DarkSpace X RuleGuard

Offline-first HTTP classifier for agentic customer-support transcripts.
Combines deterministic rule scoring with a lightweight CPU semantic gate (sentence-transformers MiniLM). Designed for reproducibility: fixed dependencies, declarative Dockerfile, no network calls on the classification hot path.

Repository: github.com/FratresMedAI/DarkSpace-X-RuleGuard

Methodology (summary)

• Rule layer: High-precision patterns for prompt-injection framing, credential phishing, tool-argument abuse, and multi-turn orchestration smuggling. Pattern families were informed by published survey and taxonomy literature on jailbreak and indirect injection techniques (e.g. crescendo, translation smuggling, synthetic tool traces), cross-checked against benign enterprise phrasing to control false positives.
• Semantic layer: Embedding similarity for fuzzy alignment; runs fully on CPU in the default container.
• Mini-Templar wrapper: Optional “rescue” path for weak multi-signal cases using corroboration and entropy heuristics (see mini_templar/core.py).

Quick start (Docker)

docker build --platform linux/amd64 --provenance=false --sbom=false -t darkspace-ruleguard:local .
docker run --rm -p 8080:8080 darkspace-ruleguard:local

Verify:

curl -s http://127.0.0.1:8080/health
curl -s http://127.0.0.1:8080/ready
curl -s -X POST http://127.0.0.1:8080/v1/classify \
  -H 'Content-Type: application/json' \
  -d '{"conversation":[{"role":"user","content":"Please summarize Q3 earnings in two bullets."}]}'

HTTP API

See docs/API.md for request/response schema and all supported POST routes (compatible with common classifier probes: /v1/classify, /classify, /api/v1/classify, root POST /, and a catch-all POST).

Local verification

python -m venv .venv && source .venv/bin/activate  # or Windows equivalent
pip install -r requirements-dev.txt
# Skip extended neural-mirror smoke during pytest (not required for rule unit tests):
export DARKSPACE_SKIP_MIRROR_STARTUP=1   # Windows: set DARKSPACE_SKIP_MIRROR_STARTUP=1
export PYTHONPATH="$PWD"                  # Windows: set PYTHONPATH=%CD%
pytest safeguards_adapter/tests/ mini_templar/tests/ -q

Offline evaluation

Single-turn JSONL:

python scripts/eval_labeled_dataset.py tests/fixtures/survey_literature_curated.jsonl --format jsonl --judge no-llm-judge

Multi-turn labeled JSON export:

python scripts/eval_labeled_dataset.py path/to/labeled.json --format conversation --judge no-llm-judge

Security & privacy

See SECURITY.md. Classification is stateless with respect to remote services; SQLite audit hooks initialize locally for compatibility with legacy middleware—treat deployments as ephemeral or mount an empty volume for audit_log.db if persistence is undesired.

License

Apache-2.0 — see LICENSE.

Citation / audit

If you use this system in research or assurance reports, cite the repository URL and the commit SHA you reproduced.

For UK AISI, CISA-style, or independent review: full reproduction = Docker build from this tree + checksum of lockfiles (requirements-docker.txt as pinned in image) + documented ENV (see Dockerfile).