Evidence foundation
Inputs retain origin, timing, version and transformation context so downstream work can be traced back to its source state.
The platform
Built to turn evidence into governed intelligence.
A shared operating system for specialist agents. Each domain remains independent while provenance, uncertainty, review and validation follow a common standard.
Four shared layers
The parts that should be common stay common.
Agents specialize in their own subject. The platform provides the disciplined path around them so new capabilities do not require a new safety model every time.
Domain agents
Focused agents transform approved inputs into probabilistic outputs with explicit horizons and uncertainty.
Decision review
Data quality, event meaning, counter-evidence, risk and executability are challenged before action is considered.
Controlled evidence
Backtests and paper environments remain distinct from any future live mode, with results evaluated rather than advertised.
A deliberate flow
Every stage has one job.
Relevant state
Collect what the domain needs, with freshness and quality boundaries.
Probability first
Forecast before market information can contaminate the underlying view.
Edge after costs
Only then compare the forecast with a precisely defined opportunity and its risks.
Evidence over stories
Preserve the decision and evaluate what happened without rewriting history.
Confidence is not a position size.
A strong forecast can make a decision eligible. It does not bypass liquidity, cost, concentration or portfolio limits. Conviction and sizing remain separate questions.
Agent-to-agent intelligence
Useful outputs can travel. Hidden assumptions cannot.
Weather.Cal is intended to become an upstream intelligence producer for agents such as Power.Cal. Shared outputs are treated as explicit, versioned data products — not as an informal copy from another agent's private runtime.
- A
Producer responsibility
The originating agent owns what the data means, how it was calibrated and where uncertainty remains.
- B
Consumer responsibility
The receiving agent binds the exact input version and adds uncertainty from its own domain.
- C
Platform responsibility
The common layer preserves lineage, policy, isolation and a reproducible decision path.
Designed to scale in stages
Prove the system before expanding the system.
The architecture is prepared for more agents, regions and compute capacity, while actual expansion remains gated by coverage, reliability, calibration and usable evidence.