Skip to content

Six invariants

SWAT+ hydrology is the case study, not the point. The exportable contribution is a small set of invariants for agent-governed scientific pipelines — the properties that make any AI-operated computational science workflow trustworthy and auditable. swatplus-builder is one concrete implementation of them.

The invariants

  1. Typed contract before execution. Work begins from an explicit, machine-checkable contract (task, scope, requested claim tier, preconditions) — not from free-form intent. Policy is validated before compute is spent.

  2. Claims, not results — tiered, gated, artifact-backed. The unit of output is a claim with a tier, not a bare number. Every claim is backed by gate results and an artifact pointer. The headline metric is never the authority.

  3. Locked baseline → candidate → independent verification. Improvement is measured against a sealed baseline, and the reported value comes from an independent rerun of the promoted artifact — never from the optimizer's own trajectory.

  4. Fresh-evidence enforcement. Outputs must be provably produced by this run. Stale artifacts are unscoreable; a metric computed on yesterday's outputs cannot promote a claim today.

  5. Provenance-or-degrade. Every fallback (synthetic weather, fallback soils, auto-selected outlet) lowers the achievable claim ceiling and is recorded. You can still run with degraded inputs — you just cannot claim as much.

  6. Machine-readable refusal. When the evidence does not support a claim, the system refuses with a typed reason and a next action, in the same structured form as a success — so refusals are auditable and programmable.

Why frame it this way

These six properties are domain-neutral. A flood-frequency tool, a remote-sensing retrieval pipeline, or a parameter-estimation service could each implement the same invariants with entirely different gates. That is the generality claim: the governance core — contracts, claims, tiers, gate protocol, evidence I/O, refusal types — is separable from the hydrology, and the hydrology is just the first set of gate implementations registered against it.

Status of the generality claim

Demonstrating the invariants in a second, non-hydrology domain (a small toy reference) is planned work, not a completed result. See the project roadmap on GitHub and Honest status. This page states the framework; it does not claim the second-domain demonstration is already done.