Evidence and claim boundaries

Use this page when a passing test, generated artifact, provenance record, or investigation must be translated into a claim without overstating what it proves.

Evidence classes¶

The terms below answer different questions and must not be collapsed.

Object	Question answered	What it cannot prove
Scientific contract	What behavior, assumptions, and failure boundary does a callable claim?	That every implementation or downstream use satisfies the contract
Validation target	Which executable check probes a named part of the contract?	That untested domains or models are valid
Evidence artifact	What metrics and comparisons were emitted under a recorded environment?	That the scientific model is adequate
Provenance	Which sources, inputs, methods, hashes, and environment produced or support a record?	Authority, correctness, or sufficiency by itself
Investigation	How do prediction, computation, audit, and limitation connect for a bounded example?	General validity outside the fixture
Warranted scientific claim	What conclusion is supported after assumptions and evidence are considered?	Any broader conclusion not named by that chain

From contract to claim¶

Let $C$ be a contract, $V$ a set of validation results, $P$ provenance, and $A$ the scientific assumptions. A claim is warranted only within their intersection:

\mathcal{W} \subseteq \mathcal{C}\cap\mathcal{V}\cap\mathcal{P}\cap\mathcal{A}.

(1)

This is a reasoning relation, not a numerical formula. More passing tests do not compensate for a false scientific assumption, and detailed provenance does not turn a weak comparison into strong validation.

Fail-closed reasoning¶

Missing evidence remains visible. An unresolved contract ID, stale generated artifact, failed comparison, absent source, or invalid environment gate must not be converted into a positive claim. No standalone indexed artifact is more informative than silently inferring evidence from a passing example.

Thresholds belong to the method-specific contract. Tightening or weakening a tolerance without an error model changes the claim. Informational telemetry must be distinguished from pass/fail comparisons so that a finite number is not mistaken for acceptance.

Audit procedure¶

State the proposed claim in one bounded sentence.
Link every noun in the claim to a contract or declared scientific assumption.
Identify independent validation targets and their tested domain.
Inspect artifact freshness, metric definitions, units, thresholds, and environment policy.
Inspect runtime and source provenance separately.
List known limitations and negative results.
Narrow the claim until every remaining clause is supported.

Worked boundary¶

A rootfinding artifact can show residual, bracket width, status, and evaluation count for benchmark fixtures. A gradient artifact can compare an implicit rule with analytic and finite-difference sensitivities. Together with a uniqueness and smoothness argument they may support the fixture’s local derivative. They do not prove that an arbitrary physical residual has one branch or that a downstream parameter is identifiable.

Where the claim stops¶

Evidence infrastructure makes reasoning inspectable. It does not automate the scientist’s responsibility to judge source adequacy, model assumptions, or external validity.

Connected ideas¶

Use Models, inference, and information, Root-finding, Provenance for reproducible research, and Executable research investigations.