Maritime risk intelligence
Risk scores you can defend to a regulator.
Lamna scores vessels, routes, and compliance posture with a deterministic, versioned, property tested core. Agents explain every number and cite their evidence. Re-run any assessment and get the identical result, proven by its input hash.
Deterministic core
Vessel, route, and compliance scores come from versioned, unit tested, pure functions. Worsening any input never lowers risk; a property test suite enforces it. Weights are configuration with a version stamped on every assessment.
Agents that explain, not decide
Claude reads the same factor contributions you do and writes the analyst narrative. It can propose an adjustment inside a hard bound, with a stated reason, logged. It cannot override the score.
Reproducible by construction
Every assessment records a hash of every input that fed it. Same inputs, same output, byte for byte. Every model call is logged with prompt, response, latency, and content hash.
Architecture
Every feed enters behind an interface. Scoring math never touches IO. The agent layer sees the same evidence the API returns to you.
Built for buyers who carry the risk
Marine insurers, P&I clubs, and compliance teams do not need another opaque ML score. They need a number with a paper trail: which factors moved it, which weights version scored it, which model explained it, and proof the same inputs give the same answer tomorrow. That is the product.
- Scoring functions
- pure, versioned, tested
- Agent adjustments
- bounded, logged, cited
- Assessment re-runs
- byte identical
- Model call audit
- prompt, response, hash