).

What's in this repo

concepts/     212 atomic concept notes (typed links, facets, hashed sources)
flashcards/   1,112 spaced-repetition cards with stable IDs
staging/      quarantined inflow — including every source the gate declined, with reasons
.kb/          the engine: schemas, rubrics, policy, CLI, tests — plus the corpus's judgment records
eval/         the pre-registered method behind the generated corpus
.claude/ .github/ .windsurf/ .agents/ AGENTS.md   ← generated harness adapters

Status

Early, honest v0.x. The engine is tested (20 suites, both-polarity gate tests, CI-enforced) and in daily use; the schema may still change with versioned migrations. Scope and non-goals: docs/SCOPE.md; decisions: docs/adrs/; the record — launch note, why "Chancery", the parallel run — is history, deliberately out of your reading path.

If you point kb at a corpus of your own and something breaks or chafes — please open an issue. Knowing whether anyone else wants governed knowledge is, quite literally, this project's next milestone.

Related

• llm-council — multi-model deliberation (library + MCP + HTTP). A sibling project; kb can use it as one judgment supplier among several, never a hard dependency.
• Karpathy's LLM Wiki gist — the pattern's most-cited articulation. kb's architecture predates it independently (Feb 2026); the convergence is evidence for the shape. Where the pattern says the LLM should do the bookkeeping, kb says: the bookkeeping is exactly the part that must be deterministic.

License

MIT © 2026 Chris Joseph (amiable-dev)