Verification Matrix

The matrix below names the strongest relationship the repository currently supports for each external reference. A stronger row must not be inferred from a weaker one: successful execution is not numerical agreement, and a checked study fixture is not a package-wide live parity surface.

Read each row as a ceiling on the public statement, not as a score. The reference name identifies an external owner; the verification mode identifies how that owner was observed; the strongest claim states the largest supported scope; and the primary record contains the denominator, failures, and limits.

Reference	Verification mode	Strongest current claim	Primary record
APE	live R registry; checked study fixtures	180 governed cases across 35 function shards, plus named study comparisons	APE correspondence
GEIGER	live R registry	48 governed cases across 17 model shards with optimizer and parameterization review	GEIGER correspondence
phytools	live R registry	65 governed cases across 21 function shards, with distribution rules for stochastic outputs	phytools correspondence
APE + nlme	checked R-derived fixtures	PGLS coefficients, likelihoods, model comparisons, and selected diagnostics for named fixtures	study-specific R references
treeio + tidytree	governed study evidence	structural node mapping and tree–trait join observations for the primate longevity fixture	study-specific R references
ggtree	source and plot-intent provenance	plot-only study fragments	study-specific R references
MAFFT + trimAl	external execution adapters	validated invocation, captured configuration and output, parsed alignment diagnostics	alignment and trimming
IQ-TREE2 + FastTree	external execution adapters and cross-engine workflow	complete run bundles and comparable topology/likelihood fields where meanings align	tree inference
MrBayes + BEAST	external execution and parser adapters	prepared or executed analyses with trace/tree diagnostics and explicit engine ownership	Bayesian inference
MCMCglmm	source-backed study boundary	PCM4 and PCM5 bundles retain not_comparable verdicts and open posterior follow-through; no posterior parity claim	study-specific R references

Mode, Outcome, And Verdict

Term	Applies to	Read from
verification mode	strength and ownership of the comparison relationship	this matrix and the tool-specific guide
case outcome	passed, failed, or skipped execution of a registered parity case	parity summary, observations, and failure artifacts
evidence verdict	governed scientific conclusion for identified claims	Evidence Book manifest and indexes
resolution state	open follow-through, blocker, or accepted bounded state	verdict workflow and study boundary records

Do not translate a mode into an outcome or an outcome into a verdict. They answer different questions and may change on different schedules.

Coverage Requires The Full Case Population

A passing-case count is uninterpretable without the selection and execution denominators.

Count	Membership rule	Required visibility
registered	every case defined by the governed registry revision	operation, parameterization, fixture and support state
selected	cases included by the declared filter before execution	filter, exclusions and unsupported boundaries
executed	selected cases for which both required lanes reached a comparable observation stage	failures, missing tools and incomplete outputs
passed	executed comparable cases satisfying every registered rule	keyed observations and rules
failed	executed comparable cases violating at least one rule	mismatch artifacts and owning observation
skipped or not comparable	selected cases without an eligible comparison	exact absence, capability, environment or identity reason

Publish all counts and their case keys. A pass rate over executed cases can be useful, but it must not hide selected cases that never became comparable. A larger registry is stronger coverage only when its new cases retain explicit outcomes and comparable observation contracts.

Verification Has Independent Axes

Do not read the modes as one universal maturity ladder. Strength depends on at least four independent properties:

Axis	Weaker evidence	Stronger evidence
execution freshness	frozen source or fixture	live execution under recorded versions
comparison precision	provenance or structural compatibility	named observations with explicit tolerances
scientific scope	one operation or adapter contract	claim-bound study evidence
reproducibility	prose description	retained inputs, raw outputs, manifest, observations, and failures

A checked fixture can be more reproducible than an undocumented live run. A live run can be fresher while remaining scientifically narrower than a study bundle. An external adapter can have excellent execution evidence without any numerical correspondence claim. Report the axes that matter instead of assigning one undifferentiated strength label.

What Each Mode Establishes

flowchart TD
    provenance["Source provenance"] --> fixture["Checked fixture"]
    fixture --> live["Live reference case"]
    adapter["External adapter"] --> execution["Governed execution"]
    live --> evidence["Evidence Book claim"]
    execution --> evidence
    fixture --> evidence

    provenance_note["Names the original owner"]
    fixture_note["Replays frozen observations"]
    live_note["Executes both implementations now"]
    execution_note["Captures an external engine run"]
    evidence_note["Adds sources, checks, artifacts, and verdict"]

    provenance --- provenance_note
    fixture --- fixture_note
    live --- live_note
    execution --- execution_note
    evidence --- evidence_note

• Source provenance identifies where an operation or study originated.
• Checked fixture compares Bijux with frozen, versioned reference outputs.
• Live reference case executes the named external package and Bijux in the same governed run.
• External adapter prepares, invokes, captures, and parses a program that remains the computational owner.
• Evidence Book claim binds a scientific statement to inputs, sources, checks, results, freshness, and a reviewable verdict.

Reading A Claim Safely

Start with the verification mode, then inspect the named operation and compared fields. Confirm parameterization, taxon order, state encoding, rooting, units, likelihood convention, and support meaning before interpreting a passing case. If a scientific claim cites the run, continue to the Evidence Book verdict and its freshness. Finally, read the non-claims: they define what neither record can be used to conclude.

A row may advance only when the stronger mode is implemented and governed. Documentation language alone cannot promote provenance to a checked fixture, an adapter to parity, or an analytical match to rendered-figure equivalence.

Select The Record From The Intended Statement

Intended statement	Minimum owning record
Bijux can prepare or invoke a named external engine	adapter contract with executable discovery, command, process state, and native outputs
one registered operation corresponds for named cases	parity registry plus selected-case summary, keyed observations, rules, and failures
one checked fixture reproduces a reference observation	fixture provenance, immutable reference value, Bijux execution, and comparison record
one scientific result corresponds under a study contract	Evidence Book claim, complete observation denominator, verdict, limitations, and freshness
two ecosystems are generally equivalent	no matrix row supplies this claim; define and govern a finite operation and population first

Choose the record whose scope matches the sentence. More operational detail does not compensate for a missing scientific claim, and a study verdict does not expand to every function offered by its reference package.

Revision Rule

Registry sizes and strongest claims describe the checked-in revision. When a registry, normalization rule, reference version, tolerance, or governed fixture changes, regenerate its summary and reconsider the matrix claim. A larger case count is not automatically stronger evidence if new cases are skipped, incomparable, or governed by weaker observations.

The citation should retain the repository revision, verification mode, reference and Bijux versions, selected operation and population, observation rule, complete denominator, strongest supported claim, and linked evidence verdict when one exists. Without the revision and denominator, the matrix row is navigation rather than a reproducible correspondence statement.