Development of Mini X Templar Service for Gray Swan by Kyle BeanDevelopment of Mini X Templar Service for Gray Swan by Kyle Bean

Development of Mini X Templar Service for Gray Swan

Kyle Bean

Kyle Bean

Maturity

Field Value Status Arena entry Maturity Low Field-ready No — not clinical, not field-certified Last reviewed 2026-08-07
Gray Swan Safeguards Arena blue-team entry. Low maturity is not a flaw; unverified absolute claims are.

Mini X Templar (MxT)

Separate product and repo from DARKSPACE Safeguards: https://github.com/FratresMedAI/darkspace-gray-swan-blue (adapter-only; source to sync safeguards_adapter/ from when you cut releases). This repository is Mini X Templar (MxT) alone: FastAPI service with MxT lift, optional corroboration rescue, and entropy heuristics on top of the vendored rules tree in safeguards_adapter/.
Gray Swan contract: POST /v1/classify and POST /classify, body {"conversation":[{"role":"...","content":"..."}]}, response violation + confidence.
Health: GET /health on port 8080.
No LLM / no outbound network on the hot path.

Docker

Repo root must be this folder (the one that contains Dockerfile, mini_templar/, safeguards_adapter/).
Build + run locally (PowerShell):

Build + push to Docker Hub (after docker login):

Default Hub image: occisorleonum/mini-x-templar-gray-swan:grayswan-2026-04-24 — use that URL in Gray Swan (port 8080).
Manual one-liner:

Optional env: SAFEGUARDS_BLOCK_THRESHOLD (default 5.5), MINI_CORROBORATION_RESCUE=1, and other MINI_* vars (see mini_templar/core.py).

Tests

From repo root:

Relationship to Darkspace

The safeguards_adapter/ tree here is a vendored copy of the pattern engine used by DARKSPACE. When you cut a release, sync that folder from the Darkspace repo (or cherry-pick commits) so both stay aligned.
Like this project

Posted Aug 11, 2026

Built Mini X Templar (MxT), a FastAPI safeguard service for the Gray Swan Safeguards Arena. It classifies conversations for violations and confidence through a local rules-based path, with no LLM or outbound-network dependency. 1st Place, $15,000, judged by UK AISI.