Skip to content

Operating scientific workflows

Core operations begin with an explicit scientific question and end with an artifact whose inputs, policy, exclusions, QC, and limitations can be reviewed. The package can be used directly from Python or the CLI; long-running execution, provider management, checkpoints, and replay belong to Runtime.

flowchart LR
    Q["scientific question"] --> I["inspect and validate inputs"]
    I --> P["declare policy"]
    P --> X["execute Core operation"]
    X --> V["review QC and exclusions"]
    V --> D{"contract satisfied?"}
    D -->|yes| A["retain result and provenance"]
    D -->|no| F["retain refusal or failure evidence"]

Begin by workflow family

Work First route Review before continuing
FASTA and digestion common workflows header interpretation, contaminants, decoys, protease policy, peptide-space caveats
spectra and raw-format inspection CLI and I/O owner guides format support, metadata, centroid/profile assumptions, rejected spectra
search-result import and FDR identification owner and artifact contracts engine provenance, score orientation, decoys, thresholds, audit trail
protein inference and LFQ inference and quantification guides shared peptides, grouping, missingness, normalization, reproducibility
DIA, PTM, targeted, or multiplex workflow-family owner family-specific acceptance, ambiguity, interference, and evidence ceiling
interpretation and review interpretation and review owner source context, computed result versus judgment, downstream authority
benchmark or public claim benchmark catalog and lineage asset provenance, holdouts, acceptance bars, transfer limits

Environment and installation

Install only the extras needed by the selected workflow. Optional scientific engines and file readers must fail explicitly when unavailable rather than changing the analysis silently. Installation and setup documents supported environments and extras; deployment boundaries explains when Runtime is required.

Output discipline

Choose output paths before execution and retain the primary machine-readable artifact, not only console output. For repository work, local run products belong under artifacts/. A complete retained result includes:

  • source and configuration identity;
  • the primary result and rejected-input ledger;
  • QC and diagnostic artifacts;
  • external-engine provenance where applicable;
  • schema and producer identity;
  • refusal or failure evidence when no accepted result was produced.

Failure interpretation

Failure class Meaning Safe response
parse or schema rejection input cannot be interpreted under the declared contract correct or explicitly transform the source; preserve rejection details
scientific refusal evidence or assumptions cannot support the requested operation narrow the question, provide missing evidence, or stop
optional capability unavailable required dependency, engine, or environment is absent install the declared extra or select an explicit alternative
QC or acceptance miss computation completed but result does not meet the scientific bar retain artifacts; do not promote the claim
operational interruption process did not complete use Runtime recovery if checkpoints or replay are required

Failure recovery distinguishes recoverable operational conditions from scientific invalidity. Observability and diagnostics maps logs and artifacts to the owning contract.

Scaling without changing meaning

Chunking, parallelism, caching, and streaming must preserve the serial scientific contract: stable identity, ordering where promised, policy, aggregation, rejections, and QC. A faster path requires equivalence evidence; it cannot use performance pressure to weaken validation or omit provenance.

See performance and scaling for supported routes and security and safety for untrusted files, external tools, and resource boundaries.

Publication boundary

A package release can be operationally sound while a workflow family remains scientifically bounded. Release and versioning connects API and artifact compatibility to distribution changes. Public workflow claims additionally require current benchmark lineage, acceptance, rerun, grounding, decision, and consequence evidence.