Kavara Kirk

Kavara Kirk is the first algorithm in the Ulysses family — an algorithmic primitive for detecting structural change in raw, unlabeled data streams. Think of it as a Kalman filter for the non-Gaussian, non-stationary world. It maintains a running structural state, updates that state as each new observation arrives, and returns a scalar score measuring how much the current observation diverges from the accumulated structure. No labels. No training. No GPU.

The engine runs on the CPU you already have. Order-book snapshots score in under a millisecond on modern Intel and AMD server-class hardware. State is small enough (kilobytes) to embed in confidential-compute enclaves, so customer data never crosses the trust boundary in cleartext.

Available via MCP: score single or batched observations against a registered model, retrieve engine attestation (cryptographic proof of the exact binary running), browse capabilities and case studies, inspect usage and inference-unit balance.

The finance domain has been characterized end-to-end — a bit-exact reproduction of a published FY24 US-equities result runs in seconds against our production sealed engine. Other domains (industrial telemetry, network traffic, biosignal streams) are structurally compatible but require characterization first.

Output is a scalar structural-change score per observation, with the engine attestation attached. No black-box embeddings, no opaque logits — a single number you can reason about, backtest, and audit.

Prepaid inference units, metered per call. First checkout is $500 → 50,000 IU. Pricing is public via the kirk_pricing tool. The sealed engine ships as Intel TDX, AWS Nitro, and AMD SEV-compatible; every response carries a cryptographic attestation of the running binary.