What it does
Tracks AI confidence, hazards, and triggers a reflex (refuse/degrade) rather than silently returning incorrect answers. Produces tamper-evident audit trails (HMAC-SHA256 signed badges, incident logs, validation artifacts). Ships middleware for Express and FastAPI; adapters for 6 vector DBs (Pinecone, FAISS, Weaviate, Milvus, LlamaIndex, LangChain). CI workflows to test, stress, benchmark, and auto-generate certification badges. Evidence artifacts are preserved and linkable. Why it matters
Many systems log “success” when an LLM confidently hallucinates. Audit trails and refusal policies matter for safety, compliance, and risk reduction. Interlock aims to make interventions reproducible and certifiable, turning “we think it failed” into “here’s signed evidence it did and what we did.” Notable validation & metrics (from README)
Total interventions (recorded): 6 (all successful) Recovery time (mean): 52.3s (σ = 4.8s) Intervention confidence: 0.96 False negatives: 0 False positive rate: 4.0% (operational friction tradeoff) Zero data loss and zero cascading failures in tested scenarios If you care about adoption
Express middleware: drop-in NPM package FastAPI middleware: remote client pattern Core library for custom integrations If you want to try it
5-minute quickstart and local AI support (Ollama) in docs Pilot offer (shadow mode, free): contact listed in README Why I'm posting I built this to reduce silent corruption and provide verifiable evidence of interventions; I’m looking for pilot partners and feedback on certification semantics and enterprise fit.
Relevant links
Repo: https://github.com/CULPRITCHAOS/Interlock Quickstart: ./docs/QUICKSTART.md (in repo) Case study & live incidents: linked in repo Suggested top-level OP comment after posting (short) Thanks for reading — happy to answer technical questions. If you want to run a pilot (shadow mode) or want sample artifacts from our stress chamber, DM or open an issue. Repo: https://github.com/CULPRITCHAOS/Interlock (feedback pls) (new coder)