Skip to content

Domain support

Anarres separates software support from scientific validation. Format support means the software can read and validate the data. Predictive evidence means a measurement or decision was tested against later outcomes. One does not imply the other.

Conformance levels

  1. Protocol: the domain has a public measurement or experiment contract.
  2. Profile fixture: synthetic records exercise domain-shaped coordinates through schema validation and the native CLI.
  3. Native adapter: released evidence is projected into profiles with exact provenance and adapter conformance tests.
  4. Source runtime: the adapter can qualify a declared native source format through the unified source envelope and, when released, run a pinned starting-policy probe.
  5. Predictive evidence: a domain study evaluates a frozen measurement or decision against downstream outcomes under its native contract.

The first four levels concern interoperability. The fifth is scientific evidence and remains specific to the study, policy, target, and budget.

Native-adapter status does not imply source-runtime status. Most current adapters reproduce released evidence. They become raw environment runners only after declaring accepted source formats and passing qualification and probe conformance on the real domain runtime. An adapter may release source qualification before it releases policy execution. anarres inspect <source> reports both capabilities separately, and anarres adapter doctor <name> is the authoritative adapter-level check.

Current status

Mathematics

  • Interface: Native grouped-probe processor
  • Native sources: Completed raw-attempt streams
  • Validation: Four-scale environment-utility study

Formal verification

  • Interface: Installable Vero adapter
  • Native sources: Source qualification and model-free runtime preflight
  • Validation: Training comparisons across Dafny and Verus. Vero remains pre-RL.

Robotics

  • Interface: Installable CUPID adapter
  • Native sources: Can v1.4.1 dataset qualification
  • Validation: Prospective PushT and Can panels, plus retrospective multi-source studies. Square candidate outcomes remain unopened.
  • Interface: Installable legal adapter
  • Native sources: Pinned LAB task-set qualification
  • Validation: Prospective SylloGym panel and a separate LAB regression-risk study

Materials science

  • Interface: Installable materials adapter
  • Native sources: OMat24 atomistic-bank qualification
  • Validation: Supervised phonon studies and a scored MLIP prospective test

Molecular design

  • Interface: Installable molecular adapter
  • Native sources: Reaction-bank qualification and a pinned dry-run policy probe
  • Validation: Prospective RxnFlow environment-selection, antibody-acquisition, and protein-acquisition panels

Vulnerability exploitation

  • Interface: Installable CyberGym adapter
  • Native sources: Sanitized level 0--2 panel qualification
  • Validation: Policy-relative scaffold measurement. No environment comparison has been released.

Agentic finance

  • Interface: Installable finance adapter
  • Native sources: Pinned Business Bench process qualification
  • Validation: Retrospective source tests. No prospective environment outcome has been released.

The profile schema can transport additional namespaced coordinates without a Rust change. That extensibility is not evidence that an unlisted adapter works or that one domain's predictor transfers to another.

Release rule

A domain moves to native-adapter status only when its checked-in workflow:

  • reads a pinned native measurement artifact
  • verifies source, subject, contract, and artifact identity
  • preserves invalid, missing, and infrastructure outcomes
  • emits versioned coordinate definitions with units and estimators
  • passes schema and native CLI conformance tests
  • retains the originating scientific record as authoritative

Predictive status requires a separate outcome-bearing experiment. Synthetic fixtures never satisfy that requirement. Umbrella programs do not receive generic adapters. Their concrete studies expose domain-specific commands when a stable native producer is available.