# Schema and specification alignment The `schema/cars-0.1.schema.json` file is an **experimental L0 subset** of the CARS 0.1 prose draft. It validates structural shape, mandatory declarations, URI syntax and a limited set of enumerations. It does **not** verify semantic dependency graphs, adapter resolvability, chemical representations, units, dependency licenses, sandbox security, execution success or scientific evidence. The sample `workflow.cwl` reference is intentionally non-executable and must be replaced before claiming CARS-Execute support. Before a normative release: resolve extensibility policy, semantic identifiers, registry governance, unit vocabulary, adapter contract, signed provenance and compatibility guarantees. ## 2026-10-20 validation-status erratum and migration The initial release ZIP already contained the mismatch: Specification §5.3 named four evidence assertions, while Schema used `not_assessed`, `under_review` and `reviewed`. The CLI delegated to that Schema. Its two original tests covered only the `not_assessed` example and a missing top-level field, so they could not detect this drift. No earlier design rationale for the three-value enum is present in this source bundle; review workflow and scientific evidence had been conflated. The corrected enum is `not_assessed`, `benchmarked`, `independently_reviewed`, `experimentally_supported`. Meanings and evidence requirements live in §5.3. The conformance suite directly parses that section's table and compares its values to Schema; it also checks the top-level MUST fields and all official example files, including the inline specification example. Compatibility impact: manifests with `under_review` or `reviewed` that passed the initial Schema now fail L0. `under_review` says only that a process is underway; `reviewed` identifies neither reviewer independence nor scientific scope. Neither has a lossless mapping, including to `not_assessed`. The CLI reports an error and migration guidance, without rewriting the input or selecting a replacement. Recipe owners must inspect the underlying evidence, explicitly select the appropriate assertion, cite the assessment, and retain their original declaration in revision history. If evidence cannot be established, do not auto-upgrade or silently reset the claim; resolve it with the owner before revalidation. Version any published recipe content change under the recipe's own `version` policy. The sole bundled official recipe and its inline copy use `not_assessed` and are unchanged. The ChemAIAtlas public-page copy also uses `not_assessed`. The separate Desktop M0 archives have no `scientific.validation_status` declaration; they are not additional conforming specification examples. No surveyed local recipe uses the two legacy values. Third-party usage cannot be inferred from this local scan. Keep `cars_version: "0.1.0"`, the Schema path and `$id` for this community-draft erratum: it restores the original prose contract without changing the manifest shape or evidence model. This is still a change to the accepted input set, not full backward compatibility with the faulty Schema. Before publication, announce this erratum, identify the corrected validator/Schema artifact by revision or hash, and notify consumers pinning the initial Schema. Do not overwrite an immutable published release artifact. A future field/model change needs a new specification version; no ratified versioning/deprecation policy exists yet. ## Other alignment findings (reported, not changed) - §5.3 lists six `result_kind` values, but input/output property definitions in Schema accept any nonempty string. The L0 subset does not enforce that list. - §5.1 calls recipe `version` SemVer, but the Schema regex accepts invalid numeric prerelease identifiers such as `1.0.0-01` and empty identifiers such as `1.0.0-alpha..1`. Tightening this is a separate compatibility change. - §5.1 describes `title` as human-readable and optionally localized; Schema requires a nonempty localization object. A plain string is rejected. The prose leaves the representation ambiguous and should be made explicit separately. - The initial data-type registry in §5.2 is descriptive; Schema allows arbitrary nonempty I/O type strings. This is an open registry boundary, not an implemented closed enum. No defaults are declared in Schema. Top-level MUST/SHOULD/MAY requiredness agrees with Schema. Adapter types and modes agree with §9; nested field requirements are mostly specified only by Schema, leaving prose gaps. - Evidence sufficiency, reviewer independence, references, units and semantic checks remain outside L0. Accepting a declared evidence status without proving it is an explicitly documented limitation, not scientific certification. The stale §6 claim that the inline example had never been Schema-validated was updated now that the conformance suite validates it. No other constraints were changed as part of this erratum.