Evidence Consumer Boundary

Evidence governance is a downstream consumer of runtime artifacts. Ordinary analysis must not require a study dossier, claim registry, or evidence index; Evidence Book workflows may depend on stable runtime and artifact contracts to evaluate named claims.

One Run, Three Acceptance Decisions

Decision	Owner	Accepts	Does not establish
execution acceptance	workflow or engine adapter	identified process state, required native outputs, and captured diagnostics	scientific suitability or correspondence
result acceptance	scientific method and consumer	admitted population, model, structured values, uncertainty, exclusions, and intended use	agreement with a reference implementation
claim acceptance	Evidence Book bundle	complete claim dependencies, comparison rules, verdict, limits, and freshness	broader runtime correctness or unobserved methods

These decisions may differ without contradiction. A complete execution can be scientifically rejected; a valid result can have no comparison claim; a structurally current bundle can retain not_comparable. Preserve the three statuses rather than replacing them with a single “validated” label.

flowchart LR
    runtime["Scientific runtime"]
    result["Typed result"]
    bundle["Artifact bundle"]
    evidence["Evidence governance"]
    indexes["Verdicts, freshness,<br/>coverage indexes"]

    runtime --> result --> bundle --> evidence --> indexes
    evidence -. must not control .-> runtime

Dependency Rule

The runtime owns scientific objects, computation, typed diagnostics, and serialization. Evidence governance owns source mapping, claim identifiers, check definitions, acceptance rules, verdicts, and freshness. A dependency in the opposite direction would make ordinary execution depend on a selected study and would blur computation success with claim adjudication.

Consumption Contract

Evidence consumer needs	Runtime or artifact provider supplies
stable operation identity	documented method, workflow, engine, and version
reproducible input scope	input hashes, taxa, reconciliation, partitions, transformations
comparable configuration	model, priors, costs, search/sampling policy, seeds
observation source	typed result fields and machine-readable ledgers
diagnostic qualification	convergence, support, uncertainty, warnings, exclusions
output integrity	manifest, schema version, inventory, checksums where governed

Evidence tooling should reject or mark unresolved any claim whose required runtime field, artifact, source, or comparison owner is missing. It must not reconstruct critical scientific values by scraping prose or figures.

Promotion Protocol

flowchart LR
    result["Typed runtime result"]
    bundle["Complete artifact bundle"]
    observation["Normalized and raw<br/>comparison observation"]
    adjudication["Acceptance rule<br/>and verdict"]
    publication["Fresh claim index"]

    result --> bundle --> observation --> adjudication --> publication

Promotion is claim-specific. The evidence consumer must name the result field or artifact, the reference owner, any normalization, the tolerance or decision rule, and the claim that consumes the observation. Missing comparison data produces not_comparable; it does not become a runtime failure or an “open” verdict.

The reverse implications are invalid:

• a published verdict does not make the Evidence Book the runtime owner;
• a matched claim does not validate fields the claim did not observe;
• a fresh bundle does not imply a favorable verdict;
• a complete runtime result does not imply a reference comparison exists.

Material Changes And Freshness

A change to a consumed method, parameterization, input, dataset, external reference, parser, schema, or acceptance rule can invalidate the relationship between an old artifact and its claim. Freshness evaluation records that dependency change. It does not rewrite the historical result; it prevents the old verdict from being presented as current.

Freshness must follow material dependencies rather than file timestamps alone. A prose-only clarification can leave a verdict current; a default, parser, dataset, tolerance, or reference-version change can require new observations even when the public function name is unchanged.

Valid Independence

• Runtime capability can exceed Evidence Book coverage.
• A complete runtime bundle can have no claim mapping.
• A governed claim can remain not_comparable while its bundle is structurally valid and current.
• Evidence indexes can be regenerated without changing runtime behavior.
• Runtime tests and benchmarks can support bounded assurance without becoming study verdicts.

See the runtime and evidence boundary for promotion requirements and the artifact consumption guide for reviewing a persisted run before claim adjudication.