Evidence Book

The Evidence Book is the review layer between scientific source material and public claims. It records what was run, with which inputs and versions, which checks were applied, and which governed comparison verdict each bounded claim currently carries.

flowchart LR
    source["Course or scientific<br/>source"] --> study["Study dossier"]
    study --> bundle["Evidence bundle"]
    bundle --> checks["Declared checks<br/>and tolerances"]
    checks --> verdict["Claim verdict"]
    verdict --> resolution["Resolution workflow<br/>when work remains"]
    verdict --> indexes["Freshness, integrity,<br/>coverage, parity indexes"]
    resolution --> indexes
    indexes --> docs["Bounded public claim"]

What A PCM Represents

“PCM” identifies a comparative-methods study family reconstructed from source teaching material. Each family has a scientific progression—prepare a population, fit named models, inspect diagnostics, and interpret bounded quantities—and an evidence progression that asks which of those steps can be reproduced and compared at the current repository revision.

A PCM is therefore neither a runtime module nor a single benchmark. It is a dossier containing several claim-sized bundles. PCM1 and PCM2 currently carry governed numerical comparisons; PCM3 through PCM5 currently preserve source, input, estimand, and closure contracts without claiming missing executions as results. The numbering describes the source study sequence, not increasing software quality or transferable validation.

Follow One Scientific Statement

Consider the statement “the registered PCM1 phylogenetic-signal observation corresponds between the reference analysis and Bijux.” The Evidence Book does not accept that sentence because a signal function exists or because two printed values look close. The review path is:

1. identify the exact PCM1 claim and its admitted 75-species population;
2. resolve the prepared trait table and tree to governed input identities;
3. confirm both executions, implementations, versions, and primary outputs;
4. select the signal observation declared by the claim dependency map;
5. apply the predeclared numerical rule and retain the complete denominator;
6. read the resulting claim verdict with its limitations;
7. confirm that material dependencies remain current at this revision.

That chain supports only the registered observation and population. It does not establish that all signal estimators agree, that the model is biologically adequate, or that every PCM1 figure is equivalent. Every study page follows this same pattern while changing the scientific question and evidence state.

Two Reading Layers

Use these public pages to understand study questions and limitations. Use the checked-in evidence-book/ for primary bundle records, registries, sources, and repository-wide review reports. The narrative explains; the checked-in record adjudicates.

For a first review, follow Reading an evidence bundle from identity through machine observations and freshness. Use Claim verdicts and resolution to distinguish observation, claim, bundle, repository, and follow-through statuses.

The Unit Of Evidence Is A Bounded Claim

A PCM study is a dossier that organizes related scientific questions. It is not a single test and does not receive one transferable “pass” label. Its bundles register claims; claims declare required observations; observations carry exact identities, rules, and outcomes.

flowchart LR
    study["Study dossier<br/>scientific context"] --> bundles["Evidence bundles<br/>comparison modes"]
    bundles --> claims["Bounded claims<br/>population and method"]
    claims --> observations["Required observations<br/>keys and denominators"]
    observations --> rules["Predeclared checks<br/>exact or tolerance"]
    rules --> verdicts["Claim verdicts<br/>with limitations"]

    verdicts -. summarized without promotion .-> study

Level	Owns	Must not be used as
study	source context, scientific progression, and the inventory of questions	one vote that overrides mixed claim states
bundle	one comparison mode and its governed files	proof for quantities outside its registered claims
claim	exact statement, population, method, dependencies, and verdict	a general capability badge
observation	compared scalar, row, topology, node, artifact, or absence	a claim verdict without the declared aggregation rule
resolution	cause, owner, follow-through, and closure evidence	a replacement for the current verdict

This model allows PCM2 to retain matched claim-scoped bundles and unexplained scalar mismatches without contradiction. It also allows PCM3–PCM5 to be complete design dossiers while remaining numerically not_comparable.

At A Glance

• studies: 5
• evidence bundles: 43
• teaching studies: 2
• migration studies: 2
• freshness statuses: current=43
• coverage gaps still tracked: 28
• foundational numerical trust: bounded
• reviewer readiness: bounded
• maturity tier: reviewable_but_incomplete

These values describe the checked-in revision reviewed on 2026-07-22. The freshness report is authoritative when the generated state changes.

Understand The Five-Study Evidence Progression

The study families are related analytical lessons, not five interchangeable votes on one implementation. PCM1 establishes a reusable primate population; PCM2 changes its covariance and model comparisons; PCM3 changes the regression design; PCM4 introduces posterior and phylogenetic random-effect estimation; PCM5 changes response families and derives quantities from posterior draws.

flowchart LR
    pcm1["PCM1<br/>population · signal · ancestors"]
    pcm2["PCM2<br/>GLS · PGLS · modes"]
    pcm3["PCM3<br/>contrasts · interactions"]
    pcm4["PCM4<br/>Bayesian mixed models"]
    pcm5["PCM5<br/>counts · infection · repeated data"]

    pcm1 --> pcm2 --> pcm3
    pcm4 --> pcm5

Study	Scientific contribution	Evidence contribution at this revision
PCM1	shows why type repair, duplicate aggregation, pruning, and tree–trait reconciliation are part of the analysis	governed preparation lineage plus matched or tolerance-matched signal and ancestral observations
PCM2	shows how residual phylogenetic covariance changes regression, evolutionary-mode comparison, and diagnostics	claim-scoped matches coexist with one non-comparable bundle and a broader 12-row unexplained mismatch population
PCM3	makes factor coding, transformations, interactions, and Martins covariance explicit before coefficients are compared	complete comparison design and closure locations; no governed numerical comparison yet
PCM4	separates Gaussian, Bayesian, PGLS, and phylogenetic random-effect estimands	source, population, covariance, prior, chain, and output contract; no governed posterior yet
PCM5	distinguishes Gaussian, count, infection, and repeated-measure models and their draw-wise derived quantities	model-family and derivation contract; no governed generalized mixed-model posterior yet

The absence of a favorable PCM3–PCM5 verdict does not erase their value as design records. It does bound that value: they prevent a future comparison from silently changing the population, estimand, scale, or denominator, but they do not supply coefficients, draws, diagnostics, or parity conclusions.

Find The Record From The Statement

Statement you want to make	First record to open	Continue only when
an input table or tree was prepared correctly	preparation claim and its input/result manifests	transformations, exclusions, taxa, checksums, and verdict cover that exact object
a coefficient or likelihood corresponds with a reference	claim-level scalar/row observations and check	model identity, compared field, tolerance, and full eligible denominator are present
a topology or ancestral value corresponds	structurally keyed observation ledger	taxa, rooting, node/clade identity, model, and tree revision align
a posterior quantity corresponds	governed draws or summaries plus chain diagnostics	priors, parameterization, chains, transformations, exclusions, and verdict are governed
a figure is supported	machine-readable analytical rows and rendering provenance	every displayed value resolves to the analytical result; plot-only limits remain explicit
a study conclusion can be cited	exact claim manifest, current indexes, and freshness record	the statement stays inside the claim population, method, observation, and verdict

Begin with the statement, not the most favorable dashboard row. If no record owns the exact quantity and population, the statement is unsupported even when a neighboring bundle is matched.

Resolve A Value Through The Record Graph

flowchart RL
    citation["Paper, report<br/>or release statement"]
    guide["Public study guide<br/>and bounded interpretation"]
    verdict["Claim manifest<br/>verdict and freshness"]
    observation["Machine result<br/>observation and check"]
    execution["Reference and Bijux<br/>execution identity"]
    inputs["Source, datasets<br/>and input manifest"]

    citation --> guide --> verdict --> observation --> execution --> inputs

The trace must remain one-to-one at the claim boundary. A study guide can summarize several bundles, but a cited number must resolve to the observation that owns it, and that observation must resolve to the exact computation and inputs. When any edge is absent or stale, stop before the missing edge.

Current Verdict Ledger

This table counts bundle-manifest verdicts. It does not replace observation-level ledgers stored inside bundles.

Study family	matched	matched_with_tolerance	not_comparable	Interpretation
primate longevity signal	8	1	0	analytical and structural comparisons are governed; numerical tolerances and plot-only limits remain explicit
primate PGLS and signal	3	6	1	most declared comparisons match; one bounded reference surface remains non-comparable
primate continuous and categorical regression	0	0	8	records and closure locations exist, but lecture-level comparisons are not yet adjudicated
cichlid Bayesian linear and phylogenetic mixed models	0	0	7	MCMCglmm posterior execution is not reproduced by a governed comparison lane
cichlid phylogenetic generalized linear mixed models	0	0	9	generalized mixed-model posterior execution remains external and non-comparable

not_comparable is not a failed numerical tolerance. It means the repository does not currently own a valid comparison under aligned execution, parameterization, and outputs. The bundle remains useful because it identifies the source, data, intended claim, and exact missing boundary.

The complete governed verdict vocabulary is matched, matched_with_tolerance, mismatch_explained, mismatch_unexplained, and not_comparable. Open follow-through and named blockers belong to the separate resolution workflow.

Distinguish The Kinds Of Absence

An absent value or comparison has an owning layer. Preserve that layer instead of translating every gap into not_comparable, failed, or “open.”

Observed absence	Owning record	Defensible conclusion
source or governed input is missing	provenance or input manifest	the proposed claim cannot be reconstructed from the dossier
reference or Bijux execution did not occur	execution record	no observation exists for that computation lane
process ran but required output is missing or malformed	execution and output inventory	this execution failed or is incomplete
both outputs exist but estimands or identities do not align	comparison observation	the proposed quantities are not currently comparable
comparable values violate the registered rule	comparison observation	a numerical or structural mismatch was observed
comparison exists but no claim consumes it	claim dependency record	the observation is not yet evidence for the public statement
claim record exists but a material dependency changed	freshness record	the prior verdict cannot be cited as current

Only the governed bundle assigns the claim verdict. A reader may encounter several absence types in one study; report each one at its owner rather than choosing a single favorable or pessimistic label for the whole dossier.

What The Evidence Supports At This Revision

Study	Safe use	Required qualification
PCM1	cite the governed 75-taxon preparation, tree–trait alignment, signal fit, and named ancestral observations	numerical conclusions are conditional on the prepared primate data, tree, and declared model; figure equivalence is outside the analytical verdict
PCM2	cite matched baseline GLS, Pagel-λ, signal, and other claim-scoped bundles	retain the 12 unexplained scalar mismatches; do not describe the study as uniformly matching
PCM3	inspect the design-matrix and claim decomposition needed for multivariable comparison	all eight claims are not_comparable; no lecture-level coefficient parity is established
PCM4	inspect source, data, model intent, and missing posterior contract	all seven claims are not_comparable; no governed primary posterior output exists
PCM5	inspect generalized mixed-model estimands and closure requirements	all nine claims are not_comparable; no governed primary posterior output exists

This table is a decision aid, not a replacement for the manifests. A paper, notebook, or release statement must cite the exact claim and bundle whose scope contains the assertion.

Observation-Level Mismatch Disclosure

The generated parity dashboard currently retains 12 mismatch_unexplained scalar rows for PCM2 under evidence-001, while that bundle's manifest verdict is matched for its narrower workspace-reload claim. This is not erased by the bundle ledger above. Reviewers making a PCM2-wide statement must inspect both levels and carry the unresolved scalar rows into the conclusion.

The distinction matters: a bundle verdict applies to its registered claim, while a scalar table can aggregate observations with broader scientific debt. The location and scope of those observations still require scrutiny; a matched bundle headline must never be used as proof that every attached scalar comparison matched.

The 12 PCM2 rows are not an abstract warning. They cover four transformed-tree branch-length observations, two early-burst fit observations, one Brownian-versus-early-burst likelihood-ratio observation, four ancestral estimate vectors, and one residual diagnostic. The PCM2 guide names the observed deltas and separates them from the claim-scoped bundle verdicts.

Review A Claim In This Order

1. Find the study and claim identifier.
2. Open the bundle manifest and input manifest.
3. Inspect source provenance and runtime or engine versions.
4. Read the declared checks and acceptance criteria.
5. Inspect machine-readable results before the reviewer summary.
6. Confirm the verdict in the current indexes and freshness report.
7. Carry exclusions, mismatches, the current verdict, and separate resolution state into any downstream statement.

The order matters. A reviewer summary is a projection of structured records; it should not be used to bypass the manifest, check definitions, or observation-level results.

Cite The Record, Not This Summary

Use the public study guide to orient the scientific question. For an auditable claim, cite the bundle manifest and the machine-readable result that contains the observation. If a value appears only in narrative prose, it is not a new evidence source; it must trace to a governed result file.

When a generated index and a bundle disagree, stop at the contradiction. Do not choose the more favorable status. Rebuild or investigate the owning record before using the claim.

Repository-Wide Control Reports

Report	Question it answers
evidence index	which bundles and claims exist?
integrity report	are declared records structurally coherent?
freshness report	do records still match their material dependencies?
parity dashboard	which scalar comparisons match, match within tolerance, or mismatch?
verdict workflows	which mismatches and non-comparable records require explicit follow-through?
coverage gaps	where are claim or workflow proofs incomplete?
false-confidence audit	where could presentation overstate evidence strength?
scientific debt register	which unresolved boundaries require durable follow-through?

What A Bundle Proves

A bundle proves only that its declared checks and acceptance rules produce its recorded verdict for identified inputs, sources, versions, and configuration. It may support a larger claim only when the claim mapping says so. It does not certify untested datasets, models, external-engine versions, or biological interpretations.

Evidence Strength Is Multidimensional

flowchart TB
    provenance["Source provenance"]
    identity["Input identity"]
    execution["Executable comparison"]
    observations["Declared observations<br/>and tolerances"]
    verdict["Claim-scoped verdict"]
    freshness["Current material<br/>dependencies"]

    provenance --> verdict
    identity --> execution --> observations --> verdict
    freshness --> verdict

A bundle can be structurally complete and current while remaining not_comparable. It can also contain passing observations while leaving a larger claim unresolved. Report each axis rather than replacing the record with one unqualified word such as “validated.”

Read Contract-Only Evidence As A Design Record

PCM3, PCM4, and PCM5 include bundles whose present value is design-level rather than numerical adjudication. Read these records by the layer they genuinely own:

Retained layer	What a reviewer gains	Claim still unavailable
source provenance	exact lecture/source program and scientific context	that the source program was executed successfully in the governed lane
governed datasets and input manifests	byte identity, schema, tree/table population and dependencies	that every model consumed the admitted inputs correctly
claim decomposition	stable estimands and one bundle per bounded question	any favorable comparison verdict
reference.R and analysis.py contract wrappers	executable statement of expected ownership, files and absent outputs	reference or Bijux scientific computation
checks and result manifests	structural completeness, freshness and explicit zero-primary-output state	coefficient, posterior, diagnostic or topology agreement
closure contract	exact executions, observations and acceptance rules still required	completion merely because the missing work is well specified

This is durable scientific infrastructure: it prevents later comparison from silently changing the question, inputs, or denominator. It is not numerical evidence. A future execution must create new governed primary outputs and observation records; it must not reinterpret the contract wrapper as an earlier run.

flowchart LR
    design["Governed design<br/>source · inputs · claims"]
    execute["Reference and Bijux<br/>scientific execution"]
    observe["Aligned observations<br/>diagnostics · denominators"]
    adjudicate["Acceptance rules<br/>and verdict"]
    current["Current PCM3–PCM5<br/>contract-only boundary"]

    design --> execute --> observe --> adjudicate
    current -. owns .-> design

The dotted edge locates the present evidence. It does not imply partial credit for the unexecuted comparison stages.

Non-Comparable Evidence Is First-Class

PCM4 and PCM5 remain valuable precisely because their external posterior boundary is visible. Their bundle verdict is not_comparable; the associated follow-through remains open until a governed posterior comparison exists. Use evidence honesty and limits when translating these records into public language.