Validation Tools

This document describes a comprehensive validation framework for a clinical-grade whole genome sequencing (WGS) pipeline, implemented through a synthetic data generation toolkit referred to as the Validation Tools. The system is designed to validate the pipeline end-to-end—from raw sequencing reads (FASTQ) through alignment, variant calling, annotation, and clinical reporting—using controlled, biologically realistic datasets with known ground truth.

The Validation Tools integrate curated pathogenic variants from ClinVar, genomic coordinates from Ensembl GRCh38, and mutation simulation via BAMSurgeon. Each gene within the clinical reporting panel is explicitly tested using representative pathogenic variants to ensure both computational correctness and clinical interpretation accuracy.

The framework goes beyond single nucleotide variants (SNVs) to encompass a multi-class variant validation regime. In addition to SNV insertion via addsnv.py, the toolkit supports small insertion and deletion (indel) simulation via addindel.py, structural variant (SV) simulation via addsv.py, and large deletion events that approximate copy-number loss (CNV-like) scenarios constructed using the same structural variant interface. Each gene defined in the clinical configuration is now exercised across multiple mutation modalities, broadening the analytical validation surface and bringing the framework into closer alignment with CAP and CLIA-style analytical validation expectations.

The validation tools constitute a deliberate movement toward full-spectrum genomic variant validation, recognising that clinically reportable findings span the full continuum from point mutations to large genomic rearrangements. By exercising the pipeline against synthetic datasets representative of each major variant class, the framework now provides materially broader analytical validation coverage while preserving the deterministic, reproducible character that underpins its evidentiary use within a regulated clinical environment.

2.1 Primary Goal

The primary objective is to establish a robust, reproducible, and clinically defensible validation framework for a WGS pipeline by ensuring:

• Accurate detection of known pathogenic variants
• Correct classification of variants (e.g., pathogenic, likely pathogenic)
• Accurate mapping of variants to genes and diseases
• Correct downstream clinical interpretation and reporting

2.2 Validation Scope Defined by reference.json

The validation targets are explicitly derived from the clinical configuration file (reference.json; see Appendix A), which defines:

• Disease panels
• Gene sets per disease
• Classification mapping rules
• Clinical output labels

2.3 Expected Clinical Behaviour

The pipeline is expected to classify variants according to defined rules:

Additionally, variant consequences such as frameshift_variant, stop_gained, and splice_acceptor_variant must be interpreted as deleterious events contributing to clinical risk.

Figure 1 — Mapping of ClinVar significance to clinical reporting outputs.

2.4 Expanded Variant Class Validation

The current implementation of the Validation Tools extends analytical validation beyond the single nucleotide variant (SNV) class to encompass small insertions and deletions, structural variants, and large copy-number-loss events. This expansion is motivated by the observation that a clinical reporting pipeline must demonstrate fidelity across the full spectrum of reportable variant classes; restricting validation to SNVs alone leaves substantive portions of the analytical surface unverified. The framework now exercises each gene defined in reference.json using a coordinated battery of simulated variant modalities, summarised below.

The implication of this expansion is twofold. First, each gene is now tested across multiple mutation modalities rather than against a single representative SNV, yielding a more comprehensive interrogation of pipeline behaviour at the locus level. Second, the framework is no longer limited to SNV detection; the analytical scope formally encompasses the indel, structural and CNV-like axes that together define the clinical variant taxonomy.

Representative Gene Targets

3.1 Overview

The Validation Tools implement a deterministic validation workflow incorporating a multi-class mutation simulation stage:

Figure 2 — Deterministic validation workflow from reference configuration through multi-class mutation simulation to output validation.

The multi-class mutation simulation stage now subsumes four distinct simulation pathways executed against the extracted regional BAM: SNV insertion via addsnv.py; indel insertion via addindel.py; structural deletion simulation via addsv.py; and CNV-scale deletion simulation, also realised through addsv.py using interval lengths representative of copy-number loss. The downstream FASTQ generation, pipeline execution, and output validation stages are invoked once per simulated dataset, producing a per-class validation record for each gene.

3.2 Core Components

Figure 3 — Core components of the Validation Tools and their interactions.

3.2 Core Components (continued)

3.2.1 Reference Configuration (reference.json)

Defines:

• Gene panels
• Disease associations
• Classification mappings
• Reporting logic

This file serves as the ground truth specification for validation.

3.2.2 ClinVar Dataset

• GRCh38-aligned VCF
• Contains clinically curated variants
• Includes fields: CLNSIG (clinical significance), GENEINFO (gene mapping)

Used to extract pathogenic and likely pathogenic variants.

3.2.3 Ensembl GRCh38 Reference Genome

Provides genomic coordinates and DNA sequence context, required for BAM realignment and variant insertion.

3.2.4 BAMSurgeon

A mutation simulation framework that modifies sequencing reads directly. Key tools:

• addsnv.py — single nucleotide variants
• addindel.py — small insertions/deletions
• addsv.py — structural variants and CNV-like large deletions

Produces modified BAM files containing realistically altered sequencing reads.

3.2.5 Picard (BAM Integrity Layer)

The validation system incorporates the Picard toolkit (Broad Institute) as a mandatory dependency within the BAMSurgeon processing layer. BAMSurgeon does not operate independently; it delegates critical BAM handling tasks to Picard during every mutation workflow, and the absence of Picard at runtime constitutes a hard failure mode of the pipeline rather than a degraded one.

Functionally, Picard is responsible for reconstructing BAM files following synthetic mutation insertion, for enforcing correct read pairing and flag consistency across the modified alignment, for maintaining strict compliance with the BAM format specification, and for preventing downstream failures during FASTQ regeneration and subsequent re-alignment. Each of these responsibilities is invoked transparently by BAMSurgeon during the SNV, indel and structural variant injection pathways described in Section 4A.

3.2.6 Validation Toolchain Requirements

The end-to-end validation pipeline depends on a defined set of core tools, each of which is required for correct execution. BAMSurgeon performs the injection of synthetic SNVs, indels, structural variants and CNV-scale deletions. Picard (Broad Institute) is mandatory for BAM reconstruction and validation during mutation injection and is invoked internally by BAMSurgeon. SAMtools provides regional extraction, sorting, indexing and FASTQ conversion. BWA ensures alignment compatibility and read structure handling for both the regional substrate and the regenerated synthetic reads. Python 3 orchestrates the pipeline and generates the per-gene variant plans that drive each simulation pathway.

Dependency presence is verified prior to execution. Missing components — most notably Picard — trigger immediate termination with explicit error logging, which prevents the silent generation of incomplete or structurally invalid validation datasets and preserves the reproducibility guarantees on which the framework depends.

4.1 Gene Extraction

All genes defined in reference.json (see Appendix A) are extracted to form the validation set. This ensures complete coverage of the clinical reporting panel.

4.2 Variant Selection

For each gene:

• Query ClinVar VCF
• Filter variants where CLNSIG contains "Pathogenic" or "Likely_pathogenic"
• Match gene symbol via GENEINFO
• Select representative variants
• Generate four mutation scenarios per gene: SNV, Indel, SV deletion, CNV-like deletion

4.3 Regional Extraction

To optimise file size and performance, a ±2 Mb region surrounding the variant is extracted:

samtools view -b input.bam chr:start-end > region.bam

This preserves local sequencing context and realistic coverage. Resulting FASTQ sizes are approximately 20–50 MB per gene per variant class. See Appendix B for the full inventory of command-line tools used.

4.4 Variant Injection

Variants are introduced using the appropriate BAMSurgeon utility for each class (see Section 4A for full detail):

addsnv.py    -v snv.txt    -f region.bam -r reference.fa --picardjar /opt/picard/picard.jar
addindel.py  -v indel.txt  -f region.bam -r reference.fa --picardjar /opt/picard/picard.jar
addsv.py     -v sv.txt     -f region.bam -r reference.fa --picardjar /opt/picard/picard.jar

Every BAMSurgeon mutation process is executed with an explicit Picard dependency: the Picard JAR is supplied at runtime via --picardjar, and during execution Picard performs the BAM rewriting that follows mutation insertion, the read-level validation and consistency checks required to preserve format integrity, and the cleanup of intermediate artefacts produced by the injection step. Without this Picard-mediated rebuild, mutation generation will fail and no FASTQ outputs will be produced.

Each injection process modifies sequencing reads, preserves base quality distributions, and maintains realistic sequencing artefacts characteristic of genuine instrument output.

4.5 FASTQ Regeneration

Following mutation injection, BAMSurgeon invokes Picard internally to rebuild and validate the mutated BAM, ensuring that read pairing, flags and header metadata remain compliant with the BAM specification. The validated BAM is then converted into paired FASTQ files (R1/R2) for downstream re-entry into the clinical pipeline:

samtools fastq -1 R1.fastq.gz -2 R2.fastq.gz

Each gene now produces multiple FASTQ datasets, each representing a distinct variant class (SNV, indel, structural deletion, and CNV-like deletion). The resulting synthetic sequencing reads are indistinguishable from real input data and contain the corresponding known pathogenic perturbation for the class under test.

4.6 Naming Convention

Each FASTQ dataset is self-describing and now encodes the mutation type explicitly:

GENE__TESTTYPE__VARIANTID__COORDINATES__EXPECTED-Positive_HighRisk_R1.fastq.gz

Where TESTTYPE is drawn from the controlled vocabulary:

• SNV — single nucleotide variant arm
• INDEL_DEL — small synthetic deletion arm
• SV_DEL — 10 kb structural deletion arm
• CNV_LOSS — 100 kb CNV-like deletion arm

4.7 Truth Set Generation

A structured truth file is generated, with the schema extended to support multi-class validation:

Gene | TestType | Chrom | Start | End | Ref | Alt | ClinSig | ExpectedConsequence | ExpectedStatus | RiskLabel

The expanded truth set supports multi-variant per gene validation, variant-class-specific validation, and downstream performance metric calculation, while remaining suitable for automated validation, output comparison, and audit traceability.

This section formalises the four simulation pathways constituting the multi-class variant arm of the Validation Tools. Each pathway shares the same regional extraction substrate and FASTQ regeneration tail, differing only in the BAMSurgeon utility invoked, the variant specification supplied, and the coordinate footprint of the injected lesion.

4A.1 SNV Simulation

Single nucleotide variants are derived directly from ClinVar pathogenic and likely pathogenic entries and inserted using BAMSurgeon addsnv.py. Because the injected substitution is drawn from a curated clinical resource, the resulting synthetic dataset retains demonstrable clinical relevance; the validation outcome therefore tests not only computational detection but also the downstream interpretive behaviour expected of a clinically reportable variant.

4A.2 Indel Simulation

Synthetic 1 bp deletions are introduced near each gene locus and generated using BAMSurgeon addindel.py. These constructs are designed to mimic frameshift variants and loss-of-function mutations, and are therefore well-suited to interrogating the pipeline's handling of consequence classes—frameshift_variant, stop_gained, and analogous truncating effects—that drive a significant fraction of clinically actionable findings.

4A.3 Structural Variant Simulation

Structural variants are introduced as 10 kb deletions centred around the gene region, generated using BAMSurgeon addsv.py. The intended scope of this arm is the validation of large genomic disruption detection and breakpoint handling within the pipeline, exercising the alignment, soft-clipping, and split-read interpretation behaviours that distinguish structural variant calling from small variant calling.

4A.4 CNV Simulation (CNV-like Loss)

Copy-number loss is approximated by 100 kb deletions introduced as a proxy for copy-number-loss events, again generated using BAMSurgeon addsv.py with an extended interval. These constructs support end-to-end validation of the CNV-loss reporting pathway at the structural level.

Per-Gene Simulation Matrix

5.1 End-to-End Testing

Each generated FASTQ is processed through the full pipeline:

FASTQ → Alignment → Variant Calling → Annotation → Clinical Interpretation → Report

Validation checks: variant detection · annotation accuracy · classification correctness · report output alignment.

5.2 Failure Conditions

• False negatives (missed variants)
• Misclassification (e.g., VUS instead of pathogenic)
• Incorrect gene mapping
• Annotation errors
• Reporting logic failures

5.3 Variant Class Coverage

Validation now explicitly tests detection of single nucleotide variants, detection of small indels, detection of structural deletions, and detection of large-scale genomic loss events. The corresponding truth set entries (see Section 4.7) carry an explicit TestType field, allowing per-class accuracy, sensitivity, and specificity to be aggregated independently and reported in the validation summary.

The practical consequence of this expansion is a significant improvement in analytical sensitivity testing across the clinical variant taxonomy, increased variant-class robustness through systematic per-class exercise, and enhanced pipeline completeness validation insofar as defects confined to a single variant class can no longer escape detection by virtue of being absent from the test corpus.

6.1 Strengths

A foremost strength of the Validation Tools is the achievement of fully automated validation, in which the entire chain from configuration parsing through synthetic dataset generation, pipeline execution, and outcome reconciliation proceeds without manual intervention. By eliminating discretionary human action from the validation loop, the framework removes a substantial source of operator-induced variability, supports continuous integration practices, and renders periodic regression testing operationally tractable at the cadence required for clinical deployment.

The framework is further distinguished by the use of clinically grounded test cases, derived directly from curated ClinVar entries with established pathogenic or likely pathogenic significance. Because each test case originates from a variant of demonstrable clinical relevance, the validation exercise is not merely a synthetic stress test of computational correctness; it constitutes a faithful interrogation of the diagnostic behaviours that determine patient-facing outcomes, thereby aligning verification activity with the regulatory and ethical expectations of clinical genomics.

The reproducibility of the generated datasets represents an additional and material advantage. Because every synthetic FASTQ is fully described by its self-encoding filename, its truth-set entry, and the deterministic invocation of BAMSurgeon against fixed reference inputs, identical datasets can be regenerated on demand for audit, retrospective investigation, or third-party review. This deterministic property satisfies the evidentiary standards expected of a clinical-grade software lifecycle and provides a defensible substrate for accreditation.

By construction, the framework supports end-to-end pipeline verification, exercising every stage from raw read input through alignment, variant calling, annotation, and final clinical interpretation. This holistic scope ensures that defects arising at the boundaries between components—precisely the integration faults that unit-level testing tends to overlook—are surfaced under realistic operating conditions, providing assurance that the system performs correctly as a whole rather than only in its constituent parts.

The use of realistic sequencing simulation via BAMSurgeon, applied to authentic background BAM files, preserves the empirical distributions of base quality, coverage depth, mapping artefact, and strand bias that characterise genuine instrument output. Critically, these distributions are preserved across the full multi-class injection regime—SNV, indel, structural and CNV-like deletion arms alike—ensuring that the validation environment faithfully reproduces the analytical conditions encountered in routine clinical operation regardless of the variant class under test.

The framework now affords multi-class variant validation per gene, exercising each clinical locus against SNV, indel, structural and CNV-like perturbations within a single coherent execution, and thereby achieving substantially closer alignment with CAP and CLIA-style analytical validation expectations than was possible under the SNV-only regime that preceded it.

6.2 Limitations

CNV modelling does not yet simulate read-depth variation. The current CNV-loss arm uses a structural deletion to reproduce breakpoint and discordant-read evidence, while depth-of-coverage perturbation is planned for a later iteration.

Structural variants are presently limited to deletions. Deletions cover the most prevalent class of reportable structural events; insertions, inversions, and translocations are scheduled for inclusion in a future release.

Duplication and amplification modelling is not yet included. Copy-number gain events will be added alongside the planned read-depth simulation work to provide bidirectional CNV coverage.

Mosaic and low variant allele fraction (VAF) scenarios are not yet included. Injected variants are introduced at constitutional heterozygous or homozygous allele fractions; titrated low-VAF injection is a planned extension.

6.3 Future Enhancements

A natural extension of the current framework is the implementation of multi-variant per gene validation, in which several representative variants per gene—spanning distinct functional consequences, exonic positions, and inheritance contexts—are exercised in concert. Such an enhancement would yield finer-grained coverage statistics, expose locus-specific behaviours that single-variant testing cannot reveal, and strengthen the statistical basis for asserting per-gene diagnostic competence.

The introduction of biallelic variant modelling would extend the framework to the validation of recessive and compound heterozygous disease mechanisms, which require coordinated injection of paired alleles at compatible coordinates. This capability is essential for genes such as MUTYH, where biallelic genotypes drive the reportable phenotype, and would enable the framework to address inheritance modes that are currently under-represented in the validation matrix.

Expanded structural variant integration represents a substantive next phase, in which duplications, inversions, and translocations are systematically constructed alongside the existing deletion arm, injected via addsv.py, and reconciled against breakpoint-aware truth sets. Achieving this will require investment in SV-specific comparators, expanded reference data, and consensus rules for partial overlap, but will materially extend validated coverage to clinically significant structural classes presently outside the framework's evidentiary perimeter.

The addition of read-depth-aware copy number variation simulation would close the most prominent remaining gap in the validation matrix by enabling controlled perturbation of coverage across defined intervals, complementing the structural CNV-loss proxy already implemented. Integration with depth-of-coverage callers, paired with truth sets describing both the perturbed regions and their expected reporting outcomes, would permit the same end-to-end discipline currently applied to SNVs and indels to be extended to fully signature-faithful CNV interpretation, thereby unifying the validation regime across variant classes.

The addition of performance benchmarking would augment the framework's correctness-oriented checks with quantitative measurement of throughput, latency, and resource consumption across pipeline stages and across the four variant-class arms. Coupling such metrics to the existing truth-set apparatus would enable simultaneous evaluation of analytical accuracy and operational efficiency, supporting capacity planning, regression detection at the performance boundary, and informed prioritisation of optimisation effort within the clinical service environment.

The expanded multi-class validation regime brings the Validation Tools into materially closer alignment with the analytical validation expectations articulated by the College of American Pathologists (CAP) and codified within the Clinical Laboratory Improvement Amendments (CLIA). Both frameworks require that a clinical assay demonstrate analytical performance across the full range of variant classes that the assay is intended to report, and that such performance be substantiated by controlled, reproducible test data with established ground truth.

The framework, as presently constituted, supports the foundational requirements of these standards in three respects. First, it provides multi-variant class validation, exercising SNV, indel, structural and CNV-like axes in a single coordinated regime, which is a stated requirement of analytical validation under both CAP and CLIA. Second, it provides end-to-end pipeline verification from FASTQ ingestion through clinical reporting, satisfying the expectation that validation address the assay as deployed rather than its constituent computational steps in isolation. Third, it provides controlled synthetic truth datasets that are deterministic, reproducible, and fully described by their accompanying truth sets, satisfying the evidentiary expectations associated with a documented validation lifecycle.

In addition to the synthetic, multi-class analytical validation described in the preceding sections, the pipeline has additionally undergone an extensive programme of real sample validation against well-characterised clinical specimens. This activity directly addresses the expectation, articulated under both CAP and CLIA, that synthetic ground-truth datasets be complemented by empirical evidence drawn from authentic biological material processed under routine laboratory conditions. Real-sample testing has been executed across the full clinical reporting panel, exercising each gene against patient-derived sequencing data and confirming that the analytical behaviour observed in the synthetic regime is preserved when the pipeline is presented with genuine instrument output.

The outcomes of this real-sample programme have been independently validated against sequencing.com, whose orthogonal interpretation of the same underlying sequencing data provides a cross-platform reference against which our pipeline calls have been systematically compared. This orthogonal confirmation step discharges the cross-method concordance expectations of clinical accreditation by demonstrating that variant calls produced by the in-house pipeline are reproducible on an independent analytical platform, and that any discordances are accountable, characterised, and within the analytical performance envelope established during synthetic validation.

With respect to quantitative performance metrics—per-class sensitivity, specificity, positive and negative predictive value, and the associated confidence intervals expected of a formal accreditation submission—a dedicated successor project is currently in flight, with a companion white paper scheduled for publication in June 2026. That work will document the substantial advances now being made in real-time quantitative performance reporting, which are surfaced and live-updated through our METRICS control panel (Monitoring Engine for Throughput, Runtime, Integrity & Clinical Systems). METRICS provides continuous, per-class instrumentation of pipeline performance against the validation truth sets and against orthogonal reference data, replacing periodic retrospective reporting with a live, auditable evidence stream suitable for ongoing accreditation oversight.

The Validation Tools provide a rigorous, scalable, and clinically aligned validation framework for whole genome sequencing pipelines. By integrating curated clinical variant datasets with synthetic read generation, the system ensures both technical accuracy and clinical reliability.

The framework has evolved from an initial regime of SNV-only validation into a multi-class genomic validation system capable of testing SNVs, indels, structural variants, and CNV-like events. This transition represents a major step toward clinical-grade validation: the system now validates a substantially broader region of the biological space relevant to clinical reporting, and it does so under deterministic, reproducible conditions amenable to audit and accreditation review.

This framework therefore establishes a practical standard for genomic pipeline validation, enabling confident deployment in clinical and research environments while providing a coherent foundation for the further enhancements identified in Section 6.3 and the additional regulatory activities outlined in Section 7.

1. Ewing, A. D. et al. BAMSurgeon: Synthetic mutation insertion into sequencing data.
2. Landrum, M. J. et al. (2018). ClinVar: improving access to variant interpretations. Nucleic Acids Research, 46(D1), D1062–D1067.
3. Yates, A. D. et al. (2020). Ensembl 2020. Nucleic Acids Research, 48(D1), D682–D688.
4. Li, H. et al. (2009). SAMtools. Bioinformatics, 25(16), 2078–2079.
5. Li, H. (2013). BWA-MEM alignment algorithm.
6. College of American Pathologists. Laboratory Accreditation Program — Molecular Pathology and Next-Generation Sequencing Checklists.
7. Centers for Medicare & Medicaid Services. Clinical Laboratory Improvement Amendments (CLIA) — Analytical Validation Requirements (42 CFR §493).

The clinical configuration artefact reference.json defines the panel scope, classification mappings, gene rules, disease definitions, and reporting logic referenced throughout this document. The full document is 1,841 lines (~78 KB) and is distributed as an accompanying machine-readable file.

A.1 Top-Level Structure

A.2 Representative Excerpt

{
  "category": "Cancer Predisposition",
  "panel_settings": {
    "default_inheritance_mode": "autosomal_dominant_or_context_dependent",
    "default_noncoding_exact_clinvar_policy": "manual_review",
    "preferred_transcript_policy": "ensembl_canonical_then_primary_protein_coding",
    "reporting_assembly": "GRCh38",
    "species": "homo_sapiens"
  },
  "classification_display": {
    "pathogenic": {
      "headline": "Clinically Significant Mutation Identified",
      "badge": "Positive",
      "severity": "high",
      "table_label": "High Risk"
    },
    "likely_pathogenic": {
      "headline": "Likely Clinically Significant Mutation Identified",
      "badge": "Positive",
      "severity": "high",
      "table_label": "Increased Risk"
    }
    /* ... vus, carrier, review, structural, negative ... */
  }
  /* ... report_sections, consequence_phrases, gene_rules,
     disease_definitions, gene_report_metadata, unique_genes ... */
}

Listing A.1 — Abridged excerpt of reference.json; refer to the distributed file for the complete specification.

The following command-line utilities constitute the operational toolchain invoked by the Validation Tools across the stages described in Sections 3, 4 and 4A. Versions are pinned in the pipeline manifest; representative invocations are reproduced below.

B.1 Representative Invocations

B.1.1 ClinVar Variant Selection

bcftools view -i 'INFO/CLNSIG ~ "Pathogenic" || INFO/CLNSIG ~ "Likely_pathogenic"' \
  clinvar.GRCh38.vcf.gz \
  | bcftools query -f '%CHROM\t%POS\t%REF\t%ALT\t%INFO/CLNSIG\t%INFO/GENEINFO\n'

B.1.2 Regional Extraction (±2 Mb)

samtools view -b input.bam chr17:43044295-47044295 > region.bam
samtools index region.bam

B.1.3 Multi-Class Variant Injection (BAMSurgeon)

# SNV arm
addsnv.py    -v snv.txt   -f region.bam -r GRCh38.fa -o injected.snv.bam   --picardjar picard.jar
# Indel arm (1 bp synthetic deletion)
addindel.py  -v indel.txt -f region.bam -r GRCh38.fa -o injected.indel.bam
# SV arm (10 kb deletion)
addsv.py     -v sv.txt    -f region.bam -r GRCh38.fa -o injected.sv.bam
# CNV-like arm (100 kb deletion via addsv.py)
addsv.py     -v cnv.txt   -f region.bam -r GRCh38.fa -o injected.cnv.bam

B.1.4 FASTQ Regeneration

samtools sort -n injected.bam -o sorted.bam
samtools fastq -1 R1.fastq.gz -2 R2.fastq.gz -0 /dev/null -s /dev/null sorted.bam

B.1.5 Re-alignment for End-to-End Validation

bwa mem -t 16 -R "@RG\tID:val\tSM:synthetic\tLB:val\tPL:ILLUMINA" \
  GRCh38.fa R1.fastq.gz R2.fastq.gz | samtools sort -o validated.bam -

Listing B.1 — Representative command invocations across all four variant-class arms; full parameterisation is encoded in the pipeline manifest.