Operator Rerun Journey

A reproducible rerun begins with a governed benchmark manifest and ends with a new run bundle whose identity, environment, state history, artifacts, and comparison result can be inspected independently. Reusing the same family name or obtaining similar summary values is not enough.

Select The Exact Lane

Open Workflow Families and record the family’s current trust status and primary run mode. Then use Benchmark Rerun Kits to identify:

• the primary and companion benchmark package roots;
• the primary and companion Runtime entrypoints;
• whether the lane is raw-executable over checked inputs or import-only;
• the scientific and operational conditions that still limit the claim.

For DDA, import_only is a material property: the rerun reopens imported search results and their review lineage. It does not rerun MaxQuant, Comet, Sage, or another search engine from raw spectra.

Preflight The Environment

Use Runtime Environment Contracts to record the supported Python and dependency bounds, operating system, provider, external tool identity, environment variables that affect behavior, and any optional dependency required by the selected lane.

flowchart TD
    manifest["benchmark package manifest"] --> inputs{"all named inputs present?"}
    inputs -->|no| refuse["refuse before execution"]
    inputs -->|yes| environment{"environment contract satisfied?"}
    environment -->|no| refuse
    environment -->|yes| execute["call public Runtime entrypoint"]
    execute --> bundle["run bundle + artifacts + diagnostics"]
    bundle --> compare["compare under declared policy"]
    compare --> verdict{"accepted, degraded, or refused"}

Refuse before execution when an input identity, checksum, required provider, or compatibility condition cannot be established. Substituting a convenient local file creates a different run.

Execute Without Rewriting Evidence

Write local output below artifacts/bijux-proteomics-runtime/; do not overwrite the tracked flagship fixtures. Preserve:

1. source revision and installed distribution versions;
2. benchmark manifest and input checksums;
3. selected public entrypoint and resolved configuration;
4. provider, environment, external tool, and dependency identities;
5. state transitions, retries, checkpoints, warnings, and refusal reasons;
6. artifact inventory, media types, sizes, and checksums;
7. terminal state and comparison-policy identity.

The Runtime Execution Boundary gives the exact manifest, entrypoint, tracked run bundle, and refusal statement for each primary family lane.

Run and serialize the DIA primary lane

The published DIA wrapper resolves its governed package inputs internally and returns a typed run report. From the repository root:

from pathlib import Path

from bijux_proteomics_runtime.workflows.benchmark_runs import (
    run_benchmark_dia_review_path,
)

report = run_benchmark_dia_review_path()
output_dir = Path("artifacts/bijux-proteomics-runtime/dia-library-review")
output_dir.mkdir(parents=True, exist_ok=True)
(output_dir / "run-report.json").write_text(
    report.model_dump_json(indent=2) + "\n",
    encoding="utf-8",
)

This executes the checked report-based DIA transformation. It is raw_executable over the packaged report inputs; it does not process vendor-native raw chromatograms. Run the matrix-shift companion through run_benchmark_dia_generalization_review_path() before making a family-level transfer statement.

Inspect The Run Bundle

Check	Acceptable evidence	Reject or downgrade when
input identity	paths and checksums match the governed manifest	files are substituted, missing, or unattributed
state history	legal transitions reach a recorded terminal state	transitions are missing or inferred from logs
artifacts	inventory resolves and checksums verify	output exists without an inventory or stable identity
environment	provider and dependency record is complete	behavior-relevant conditions are unknown
refusal	structured reason and unmet condition are preserved	failure is collapsed into an empty or successful result
replay	declared stable fields compare under the named policy	byte similarity is used without semantic comparison

Use Black-Box Run Verification for installed entrypoint checks and Runtime Artifact Stability for stable, conditionally stable, and intentionally variable fields.

Apply Replay Pressure

Run the primary and companion lanes independently. Then apply the relevant Runtime Replay Challenges: changed ordering, missing inputs, environment drift, interrupted state, artifact substitution, or family-specific transfer pressure.

Read the result in the Black-Box Benchmark Dashboard. A dashboard downgrade is evidence, not a documentation inconvenience. The allowed language is the maximum statement defended by the black-box packet.

Classify The Comparison

comparison outcome	required record	permitted interpretation
stable fields match under the named policy	primary and companion reports, policy identifier, field-level diff	the declared runtime behavior is repeatable for the checked package pair
direction survives but scope or coverage weakens	weakened-claim rows and affected artifacts	transfer is bounded; narrow the family statement
a declared claim collapses	collapsed-claim row and refusal or downgrade	exclude that claim even if the aggregate stability score is high
fields differ because they are intentionally variable	stability class and observed values	report variation; do not call it failure or parity
policy, environment, or input identity is missing	preflight refusal	no comparison claim is available

Separate Runtime From Scientific Acceptance

Runtime success establishes that the declared lane completed under recorded conditions. Core’s Flagship Acceptance Bars determine whether scientific outputs meet the family contract. Knowledge, Intelligence, and Lab can narrow the claim further through contradiction, recommendation instability, or downstream burden.

Result	Operator conclusion
primary and companion pass under the comparison policy	report bounded Runtime evidence and continue scientific review
primary passes, companion fails	report fragile transfer; do not generalize
execution completes but Core acceptance fails	report operational completion and scientific rejection separately
environment or input identity is incomplete	refuse the rerun claim
imported lane completes	report imported-result replay, not native engine execution

Preserve The Refusal

Open Runtime Rerun Refusals before summarizing the run. Record the failed invariant, the strongest statement still supported, and the evidence required to retry or widen the claim. Consult Current Capability Limits instead of converting a degraded rerun into broader public language.