council-safe — in-house jail detector, measured

An in-house CSOAI jail-escape detector, measured on the same GSPC board as every other model and published, not hidden. Detection accuracy on the frozen jail bank is reported exactly as measured (0.4789) — the number is what it is, and the swarm-candidates 0.0 row stays on the page.

A name is not a measurement: six public models on the Hub are named "refusing" and do not refuse. That is why this card leads with the signed cards rather than with the name.

No weights are published in this repository. This is a measurement record: the signed cards below are the artefact. Every hash resolves to a card you can fetch and verify yourself, with no account.

Measured axes — Ed25519-signed, re-verifiable

axis	accuracy	signed card (sha256)	frozen bank
jail-escape-detection	0.4789	cff1aaf0…36535f4	csoai/gspc-jail-goldbank
swag-30	1.0	5756982c…9615bcc	csoai/gspc-jail-goldbank
swarm-candidates	0.0	4eb8ffdb…8c540c9	csoai/gspc-swarm

Each card was fetched and confirmed 200 at 2026-09-04T04:39:12Z. To check one by hand, follow https://councilof.ai/signed/HOW-TO-VERIFY.md, or paste it into https://councilof.ai/gspc-verify — free, no account. A card is evidence of what bytes scored on a frozen bank at a time. It is not an approval, a rating, or a safety guarantee.

Citation

@misc{csoai_model_council_safe,
  title        = {council-safe — in-house jail detector, measured},
  author       = {{CSOAI Ltd}},
  year         = {2026},
  doi_status   = {Zenodo DOI 10.5281/zenodo.21991104: UNAVAILABLE since 29 Sep 2026 (Zenodo blocked the CSOAI account; appeal pending; status https://councilof.ai/interop/zenodo-status.json). Cite the served methodology: https://councilof.ai/signed/HOW-TO-VERIFY.md},
  howpublished = {Hugging Face Hub, \url{https://huggingface.co/csoai/council-safe}},
  note         = {Live board: https://councilof.ai/api/gspc. Measurement, not certification.}
}

Reuse and provenance

This repository contains a model record and supporting material; it does not distribute model weights.

Use the file's observation date, scope and licence when reusing it. Repository updates and downloads are not new measurements.

• Current GSPC board and methodology.
• Verify supported signed records and read corrections.
• Earlier page notes and dated integration announcements.

Measurement is not certification. A repository, signed record or passing software test does not by itself establish current board admission, product acceptance or benchmark leadership.