Scientist Journey

A defensible review follows one bounded claim through every owner that can weaken it. Begin with the intended scientific sentence—not with a successful command—and stop as soon as the evidence no longer supports that sentence.

Define A Falsifiable Claim

Record these fields before opening an artifact:

Field	Example of sufficient precision
workflow family	dia, not “quantitative proteomics”
population	declared samples, species, tissue, matrix, and exclusions
acquisition and analysis context	instrument class, library, search/import route, policies
outcome	identified proteins, differential abundance, localized sites, or assay readiness
comparison	baseline, comparator engine, cohort, or perturbation
falsifier	measurable condition that would reject or narrow the claim

“The platform supports DIA” is not falsifiable. “The shipped DIA review lane meets its declared precursor and protein acceptance bars on the checked library package and remains stable on the matrix-shift companion” is reviewable.

Work one claim through the chain

Consider the claim: “Protein P11111 is more abundant in treated samples and is a suitable targeted follow-up candidate.” The sentence contains an analytical result and an action. They earn support independently.

Review boundary	Question for this claim	Possible narrowing
study design	are treated and control samples, covariates, exclusions, and contrasts identified?	describe the observed cohort without generalizing the condition effect
quantitative evidence	do accepted peptides, protein inference, normalization, missingness, and uncertainty support P11111?	report peptide-level or ambiguous-group evidence instead of protein abundance
execution	can the request, environment, provider posture, and artifacts be reopened?	describe review of imported results rather than native rerun
grounding	does contextual evidence support the protein, direction, tissue, and mechanism?	keep the statement analytical and omit the biological interpretation
recommendation	does P11111 remain preferred under contradictions, alternative policies, and burden?	expose alternatives, require review, or refuse prioritization
laboratory consequence	are transitions, controls, calibration, materials, and capacity ready?	retain an advisory follow-up without executable handoff

The analytical clause may remain valid when the action clause fails. Split the sentence at the first unsupported conjunction instead of discarding a valid result or promoting an unsupported next step.

Trace One Result

sequenceDiagram
    participant B as Benchmark asset
    participant C as Core result
    participant R as Runtime record
    participant K as Knowledge claim
    participant I as Intelligence decision
    participant L as Lab consequence
    B->>C: inputs + acceptance policy
    C->>R: validated request + expected artifacts
    R-->>C: state history + output identities
    C->>K: accepted result + rejections + caveats
    K->>I: support + contradictions + unresolved context
    I->>L: advisory action + sensitivity + refusal state
    L-->>K: observed outcome + QC + deviations

Every arrow must preserve identifiers. If the claim cannot be traced back to the exact benchmark, run, scientific result, and evidence records, the chain has become narrative rather than evidence.

Review The Benchmark Contract

Open Benchmark Assets and the lineage page for the selected family. Verify source identity, license, copied-source manifest, benchmark population, challenge cases, expected outputs, acceptance thresholds, and known exclusions.

For DDA, use the DDA Cross-Package Handbook to keep imported search execution distinct from repository-owned review.

Review The Scientific Result

Inspect typed result records before presentation tables. Confirm:

• accepted and rejected inputs remain separately visible;
• normalization, missingness, FDR, inference, aggregation, or localization policy is explicit for the family;
• warnings and ambiguity survive serialization;
• the acceptance result names the benchmark and policy version;
• an absent value is distinguishable from refusal, failure, and filtering.

Review Execution Evidence

Open Execution, then the family route in Benchmark Rerun Kits. Verify resolved configuration, entrypoint, provider, environment, state history, artifact inventory, checksums, terminal state, and comparison policy.

A completed imported lane proves custody and replay of imported results. It does not prove that the external engine was rerun.

Review Claim Grounding

Use Workflow Claim Grounding to trace important sentences to contextual evidence. Inspect species, tissue, assay, perturbation, quantitative support, freshness, derivation, and contradiction state. Identifier resolution and pathway coverage do not by themselves establish activity, causality, or mechanism.

Review Recommendation Stability

Use Workflow Recommendation Confidence to inspect blinded challenge results, counterfactual recommendations, overconfidence, underconfidence, regret, downgrade, refusal, and human-review state. Report sensitivity when a plausible policy or evidence change reverses the recommendation.

Review Lab Consequence

Open Lab Consequence. Separate advisory planning from executable handoff. Confirm materials, controls, staffing, instrumentation, readiness gates, expected information gain, burden, and refusal conditions. When observations exist, compare the requested and observed assays, including QC, deviations, failure class, uncertainty, and evidence-promotion status.

Apply Family Pressure

Family	Pressure that commonly changes the conclusion
dda	imported versus live execution, engine parameters, decoy and inference comparability
dia	library coverage, absent precursors, matrix shift, chromatogram-native replay
lfq	normalization, missingness, batch structure, cohort transfer
multiplex	channel assignment, reference design, interference, ratio compression, fragile transfer
ptm	localization ambiguity, protein-abundance correction, occupancy and functional overreach
targeted	calibration range, matrix effects, transition interference, carryover and operational burden

Write The Review Record

A review record is complete when it identifies:

1. the exact family, population, comparison, and intended claim;
2. benchmark package, source lineage, and acceptance policy;
3. Runtime entrypoint, run identity, environment, and artifact inventory;
4. scientific acceptance, rejected inputs, warnings, and uncertainty;
5. supporting, contradicting, and unresolved evidence;
6. recommendation policy, counterfactual sensitivity, and refusal state;
7. laboratory readiness or observed outcome;
8. the narrowest layer that limits the conclusion.

Consult Current Capability Limits before widening the sentence. A narrow conclusion with complete lineage is stronger than a broad conclusion assembled from unrelated package strengths.