v0.1.0 - Architecture Zero

A self-hosted AI assistant platform built around a measured trust layer. Run it
on your own infrastructure, point it at your own documents, and every answer is
grounded in your corpus with citations.

git clone https://github.com/architecture-zero/architecture-zero
cd architecture-zero
cp .env.example .env    # set JWT_SECRET_KEY - the backend refuses to boot on the placeholder
docker compose up -d --build

Then claim the deployment with the code the backend prints to its own logs, and
ask your first question.

What is in it

• Hybrid retrieval - vector plus BM25, with cross-encoder reranking. The
  reranker ships inside the image, so the first request does not block on a
  download and the instance works on a box with no outbound network.
• Tiered access control - retrieval is filtered by the caller's clearance,
  and the same tier resolution gates the answer layer and the file tools, so
  one surface cannot walk around another.
• An ingest-time injection gate - untrusted content is scanned at the
  single choke point every ingestion path goes through, and quarantined rather
  than trusted.
• A judged evaluation system - the instance measures its own answer
  quality, groundedness, honesty and retrieval recall on its own corpus, and
  the trust panel publishes those measured numbers rather than hand-entered
  ones.
• A reference web client - chat plus the admin surfaces an operator
  actually needs: provider settings, the knowledge base, users and roles,
  evaluation runs, the audit log.
• Operational basics that are usually an afterthought - a first-owner claim
  code so a reachable deployment cannot be taken by whoever finds it first,
  audit logging, backups with a heartbeat the health probe reads, and a
  /metrics endpoint with its own long-lived token.

What this tag means

Ninety commits over ten days, then seven rounds of adversarial review and a
live run from a fresh clone found nothing further.

That is what it means. It does not mean correct.

The Known limitations
section enumerates what this release does not do, rather than leaving it to be
discovered. The roadmap
names the work consciously not taken.

Two things are worth stating plainly for anyone evaluating this:

The last two defects found before this tag came from running it in a browser,
not from reading the code. One of them was a shipped default that only worked
on localhost - deploy anywhere real, browse to your own server, and every
question returned 403, while the config file said the relevant setting was never
consulted. Seven review rounds and roughly a hundred review agents did not see
it. A wrong port surfaced it in one second.

A green test suite is evidence about the suite until you read what its
assertions exercise. This repository twice reported a fully passing suite over
code that suite did not touch at all. The frontend tests added for this release
cover one rule thoroughly - a message is marked unstored exactly when no server
row exists, checked across every way a stream can end - because that rule is
where the serious defects were and getting it wrong deletes conversation
history. They were mutation-tested: reverting the fix makes exactly the right
one fail. The README says what they cover and what they do not.

License

Apache-2.0.