Kirk — Unsupervised Structural Change Detection

kirk-mcp — Kavara sealed inference engine, attested with engine_sha on every response. Kirk is an explicit energy-based inference primitive. A bounded normalization property lets the energy landscape be evaluated without the O(2^n) partition-function wall that pushed the field away from explicit energy methods in the 1980s. The embedding is the energy function. Properties: no collapse by construction, online learning without pretraining, entropy differential as native surprise/anomaly signal, and composability across heterogeneous inputs. Authenticated access is either via a personal connector URL (https://kirk-mcp.kavara.ai/mcp/k_...) or CF Access service-token headers on the bare /mcp path. USAGE POLICY: - ≤200 books: use kirk_score_book or kirk_score_book_batch interactively. - >200 books OR loops: call kirk_bulk_howto — returns a Python client that scores at zero LLM token cost. Never loop the metered scoring tools from an LLM; both are capped by design. - Balance/checkout