MrBayes And BEAST Correspondence

MrBayes and BEAST own their posterior computations. Bijux owns explicit analysis preparation, governed execution, native artifact parsing, posterior diagnostics, tree-set review, and evidence packaging. The record always says whether an analysis was only prepared, parsed from existing outputs, or executed by the current workflow.

A Posterior Claim Begins With Its Sample Population

Required identity	Review requirement
analytical target	likelihood, priors, clock or tree prior, topology treatment, and monitored parameters
execution population	independent chains, resumed segments, seeds, generations or states, and terminal status
selected posterior	burn-in, thinning or subsampling, failed-chain policy, and retained draw/tree identities
diagnostic population	exact series, estimator, threshold, and chains contributing to ESS, drift, or convergence
reported quantity	parameter or clade key, transformation, estimator, interval rule, units, and retained denominator

A posterior mean without this lineage is not portable evidence. Similar means can arise from different priors, chains, burn-in windows, transformations, or failed-sample policies. Compare values only after their sample populations and estimands align.

flowchart LR
    inputs["Alignment · partitions<br/>tree · dates · calibrations"]
    prepare["NEXUS or XML<br/>validation"]
    engine["MrBayes or BEAST<br/>external posterior"]
    native["Trace · diagnostics<br/>posterior trees"]
    select["Burn-in · subsampling<br/>chain/state identity"]
    assess["ESS · drift · convergence<br/>clade/topology sensitivity"]
    evidence["Methods · tables<br/>checksummed evidence package"]

    inputs --> prepare --> engine --> native --> select --> assess --> evidence

Preparation, Execution, And Review Are Separate

State	MrBayes	BEAST	Supported statement
prepared	mrbayes-prepare writes NEXUS with model, rates, chains, sampling, and burn-in policy	beast-prepare writes XML with substitution model, clock, tree prior, dates, calibrations, chain and logging	configuration is structurally represented; no posterior exists yet
inspected	prepared NEXUS and declared inputs are reviewable	beast-xml, calibration, and tip-date commands validate the time-tree contract	the supplied configuration passes supported checks
executed	mrbayes-run invokes mb and captures native outputs	beast-run invokes beast with seed, threads, and overwrite policy	the named executable ran under the retained identity
parsed	trace, MCMC, consensus, ESS, parameter and tree commands read supplied artifacts	log, parameter, tree, consensus and diversity commands read supplied artifacts	observations come from those artifacts, not necessarily a current run
packaged	report and Bayesian evidence commands join selected records	methods, diagnostics table, and evidence package join selected records	the package is internally inventoried; scientific adequacy remains reviewable

The report must distinguish a prepared-only analysis, a parsed prior run, and an engine execution performed by the current workflow. File presence alone is not execution proof.

MrBayes Correspondence

Preparation accepts aligned nucleotide or protein FASTA, optional partitions, model and rate settings, generations, chain count, sampling/printing frequency, and burn-in fraction. Execution retains the mb executable identity, version, command, process state, timeout, native outputs, streams, manifest, and checksums.

Review commands expose distinct posterior questions:

• mrbayes-traces, mrbayes-mcmc, and mrbayes-ess parse parameter and MCMC records;
• mrbayes-parameters, mrbayes-burnin-sensitivity, and mrbayes-convergence evaluate retained trace behavior;
• mrbayes-trees, mrbayes-subsample, mrbayes-summarize, and mrbayes-consensus preserve generation identity while reviewing topology;
• mrbayes-report joins the selected posterior trees and traces into a reviewer-facing report.

BEAST Correspondence

Preparation accepts an alignment plus optional starting tree, fossil calibrations, and tip dates, with explicit clock model, tree prior, chain length, and logging interval. XML inspection retains taxon, character, calibration, tip-date, chain, and logger counts plus validation issues.

Execution records the beast executable, version, threads, seed, overwrite policy, command, timeout, streams, outputs, manifest, and checksums. Review is split by ownership:

• calibration and tip-date commands verify identifiers and dating inputs;
• beast-log, beast-parameters, beast-burnin-sensitivity, and beast-convergence retain state-indexed trace behavior;
• beast-trees, beast-subsample, beast-consensus, and beast-diversity retain state identity while reviewing the posterior tree set;
• Bayesian methods and diagnostics-table commands produce publication-facing records from explicitly selected inputs.

Run Identity And Resume Safety

Identity component	MrBayes	BEAST
prepared analysis	NEXUS checksum, taxa, partitions, model/rates, generations, chains, sampling and burn-in	XML checksum, taxa, substitution model, clock, tree prior, dates/calibrations, chain and logging
execution	mb path/version, command, streams, exit and timeout	beast path/version, command, threads, seed, overwrite, streams, exit and timeout
posterior	generation-indexed traces and trees, consensus and native sidecars	state-indexed log and trees, consensus/MCC and native sidecars
selection	burn-in and subsampling against retained generations	burn-in and subsampling against retained states
assessment	parameter ESS, trace drift, chain and clade behavior	parameter ESS, trace drift, topology, calibration and time-tree behavior

mrbayes-run --resume and beast-run --resume reuse only a verified completed manifest whose executable, inputs, configuration, and output hashes still agree. Timeout markers, killed processes, malformed posterior files, or changed identity require rejection or an explicit clean restart.

Resume therefore preserves execution lineage; it does not create a new independent chain and must not increase the denominator in a convergence claim.

Run And Review

bijux-phylogenetics adapter mrbayes-prepare \
  dataset/alignment.fasta \
  --model gtr \
  --rates gamma \
  --ngen 1000000 \
  --out artifacts/bayesian/analysis.nex

bijux-phylogenetics adapter mrbayes-run \
  artifacts/bayesian/analysis.nex \
  --timeout-seconds 7200

bijux-phylogenetics adapter beast-prepare \
  dataset/alignment.fasta \
  --tree dataset/starting-tree.nwk \
  --calibrations dataset/calibrations.tsv \
  --clock-model strict \
  --tree-prior yule \
  --out artifacts/bayesian/analysis.xml

bijux-phylogenetics adapter beast-xml \
  artifacts/bayesian/analysis.xml \
  --json

bijux-phylogenetics adapter beast-run \
  artifacts/bayesian/analysis.xml \
  --threads 4 \
  --seed 11 \
  --timeout-seconds 7200

Parser and reporting commands can review governed existing outputs without claiming that the current machine reran the posterior analysis.

Diagnostic Interpretation

Observation	Supports	Does not establish
adequate ESS under a named estimator and threshold	retained draws contain enough effective information for that parameter rule	stationarity, model adequacy, or independent-chain agreement
stable trace after selected burn-in	the named window/drift rule did not detect instability	correct burn-in for every parameter or topology
similar independent-chain summaries	selected estimands agree under the comparison rule	convergence in unobserved dimensions or correctness of the target
stable clade frequencies	selected topology summaries are insensitive to the named comparison	identical branch lengths, dates, parameters, or biological history
parser acceptance	native files satisfy the supported structural grammar	current execution, adequate sampling, or a valid biological model

A posterior comparison must align priors, likelihood, clock and tree model, parameterization, chain/state selection, transformations, and estimands before values are adjudicated. Otherwise the posteriors remain separate governed observations.

Preserve Posterior Sample Lineage

Every reported posterior quantity must resolve to the exact retained samples that contributed to it. A summary detached from chain and state identity is not repairable from its mean, interval, or consensus tree.

reported posterior quantity
  = engine run
  + chain or state series
  + burn-in rule
  + thinning or subsampling rule
  + transformation
  + estimand

Projection	Required parent record	Completeness condition
parameter summary	native trace, chain/state column, selected window, transform, and missing/non-finite handling	retained and excluded sample counts reconcile
ESS or drift diagnostic	exact series, estimator/rule, threshold, and selected window	diagnostic denominator matches the reported series
clade frequency or consensus	posterior tree set, generation/state identity, burn-in and subsampling	accepted trees and discarded trees are both counted
independent-chain comparison	distinct run identities, aligned estimand, and per-chain diagnostics	resumed segments are not miscounted as independent chains
evidence package result	all parent records plus inventory and checksums	every table or figure resolves to its selected native inputs

Changing burn-in, subsampling, topology filtering, or parameter transformation creates a new projection. Preserve it alongside its rule; do not overwrite the prior summary and keep the old interpretation.

Admit Samples Before Summarizing Them

Posterior review has an admission boundary before any pooled estimate. For each requested chain or run, retain whether it started, completed, parsed, passed the declared diagnostic policy, and contributed samples. Then retain requested, emitted, excluded, and admitted draw or tree counts for every contributing series.

An excluded chain remains in the execution population with its exclusion reason; it is not erased from the study design. A resumed segment remains part of its parent chain unless the engine contract establishes independent initialization. The pooled posterior denominator is therefore the admitted sample population, while coverage and convergence statements still disclose all requested chains and their terminal states.

Evidence Package Acceptance

adapter bayesian-evidence requires explicit input, configuration, tree, log, diagnostic, and report paths. The resulting package inventories and checksums the selected files. Acceptance means the package is complete under that inventory; it does not upgrade diagnostics into convergence or correspondence into posterior parity.

Deliberate Non-Claims

• Native run_bayesian_inference(...) has its own supported DNA contracts and does not dispatch through MrBayes or BEAST.
• Similar posterior means do not establish parity unless priors, likelihood, clocks, proposals, burn-in, and compared estimands are aligned.
• ESS and convergence thresholds qualify sampling behavior; they do not prove model adequacy or biological correctness.
• PCM4 and PCM5 retain an MCMCglmm external-posterior boundary that these adapters do not erase.