Preserving domain science¶

The Anarres CLI makes domain studies inspectable and interoperable. It does not replace their models, estimands, decision rules, or native evidence.

An environment profile is a typed projection with four responsibilities:

identify the environment, starting subject, source, and measurement contract
record the domain adapter and implementation revision that produced it
expose selected measurements with versioned definitions and units
link to native evidence and qualification records by URI and digest

The originating study remains authoritative. Reproducing a claim still requires its native protocol, records, and analysis. The CLI can show that two candidates were measured compatibly, but it cannot infer that a coordinate predicts utility unless a separate validated decision rule establishes that relationship.

Legal reasoning¶

The legal program contains two distinct scientific objects. The SylloGym study uses readiness, eligible-response availability, reward discrimination, token cost, truncation, reliability, missingness, and a coverage floor to construct a mixture. The LAB study uses task and workflow structure to estimate regression risk across held-out practice areas. Neither object is equivalent to grouped reward mixedness or effective coverage.

The legal adapter exposes the LAB result through anarres legal regressions as a read-only evaluation report. It recomputes the compact released summary and retains its validation split, support, uncertainty, and claim boundary. It does not convert the result into a candidate-environment profile.

An adapter can export selected measurements from either study, but it must use different measurement contracts and coordinate definitions. The native frozen decision, checkpoint record, regression endpoint, validation split, and uncertainty analysis remain linked artifacts rather than flattened profile attributes.

Robotics¶

The robotics program keeps target relevance, policy weakness, coverage, data-quality correction, reliability, missingness, exposure, cost, native outcomes, and guardrails separate. It also distinguishes a correct abstention from selector superiority and retains physical outcomes in their native units.

An adapter must preserve episode and task identity, policy and dataset revision, termination and truncation, and simulator or hardware validity. Candidate profiles may expose a compatible subset of measurements for comparison, while the native prospective protocol and outcome record remain authoritative.

Agentic finance¶

The finance program separates starting-policy qualification, candidate measurement, training outcomes, material-control outcomes, and retrospective cross-source analyses. A qualification profile may expose executable completion, tool-protocol reliability, mixed-scenario coverage, control violations, truncation, and realized cost under one pinned operations contract. Candidate ranking begins only after the native decision rule has been frozen and tested.

Infrastructure-invalid episodes remain distinct from failed finance work, and control violations remain visible even when executable completion improves. Published post-training comparisons can support the finance research program without becoming pre-training coordinates for the prospective operations study.

Molecular design¶

The molecular program measures fixed replay banks under one RxnFlow starting policy, synthesis environment, Q07820 oracle, and bank-size contract. Candidate profiles expose the native construction-view summaries for target reward, generated actions, oracle time, scaffold and ring-system coverage, route and reaction coverage, and distance in the frozen graph representation. The independent view qualifies the stability of the six-bank ordering and does not contribute members or get averaged into those coordinates.

Molecule-specific eligibility, conformer, and pinned-oracle atom-type failures remain explicit scientific zeroes in the native probe and are excluded from the eligible bank pool. Every other parser, receptor, toolchain, docking-process, and timeout failure remains an infrastructure failure. Profiles link the native qualification records and preserve target-training outcomes as a separate scientific object. A compatible six-bank panel therefore establishes measurement conformance, not predictive or allocation evidence.

Evolution¶

New domain functionality normally requires a new adapter or measurement convention, not a new CLI command. New coordinates can be added by publishing a definition and emitting them from the adapter. A changed estimator, unit, or scientific meaning receives a new convention revision. The CLI rejects mixed revisions in one candidate panel, allowing old and new studies to coexist.

The core schema changes only when the operational envelope itself changes, such as adding a new integrity primitive. Such changes require an RFC, migration notes, and conformance tests.

Current integration boundary¶

The grouped policy-probe workflow has a complete native path from qualified attempts to a content-addressed environment manifest. Cross-domain profiles have a validated interchange contract, Python construction helpers, native CLI conformance tests, and a real producer for the released Vero development measurement. That adapter preserves a valid zero-signal result, positive model cost, and the failed support gate as an unready profile.

The SylloGym release includes its complete outcome-blind component measurement and frozen mixture weights. Its study-local adapter projects the three compared mixtures without reading downstream outcomes. The PushT release still omits the complete per-candidate pre-training vectors and remains an evidence release rather than a profile producer. The LAB regression-risk record is a predictor evaluation over tasks, not a candidate environment measurement, and remains a separate scientific object.

The Business Bench finance adapter projects one completed Qwen3-8B starting-policy qualification. It preserves zero completion, zero mixed-group coverage, positive model cost, infrastructure failures, and the failed readiness gates. Comparable candidate profiles remain unavailable until a starting policy qualifies and the native candidate bank is measured.

The Vero formal-verification, molecular reaction-bank, SylloGym legal, Can robotics, MLIP materials, and Business Bench finance studies now have native profile producers at their respective evidence boundaries. Their profiles establish record and measurement conformance. Predictive or allocation evidence still requires the outcome protocol defined by each study. Additional producers should first remain beside their experiment and satisfy the adapter conformance workflow.