
Maturity
Field Value Status Research prototype Maturity Low Field-ready No — not clinical, not field-certified Last reviewed 2026-08-07
DarkSpace x Gray Swan integration scaffold. Research only — not field-certified.
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)
Verify:
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
Offline evaluation
Single-turn JSONL:
Multi-turn labeled JSON export:
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).1 like · 12 views · Aug 11, 2026
Developed an offline-first HTTP classifier for agentic transcripts.