Runtime And Evidence Boundary

Runtime output becomes governed evidence only through an explicit claim contract. Executing a method can produce a complete, reproducible result without establishing a study claim; an Evidence Book bundle adds the source, claim identity, checks, acceptance rules, freshness, and verdict required for adjudication.

flowchart LR
    request["Validated scientific<br/>request"]
    execution["Native method or<br/>external engine"]
    result["Typed result and<br/>diagnostics"]
    artifacts["Manifest and linked<br/>artifacts"]
    mapping["Claim and source<br/>mapping"]
    checks["Declared observations<br/>and acceptance rules"]
    verdict["Claim-scoped verdict<br/>and freshness"]

    request --> execution --> result --> artifacts
    artifacts --> mapping --> checks --> verdict

Runtime Result Contract

A runtime result identifies the method owner, inputs, configuration, status, diagnostics, and outputs. It answers whether the requested computation executed under its declared operational contract. It may also carry numerical or integration-test support, but it does not inherit a scientific claim that was never mapped to it.

Evidence Claim Contract

An evidence claim binds a scientific statement to source provenance, identified inputs, an executable or externally bounded reference, observation-level checks, acceptance rules, artifacts, a verdict, and material dependency freshness. The claim is no broader than that binding.

Promotion Requirements

Requirement	Runtime output alone	Governed evidence claim
method and software identity	required	required
input and configuration identity	required	required
diagnostics and output inventory	required	required
stable claim identifier	optional	required
scientific source provenance	optional	required
reference or closure boundary	optional	required
declared observations and tolerances	optional	required
claim-scoped verdict and freshness	no	required

Copying a report into the Evidence Book does not satisfy these requirements. Promotion requires a registered bundle whose machine-readable records connect the claim to the exact runtime or external computation.

Promotion Creates A Child Record

Evidence promotion does not mutate a runtime result into a different kind of truth. It creates a claim record that consumes an immutable result and adds scientific source, comparison, adjudication, and freshness identities.

flowchart LR
    result["Runtime result<br/>status and diagnostics"] --> artifact["Immutable artifact<br/>identity and inventory"]
    artifact --> observation["Claim observation<br/>normalization and rule"]
    observation --> verdict["Bounded verdict<br/>limitations and freshness"]

    artifact -. remains runtime-owned .-> artifact
    verdict -. links, never rewrites .-> artifact

Change after promotion	Required action
result bytes, schema, or diagnostics change	create a new result identity and reopen dependent observations
normalization or tolerance changes	create a new comparison rule and adjudication record
claim wording or population changes	register a new or revised claim dependency set
only explanatory prose changes	preserve machine values, verdict, limitations, and links exactly

This child relationship preserves both histories: the runtime record explains what happened, while the Evidence Book explains which bounded statement may consume it. Editing one to resemble the other destroys the audit boundary.

Stop At The Strongest Record Actually Owned

Evidence accumulates by adding specific records. It does not advance merely because a neighboring capability exists.

flowchart LR
    capability["Executable capability"] --> result["Identified result"]
    result --> reproducible["Reconstructable run"]
    reproducible --> comparison["Registered comparison"]
    comparison --> claim["Adjudicated claim"]

    capability -. "valid stopping point" .-> capability
    result -. "valid stopping point" .-> result
    reproducible -. "valid stopping point" .-> reproducible
    comparison -. "valid stopping point" .-> comparison

Each stop is useful when described honestly:

Strongest retained record	Defensible statement	Statement still unavailable
executable capability	the method or adapter is available under its public contract	this study result was reproduced
identified result	the named inputs and configuration produced this result and diagnostic state	an independent implementation agrees
reconstructable run	the computation can be reconstructed from retained identities and artifacts	the scientific claim passes
registered comparison	named observations were compared under a declared rule	neighboring outputs or parameterizations agree
adjudicated claim	the bounded claim has the recorded verdict and freshness state	broader scientific validity or independent replication

A missing higher layer is neither erased by a lower-layer success nor evidence that the lower layer failed. It determines the strongest language the public record can support.

Evidence Record Anatomy

A governed bundle separates facts that are often collapsed into one narrative:

Record	Owns	Review question
manifest.json	bundle identity, ownership, summary, limitations, freshness, and verdict	What is the current bounded conclusion?
claims.json	stable claim identifiers, source fragments, tags, and claim verdicts	Which statement is being adjudicated?
inputs.manifest.json	exact input files, objects, roles, and checksums	Which analytical population and tree were used?
provenance.json	source origin, retrieval or derivation, and reference identity	Where did the comparison basis come from?
checks.json	required artifacts, primary outputs, comparison mode, and validation rules	What must exist before the verdict is reviewable?
results/	governed observations, tables, summaries, and their inventory	What values and structural facts were actually retained?
report.md	human interpretation of the same bounded record	Does the prose preserve the machine-readable status and limitations?

No single file substitutes for the bundle. A result table without input identity is ambiguous; a claim without primary outputs is unsupported; a report whose wording exceeds the manifest verdict is inconsistent.

Do Not Promote By Proximity

flowchart TB
    capability["Nearby runtime capability"]
    execution["Study-specific execution"]
    observations["Retained reference and<br/>Bijux observations"]
    checks["Declared comparison<br/>and acceptance rule"]
    verdict["Claim-scoped verdict"]

    capability -. does not imply .-> verdict
    capability --> execution --> observations --> checks --> verdict

An implemented PGLS surface does not close a PGLS lecture claim until the study-specific design matrix, taxa, covariance, outputs, and checks are retained. A BEAST parser does not close an MCMCglmm posterior claim. A parity case for one parameterization does not transfer to another model carrying the same family name.

Changes That Reopen A Claim

A previously favorable verdict requires freshness review when a material dependency changes:

• the scientific algorithm, model parameterization, optimizer, or sampler;
• parsing, normalization, tolerance, or comparison semantics;
• datasets, tree topology, branch lengths, exclusions, or source provenance;
• result schemas, serialization, or governed primary-output paths;
• the reference executable, package version, or external-engine contract;
• claim wording broad enough to exceed the retained observations.

Presentation-only edits need not invalidate numerical evidence, but they must still preserve exact values, verdict vocabulary, and limitations. Freshness is an explicit dependency decision, not an assumption based on a recent file timestamp.

Valid Boundary States

• A native method can be implemented and tested without a study bundle.
• An adapter can be fully operational while the external algorithm remains the computational owner.
• A bundle can be structurally valid and current while its verdict is not_comparable.
• A matched study observation does not grant parity to neighboring models, datasets, or package surfaces.
• A released runtime does not make stale evidence current.

Use the claim reading guide to qualify public statements, validation lanes to identify the proof family, and the Evidence Book to inspect governed study verdicts.