Claim Reading Guide

Read a public claim from the noun outward. Identify exactly which method, workflow, study, corpus, or release is being described before evaluating the confidence word attached to it.

Decode The Vocabulary

Word	Minimum meaning in these docs	Look for
implemented	executable code owns the stated behavior	source and public contract
tested	a declared contract passes fixtures	test identity and invariant
benchmarked	metrics were collected on a named corpus	environment, configuration, metric, threshold
parity	declared outputs agree under stated conventions and tolerance	reference, compared fields, tolerance, verdict
evidence-backed	a governed bundle supports a bounded claim	claim ID, provenance, checks, freshness
released	published artifacts passed release gates	version, revision, package and docs alignment

None of these words means “appropriate for every scientific question.”

Trace A Claim

flowchart LR
    statement["Public statement"] --> scope["Subject and scope"]
    scope --> lane["Validation lane"]
    lane --> record["Test, benchmark,<br/>or evidence record"]
    record --> rule["Acceptance rule"]
    rule --> verdict["Verdict and freshness"]
    verdict --> limits["Exclusions and<br/>open boundaries"]

1. Extract the subject and quantifier: one method, all datasets, a named study?
2. Identify the proof lane actually cited.
3. Open the primary record rather than relying on summary prose.
4. Confirm compared inputs, versions, configuration, metric, and tolerance.
5. Check the verdict and whether the record remains fresh.
6. Carry exclusions and limitations into your own use of the claim.

Reconstruct The Claim As A Contract

A reviewable claim is the product of six bounded fields:

subject × population × method × observation × acceptance rule × revision

Removing any field changes the statement. “PGLS agrees” does not identify the traits, taxa, covariance convention, compared coefficient, tolerance, or software revision. A green verdict cannot restore information that the claim never declared.

For example, a bounded PCM2 claim can be read as follows:

Field	Bounded meaning
subject	the registered PGLS covariance observation
population	the retained mammal taxa after named reconciliation and exclusions
method	the declared Brownian covariance construction and regression parameterization
observation	the registered matrix entries or derived coefficient named by the check
acceptance rule	the stored structural rule and numerical tolerance
revision	the runtime, reference, input, and bundle identities recorded by the evidence manifest

The contract supports only the intersection of those fields. It does not imply that every PGLS model, taxon set, covariance convention, or output agrees.

Transfer A Claim Only To A Narrower Contract

A governed claim can support a downstream statement only when the downstream contract is contained within every material dimension of the source claim.

Proposed transfer	Valid only when
one observation from a multi-observation claim	the observation is registered, eligible, adjudicated, and carries the same rule and revision
one taxon subset from a study population	the claim explicitly supports subset interpretation or the analysis is rerun on a newly identified population
one model from a model family	the exact parameterization and result are inside the governed claim
a newer runtime or reference version	freshness review establishes that every material dependency remains applicable
a scientific statement from a parity result	a claim record explicitly consumes that parity observation and preserves its limitations

Changing the population, estimand, method, observation, acceptance rule, or revision creates a different contract. Similarity, monotonicity, or intuitive scientific plausibility is not an entailment rule. When containment cannot be shown, cite the original narrower claim as context and leave the new statement unevidenced until it has its own record.

Qualifiers That Change Meaning

Qualifier	Question it must answer
native	Which repository-owned implementation performs the computation?
external or adapter-backed	Which executable owns the algorithm, and what does Bijux orchestrate?
checked fixture	Which frozen reference values and versions were retained?
live parity	Were both implementations executed in the governed run?
matched with tolerance	Which observations, absolute/relative rules, and boundary conventions matched?
current	Which material dependencies were checked for freshness?
complete	Which required cases and artifacts were included, and how were failures counted?

Adjectives cannot carry scope by themselves. “Current parity” remains incomplete without the named reference, registered cases, compared fields, tolerances, and freshness basis.

Example

“Native maximum-likelihood inference is benchmarked” is incomplete. A usable version says which native entry point, corpus, model/search configuration, comparison metric, environment, and acceptance threshold were measured. If a study claim relies on that benchmark, the evidence bundle must make the link explicit.

A stronger claim need not be longer. It must use nouns that resolve to owned records and qualifiers whose exact meaning is declared.

Scope Traps

• Evidence for a nucleotide model does not transfer to protein or codon models.
• Agreement on one tree or dataset does not establish corpus-wide parity.
• A normalized adapter result does not make the external algorithm native.
• A fresh release does not automatically refresh an old evidence bundle.
• A successful workflow does not resolve weak convergence or model fit.
• A study-family summary does not make every bundle pass.

Combining Proofs

Proofs can support one claim together only when their subjects and assumptions align. A parser contract, numerical known answer, workflow integration test, and study comparison may form a strong chain for one method and configuration. They cannot be combined by counting unrelated green checks.

When combining records, preserve:

• the narrowest shared method and input scope;
• every material configuration difference;
• the weakest current verdict or missing link;
• separate ownership for native and external computation;
• the revision or version to which the chain applies.

Strongest Available Proof

Use source and API contracts to verify ownership. Use tests for interface and numerical invariants. Use benchmarks for named comparative behavior. Use the Evidence Book for study claims. Use release gates for current publication alignment. Do not cite a broader but weaker surface when a narrower primary record exists.

Apply The Counterevidence Test

A trustworthy claim states what observation would weaken it. Before accepting the wording, identify the failure record, contradictory result, dependency change, or excluded case that would force the claim to be narrowed.

Claim form	Counterevidence that must change the claim
implementation contract	a supported input violates the documented result or refusal invariant
numerical result	an independent known answer falls outside the registered tolerance
parity	any required case is missing, not executed, differently parameterized, or outside tolerance
benchmark	the registered environment or workload no longer satisfies the acceptance rule
evidence-backed study claim	a material dependency is stale, provenance breaks, or a required observation is contradictory
released surface	package, schema, docs, or artifact identity does not resolve to the admitted revision

Do not repair a contradicted claim by deleting the case, changing the denominator, widening tolerance after observation, or replacing the primary record with summary prose. Register a new claim or comparison when the subject or acceptance rule genuinely changes; preserve the earlier verdict as part of the evidence history.

When Evidence Is Missing

Downgrade the wording, not the standard. Say “implemented,” “available,” or “not yet evaluated under this study contract” according to the facts. Record the missing comparison or closure criterion. Absence of evidence is not a failure to market aggressively; it is a boundary a reader deserves to see.

If a claim cannot be traced, do not infer its proof from repository size, coverage percentage, publication status, or neighboring examples. Treat it as an undocumented or unsupported claim until the owning record is identified.

Continue with validation lanes and the Evidence Book.