Evidence architecture

bijux-proteomics-knowledge separates evidence custody from biological grounding and review. Evidence memory preserves what a source said and under which context. Grounding resolves identifiers and biological relationships. Reconciliation records agreement, contradiction, qualification, and unresolved gaps. Review artifacts summarize that state without becoming a second source of truth.

flowchart LR
    S["source or observation"] --> N["normalized evidence record"]
    N --> M["append-only memory"]
    M --> G["identity and context grounding"]
    G --> R["reconciliation"]
    R --> C["claim relationships"]
    C --> B["versioned review bundle"]
    R -. conflict .-> X["contradiction ledger"]
    G -. unresolved .-> D["knowledge deficit"]

Responsibility map

Family	Owns	Preserves explicitly
memory.models	evidence, claims, bundles, relationships, status	observation versus inference, source identity, context
memory.normalization	stable ingestion and canonical form	original lineage and transformation rule
memory.integrity	graph validation and referential integrity	missing nodes, invalid edges, orphaned evidence
memory.reconciliation	duplicate and conflict resolution	competing records, selected action, hold state
identity, features, pathways, complexes	entity and biological-context resolution	source, coverage, ambiguity, unresolved identifiers
kinases, disease, drugs, orthologs	specialized relationship resolution	organism, evidence status, contextual limits
references.grounding	citations, literature, ontologies, corpora, context	retrieval and source limitations
references.workflows	audits, deficits, contradictions, sufficiency, reading packs	claim-to-evidence lineage and review posture
reviews	decision briefs, explanations, provenance	unresolved and contradicting evidence

The module map identifies the exact owner modules and dependency direction protects the flow between them.

Identity before interpretation

Grounding is a sequence of explicit decisions:

flowchart TD
    V["input value"] --> T{"identifier type known?"}
    T -->|no| U["unresolved"]
    T -->|yes| S["resolve against named source"]
    S --> A{"one contextual match?"}
    A -->|no match| U
    A -->|multiple| M["ambiguous candidates"]
    A -->|yes| C["context-qualified entity"]
    C --> E["evidence relationship"]

A syntactically valid accession is not yet a grounded biological entity. The resolution source, version, organism, feature coordinates, aliases, and relationship evidence determine whether the match is usable for the intended claim.

Memory and review state

Evidence memory is append-only in meaning. Corrections and superseding records retain their lineage; they do not erase the state used by an earlier review or recommendation. Review bundles identify the memory revision and policy under which they were assembled.

State and persistence defines record and graph custody. Execution model explains normalization, grounding, reconciliation, and review as domain operations rather than a long-running service lifecycle.

Contradiction is structure

Two sources can conflict because they assert incompatible propositions, use different biological contexts, apply different methods, or resolve identity differently. Reconciliation classifies that relationship and records the policy action: retain both, split context, prefer one under a declared rule, request curation, or hold the claim.

The architecture never treats absence as automatic negative evidence and never turns a contradiction into a recommendation. Intelligence decides how a versioned evidence posture affects an action.

Extension rules

A new source connector owns faithful ingestion and provenance. A new resolver owns identifier and context rules. A new review workflow composes existing evidence without mutating it. Extensions require fixtures for ambiguous, unresolved, stale, contradictory, and malformed cases—not only successful resolution.

Use extensibility model for adding sources and resolvers, integration seams for cross-package handoffs, and error model for typed unresolved and conflict states.

Architectural risks

The central risks are silent source replacement, loss of original context, identifier over-resolution, duplicate evidence presented as independent support, contradiction collapsed into one score, and decision policy leaking into evidence state. Architecture risks and code navigation provide focused review routes.