6. What evidence can carry

Claims and their grounds

Evidence has a subject and a scope. A parser test, a visual inspection, a source comparison and a user trial answer different questions. A polished representation cannot establish semantic correctness. A correct calculation does not establish that its published rendering keeps the result.

Verification concerns conformity to an identified requirement, claim or reference. Validation concerns fitness for intended use. Tools also use these words for narrower operations, so the actual subject matters more than the label. Acceptance of a document's format alone does not show whether the document promises a useful product.

The Root F-02 case in chapter 9 shows why a claim needs grounds that suit its actual subject. A current caller and executable behavior challenged a plausible broad interpretation and required a narrower explanation with better support. A successful check alone would not have explained what was wrong with the earlier interpretation.

Consequence, detectability and independence

A wrong result that looks plausible can be more dangerous than an obvious failure. So numerical errors, permission mistakes and silent data loss often need stronger examination than a reversible change of wording. The value of a check depends on the failure it can expose and on the consequence of missing that failure.

A shared oracle can make many results agree while it carries the same mistaken assumption into all of them. Independent examination concerns preparation and grounds, not only different identities in a review record. It can use a source comparison, another derivation, a difficult case or scrutiny of the oracle itself.

Proportionate examination is an argument about adequate grounds under the governing conditions. It does not permit anyone to abandon an adopted criterion because the criterion is inconvenient. If a criterion is wrong, that becomes a substantive question for the appropriate decision. It is not an excuse to redefine a failed result as a pass.

Candidate identity and evidence reuse

An examination needs a candidate that people can identify. A moving branch name, a screenshot from an earlier build or a flattened extract of a source can refer to a different subject from the one under discussion. A hash can make that identity precise. It does not show that the check was adequate.

Earlier evidence can stay applicable when the conditions that support its claim stay unchanged. Those conditions can include requirements, inputs, implementation, the check and the environment. A changed document does not necessarily invalidate an executable path that was tested before. And an unchanged executable can still face a new use that its earlier tests did not address.

So the question is whether evidence applies. It is not a choice between universal reuse and universal repetition. A copied table can need checks of placement and rendering without a new derivation of unchanged numbers. A different consumer can need new interface evidence even when the copied bytes are identical.

The library delivery case separated two kinds of support that a summary can easily merge. The match between the complete archive and Git established which source was supplied. Selected operations on an extracted archive established bounded behavior. The first result could be exhaustive about file identity, while the second stayed selective about use. Neither result had to claim more than it established for the combination to be useful to a recipient.

Retained evidence serves examination or recovery. The form and duration of retention depend on that use and its governing obligations. A raw log can be essential for one claim, while a reproducible fixture and a source binding are enough for another. Missing evidence remains a limit. A narrative written afterward cannot silently become an observation made at the time.

What a failure changes

A failure can come from the product, the environment, the test or its assumptions. A passing rerun can reveal variability without proof that the cause is repaired. If someone changes an expected value only to silence the failure, they change the question and keep the appearance of an answer.

A consequential finding can invalidate one dependency without invalidating everything around it. Keeping valid work depends on understanding what changed. A simple reduction in testing does not give that understanding. The same understanding lets independent work continue while participants investigate the affected claim.

A useful examination stays open to the possibility that its own framing is wrong. The question then becomes which response fits: repair the implementation, improve the evidence, or reconsider the commitment. Those responses have different consequences and different decision rights.