Claim Verdicts And Resolution

A verdict reports the outcome of a registered comparison. Resolution reports what happens next. Keeping them separate prevents open work from weakening a real mismatch and prevents a detailed plan from masquerading as completed evidence.

Locate The Status Level

flowchart TB
    observation["Observation status<br/>one scalar, row or structure"] --> claim["Claim verdict<br/>one bounded statement"]
    claim --> bundle["Bundle summary<br/>registered claim collection"]
    bundle --> index["Repository indexes<br/>coverage and freshness"]
    claim --> resolution["Resolution record<br/>owner, blocker and closure"]

Never lift a favorable status upward without checking its denominator. A matched observation does not make its claim matched; a matched claim does not make every observation in the bundle matched; a repository readiness result does not make every scientific method validated.

Summaries Do Not Promote Verdicts

Higher-level views preserve lower-level states; they do not replace them.

Summary surface	Defensible content	Forbidden promotion
observation table	counts and identities for exact, within-budget, mismatched, absent, and excluded rows	“overall match” without the claim aggregation rule
claim record	one verdict derived from its complete required population	extension to an unregistered quantity or population
bundle report	inventory of its claims, observations, contradictions, and limitations	selecting the most favorable claim as the bundle meaning
study guide	scientific relationship among bundles and the current mixed evidence state	a study-wide pass/fail label
repository dashboard	coverage, freshness, integrity, and verdict counts	numerical or scientific validation of every runtime surface

An aggregate can say that ten observations matched and two did not. It cannot say the claim matched unless the claim contract declares those twelve rows, their eligibility, and the rule that yields that verdict. Likewise, a release gate can establish publication eligibility for its governed inventory while a scientific mismatch remains visible and unresolved.

Verdict Vocabulary

Verdict	Required meaning	Honest downstream language
matched	all claim-required comparable observations satisfy the registered exact or structural rule	the bounded claim matches under the recorded contract
matched_with_tolerance	all claim-required comparable observations satisfy the predeclared numerical budget	the bounded claim agrees within the named tolerance
mismatch_explained	a comparable difference exists and its registered cause and consequence are established	the claim differs for the recorded reason; state its impact
mismatch_unexplained	a comparable difference violates the rule without an accepted explanation	the claim does not currently correspond
not_comparable	the required estimand, identity, execution, or output alignment is unavailable	no valid comparison is currently owned

not_comparable is not a polite match and mismatch_unexplained is not merely “needs review.” Neither may be converted into matched_with_tolerance by widening the tolerance after observing the result.

Keep Comparability, Agreement, And Currency Separate

Three independent decisions qualify a claim:

Axis	Question	Possible outcome
comparability	do both sides address the same population, estimand, identity, and required output?	comparable or not_comparable
agreement	for comparable observations, does the registered exact, structural, or tolerance rule hold?	matched, matched_with_tolerance, mismatch_explained, or mismatch_unexplained
currency	do the verdict and observations still resolve to unchanged material dependencies?	current or stale

The axes cannot compensate for one another. A current record may remain not_comparable; a numerically matched historical record may be stale; and a fresh mismatch remains a mismatch until a new governed observation chain is adjudicated. Public language must carry all applicable axes rather than compressing them into “valid” or “invalid.”

Read Similar-Looking Records Differently

Three result directories can all contain source descriptions, analysis code, checks, and reports while supporting different sentences:

Record state	Public sentence
aligned outputs satisfy all registered rules and freshness is current	the named claim corresponds under its exact comparison contract
aligned outputs violate a rule and the cause is not established	the named claim is currently mismatched; investigation remains open
a required execution or primary output is absent	the claim is not currently comparable; the dossier defines what must be produced

The amount of documentation in a directory does not choose among these sentences. Primary outputs, observation alignment, the verdict rule, and freshness do. A detailed contract-only bundle can be excellent evidence of what remains to run while still being no evidence of numerical agreement.

Derive A Claim Verdict From Its Required Population

Start from the claim dependency set, not from whichever rows are easiest to compare:

flowchart TD
    required["Enumerate required observations"] --> aligned{"Every required observation<br/>has aligned identities and outputs?"}
    aligned -- no --> incomparable["not_comparable"]
    aligned -- yes --> rules{"Apply every predeclared rule"}
    rules -- all exact --> exact["matched"]
    rules -- all within budgets --> tolerance["matched_with_tolerance"]
    rules -- violation with accepted cause --> explained["mismatch_explained"]
    rules -- violation without accepted cause --> unexplained["mismatch_unexplained"]

The denominator includes every observation the claim declares, including missing, excluded, failed, and non-comparable members. A verdict may summarize that population only when the aggregation rule is defined in advance and the machine record retains each member. “Most rows matched” is a descriptive count, not one of the governed claim verdicts.

When exact and tolerance rules coexist, report the strongest vocabulary that is true for the complete claim population. Do not label the claim matched because its exact subset passed while required tolerance-based observations were merely within budget.

Apply Verdicts To The Present Studies

The same repository can legitimately retain several evidence states because each state belongs to a different claim and observation population.

Present record	Correct interpretation	Incorrect compression
PCM1 focused preparation bundle marked matched	the named preparation invariant corresponds for its registered inputs	every PCM1 output is exact or every figure is equivalent
PCM1 aggregate bundle marked matched_with_tolerance	all claim-required comparable observations meet their predeclared budgets	numerical identity or proof on unrelated primate data
PCM2 focused matches beside 12 broader mismatch_unexplained scalar rows	bounded bundle claims and unresolved broader observations coexist	PCM2 “passes overall” because most bundles are favorable
PCM2 intercept sweep marked not_comparable	no governed aligned observation currently supports that comparison	the two implementations were compared and disagreed
PCM3–PCM5 contract-only bundles marked not_comparable	the intended claim, inputs, source, and missing execution boundary are governed	wrappers, nearby methods, or complete manifests constitute numerical evidence

These examples are not exceptions to the vocabulary. They demonstrate why a verdict must travel with its claim identifier, population, method, observations, and limitations.

Keep Other Statuses In Their Own Records

Execution states such as blocked, failed, partial, and complete belong to run records. Freshness states belong to dependency review. Coverage gaps describe missing proof. Plot-only, workflow-only, artifact-only, and seeded-input-only classify tracked surfaces that are not analytical correspondence. Preserve these dimensions beside the verdict instead of forcing them into it.

Resolve By Earliest Owning Cause

Problem	Owning layer	Closure evidence
source or dataset unavailable	provenance and dataset registry	governed locator, checksum, license and admitted population
reference program never executed	reference execution	versioned invocation, terminal record and primary outputs
Bijux program never executed	Bijux execution	runtime identity, invocation, diagnostics and primary outputs
estimands or identities differ	comparison design	aligned parameterization and stable observation keys
registered tolerance fails	observation	corrected implementation or accepted causal explanation plus consequence
claim lacks required observations	claim dependency map	complete eligible observation denominator
material dependency changed	freshness record	regenerated bundle and current indexes

Resolve the earliest broken edge. Editing a report cannot repair a missing execution; rerunning a program cannot repair an ill-defined estimand; marking an item closed cannot repair a failed tolerance.

Preserve Contradictions

When bundle, claim, observation, report, and generated index disagree, retain all conflicting records and stop the affected claim. Identify the owning source, rebuild projections, and review whether the contradiction changed a published interpretation. Do not choose the most favorable status or delete the row that prevents a clean summary.

Record Resolution Without Rewriting History

A resolution record names the affected claim and observations, current verdict, blocker or causal explanation, responsible owner, required execution or evidence, acceptance rule, and closure artifacts. When work closes, create new governed outputs and freshness state. Historical mismatches and non-comparable states remain part of the audit trail.

A corrected implementation may produce a matching new comparison, but it does not mean the earlier mismatch never occurred. Report both when the history matters to released artifacts or scientific conclusions.

Closure requires a new observation chain, not only a resolution-state edit. The closure record should link the corrected input, execution, result, comparison, claim verdict, and freshness projection. If an accepted causal explanation leaves the numerical difference in place, use mismatch_explained; explanation makes the discrepancy interpretable, not equal.

Write The Public Statement

Use this pattern:

Claim <claim_id> is <verdict> for <population and method> under <comparison rule>. The record excludes <limits>. Freshness is <state> at <repository revision>. Resolution state is <state and owner>.

Omit no material qualifier. If a statement cannot name its claim identifier, population, method, comparison rule, and current verdict, it is broader than the Evidence Book record supports.