Artifacts And Reproduction

A parity run is reviewable when its selection, environment, observations, and failures survive the terminal session. Bijux writes those records explicitly; the JSON printed to standard output is useful for automation but is not the whole reproducibility package.

Reproduce A Checked Fixture

The checked-fixture lane requires no R installation:

mkdir -p artifacts/parity
bijux-phylogenetics parity \
  --reference-source checked-fixture \
  --summary-out artifacts/parity/checked-summary.tsv \
  --observations-out artifacts/parity/checked-observations.tsv \
  --manifest artifacts/parity/checked-manifest.json \
  --json

Use repeated --reference-method options to select governed method shards. Add --extended only when the larger posterior tree-set fixtures are intended; recording that selection matters because the result set changes.

Reproduce A Live R Case

Live execution is opt-in and tool-specific:

mkdir -p artifacts/parity/geiger-failures
RUN_LIVE_PARITY=1 bijux-phylogenetics parity \
  --reference-source geiger-live \
  --geiger-shard fit_continuous_bm \
  --summary-out artifacts/parity/geiger-summary.tsv \
  --observations-out artifacts/parity/geiger-observations.tsv \
  --geiger-failure-root artifacts/parity/geiger-failures \
  --manifest artifacts/parity/geiger-manifest.json \
  --json

APE and phytools use the corresponding --ape-* and --phytools-* selection, executable, and failure-root options. A live label is never retained if the command silently falls back to frozen data.

Read The Artifact Set

Artifact	Review use
summary table	one row per case with reference source, selection, runtime status, and counts
observation table	exact values, differences, tolerances, and per-observation outcomes
reproducibility manifest	command context, selected lane, environment and linked outputs
failure root	case inputs, external output, diagnostics, and exception context needed to reproduce a mismatch or skip
GEIGER review reports	optimizer triage, boundary warnings, likelihood policy, model confidence, parameterization registry, and generated human/machine reports

Preserve mismatches, skips, warnings, and incomplete outputs. Removing them from an artifact directory changes the evidentiary meaning of the run even if the passing summary rows remain.

Verify The Artifact Set As One Record

After execution, check these relationships rather than opening files in isolation:

1. The manifest names the requested reference source and the exact selected cases or shards.
2. Summary counts reconcile to selected cases and agree with observation statuses.
3. Every required observation identifies its case, field, values, rule, and outcome.
4. Each failed or skipped case points to enough captured context to diagnose execution or comparison failure.
5. Tool and package versions match the environment that produced the files.
6. No downstream report silently drops a non-pass or changes its status name.

Hash or archive the complete directory when it crosses an environment or review boundary. Copying only the passing table destroys the denominator and prevents later reviewers from distinguishing absence from success.

Reconstruct The Denominator

Review registry identity and selection before reading success counts:

flowchart LR
    registered["Registered cases"]
    selected["Selected shards<br/>and cases"]
    executed["Executed cases"]
    passed["Passed"]
    failed["Failed"]
    skipped["Skipped"]

    registered --> selected
    selected --> executed
    selected --> skipped
    executed --> passed
    executed --> failed

The manifest and summary must make each transition countable. Failure roots explain non-passing cases; they must not be removed before an aggregate is reported.

External Engine Bundles

MAFFT, trimAl, IQ-TREE2, FastTree, MrBayes, and BEAST do not use the R parity command. Their adapters retain the external executable and configuration, native standard output and error, expected engine files, parsed typed results, and workflow-level manifests or reports. Review the raw engine artifacts before the normalized presentation whenever parsing or support semantics are in question.

An engine bundle establishes governed execution. A cross-engine or native comparison must separately name aligned fields and statistical meanings; the presence of two successful run directories is not itself a parity verdict.

Promotion Into The Evidence Book

Parity artifacts become scientific evidence only when a dossier also binds them to stable claim identifiers, source references, governed datasets, explicit checks, freshness records, and a verdict. Promotion does not discard the parity-level details: the Evidence Book should link or retain the exact observations that justify its claim. It also preserves the current evidence verdict and any separate open follow-through or named blocker.

Reproduction Completion

A rerun is complete when it reconstructs selection, execution identity, observations, statuses, and failure records—not merely when the command exits successfully. Differences caused by a newer external package or executable are new observations. Preserve both revisions and adjudicate the change instead of overwriting the older run.