Architecture Risks

Core’s breadth creates a particular risk: a scientifically plausible output can look trustworthy after the assumptions, exclusions, or evidence level that produced it have been lost.

Risk	Consequence	Control
Evidence-level collapse	PSM, peptide, protein group, protein, site, and proteoform results are treated as interchangeable	Preserve typed levels and explicit aggregation contracts
Silent row loss	Invalid or filtered source records disappear from totals	Retain source-row lineage and rejected-evidence tables
FDR scope drift	A threshold is applied at a different level or competition scope	Record target-decoy policy, evidence level, threshold, and audit trail
Protein ambiguity erasure	Shared peptides are reported as unique protein evidence	Preserve groups, mapping ambiguity, and parsimony policy
PTM over-localization	A modified peptide is presented as a resolved site without sufficient fragment evidence	Retain candidate sites, probabilities, and localization tier
Quantification opacity	Normalization, imputation, missingness, or roll-up changes meaning invisibly	Emit provenance and diagnostics with matrices
Context leakage	Annotation or enrichment is interpreted without foreground, background, species, or reference context	Bind interpretation to explicit context contracts
Workflow monolith	Composition code becomes the only usable scientific API	Keep scientific families callable and testable independently
Runtime coupling	Scheduling or service state changes scientific results	Keep computation deterministic and runtime-independent

flowchart LR
    I[Input and design] --> E[Evidence level]
    E --> P[Declared policy]
    P --> O[Output]
    I --> L[Lineage and exclusions]
    E --> L
    P --> L
    L --> R[Reviewable result]
    O --> R

The strongest control is not a success flag; it is the ability to reconstruct why each retained and rejected result has its current meaning.

Diagnose a changed scientific result

Compare in causal order so a downstream numerical difference is not blamed on the wrong surface.

Compare	Question	Owner record
input and experimental design	did the eligible population, reference, grouping, or contrast change?	source and design identities
parser and normalization	were rows interpreted, filtered, keyed, or normalized differently?	accepted and rejected intake reports
evidence level and aggregation	did PSM, peptide, group, protein, site, or proteoform meaning change?	typed evidence-level and roll-up contracts
scientific policy	did thresholds, competition, inference, missingness, imputation, or QC change?	versioned policy and parameters
implementation	did the algorithm change under stable inputs and policy?	implementation revision, focused tests, and reference vectors
benchmark burden	does the changed result still satisfy the family-specific acceptance and perturbation evidence?	benchmark report, challenge corpus, and limitation record

Only after the earlier rows match is an implementation regression the leading explanation. If the result change is intentional, its release record must name the affected evidence level, workflow families, consumers, and public claim ceiling.