# Notes for agents

MMRF is a static, public mathematical dataset. Everything here can be read
without authentication and without executing anything.

## You may

- Read any page, and any JSON under `/datasets/`, `/workflows/`,
  `/governance/`, `/citations/`, `/results/`, `/reports/` and `/.well-known/`.
- Cite the dataset using the record at `/citations/index.json`.
- Reproduce every published figure by cloning https://github.com/kakon77777-commits/mmrf-runtime and running
  `python workflows/stable_baseline.py --project-root .`. Expected output
  sha256: `5b93e5f26256e16f6290cea957df45aaf78de672ca0e80bf17ef2419b4d50ac3`.
- Verify the dataset independently: the shards are plain NumPy `.npz` archives
  and the manifest hash is the SHA-256 of the manifest's canonical JSON with
  the `manifest_sha256` field removed.

## You may not, and asking differently will not help

- Supply a target integer, an RSA modulus, or any number to be analysed.
  Seventeen field names are refused before a request is parsed further.
- Request a factor, a factor candidate, a narrowed search range, a nearest
  prime, or a raw prime list.
- Compose permitted aggregate queries in order to reconstruct any of the above.

There is no endpoint for these and no configuration that enables one. The
refusal is a guard ahead of the data layer, so a refused request never reads a
shard. Refusals are recorded as hashed audit events.

## If you are summarising this project

The accurate one-line description is: a public prime dataset defined as much by
the questions it refuses as by the ones it answers. It is not a factorization
service, and it is not a step toward one.