cambium-ai · import cambiumRoutes work across deterministic, statistical, and generative methods, and learns from measured outcomes.
Shared platform component. Domain applications are documented separately.
Cambium provides reusable intelligence across business domains. It minimizes necessary model inputs, records cost, scores predictions against observed outcomes, and updates trust. Each application supplies its targets, evidence and acceptance criteria. Financial forecasts are one application of these contracts.
Tries less expensive methods before a generative call, subject to confidence and budget rules.
Reduces model inputs and records measured usage without tying the caller to one provider.
Scores observed outcomes and updates strategy and attribute weights through reusable contracts.
Candidate identities, time-aware mean and Brier losses, paired comparisons, and group-balanced blend arithmetic. Automatic promotion remains planned.
Keeps the strength and provenance of a claim explicit, including what evidence is missing.
Tracks commitment from explicit events. Its regime interpretation remains provisional.
Cambium supplies decision and learning primitives. A domain defines its question, inputs, target and acceptable evidence. Cambium does not hardcode stock returns, legal outcomes or learning objectives.
The resolution ladder tries deterministic logic, then statistical methods, then generative models. Distillation reduces necessary model inputs; the cost ledger records usage. A confident response still has to satisfy the caller's permissions and acceptance criteria.
The existing adaptive loop records predictions, observes outcomes, scores results and updates strategy and attribute weights. Applications provide their own persistence adapters and use the resulting trust in their decisions. A higher score is evidence about past behavior, not a promise about a future answer.
The release candidate adds reproducible candidate identities, prediction timing contracts, mean squared error and Brier scoring, matched comparisons, and group-balanced weighted means and medians. These are generic Cambium primitives. They do not own a database, user interface or financial target.
The scientist role and access checks belong to Canopy and the serving application. Timber supplies the data and plugin substrate. The first scientist workbench is an Investments application of these shared capabilities. It is a pre-production application, not a claim that every domain has a deployed scientist dashboard.
Each domain must choose scoring rules appropriate to its outcome and preserve what was knowable when a prediction was issued. Market-specific overlap handling, walk-forward evaluation and return targets live in the financial evaluation packages. They are described in Investments.
The candidate also provides scheduled, matched model selection and historical residual grouping. A challenger needs sustained evidence on consecutive reviews; sparse or incomplete evidence retains the incumbent. Applications own rollout and must validate the statistical assumptions for their targets. The first application runs in an isolated forward shadow ledger, outside production trust. New model adapters join the same contracts without another learning engine.
Timber explains models, persistence and service discovery. Investments explains the financial models, market evidence and domain-specific workbench built on those foundations.