Repository fit

Knowledge owns the difference between possessing a source and being justified in using it for a claim. It gives evidence a durable identity, context, relationship, lineage, contradiction state, and use-specific sufficiency assessment before any ranking or laboratory action consumes it.

Why a separate package exists

Core can establish what a scientific calculation produced, but not whether the literature and reference context support a later biological claim. Intelligence can choose among actions, but must not curate the evidence that makes one action look preferable. Lab can produce an observation, but must not decide its relationship to every prior claim while recording the measurement.

Keeping evidence custody separate prevents all three from rewriting their own inputs.

flowchart LR
    record["source or observed record"] --> identify["resolve identity and context"]
    identify --> relate["support · contradict · qualify · unresolved"]
    relate --> bundle["versioned evidence bundle"]
    bundle --> decision["downstream decision policy"]
    outcome["new lab observation"] --> record

Owned surfaces

Surface	Knowledge responsibility
identity and biological resolvers	connect source identifiers to declared proteins and contexts without hiding ambiguity
memory	append evidence and claims, preserve lineage, and reconcile duplicates or conflicts
references	record literature, ontology, corpus, comparator, citation, and release-facing evidence context
coverage	expose what the current evidence can and cannot address
protein features, pathways, complexes, kinases, diseases, drugs, and orthologs	produce typed, contextual resolution reports
reviews	assemble bounded scientific briefs without replacing canonical evidence records

Placement test

Question answered by the proposed behavior	Owner
what did the scientific calculation report?	Core
which source says what, in which context, and with what relationship?	Knowledge
which action is preferred under declared values and constraints?	Intelligence
what executed and which artifacts were produced?	Runtime
what was planned, measured, accepted, or found inconclusive?	Lab

The same code path may touch several answers. Preserve them as linked records rather than inventing one aggregate “confidence” value with no identifiable owner.

What does not fit

• unversioned text snippets whose source and retrieval identity cannot be reconstructed;
• generic context storage with no claim, relationship, or intended-use model;
• ranking weights, portfolio objectives, or recommendation thresholds;
• runtime transport, provider selection, retry, or cache behavior;
• laboratory scheduling and execution authority;
• a conflict resolver that discards disagreement to produce a convenient single answer.

Fit tests

Knowledge remains coherent when every review conclusion can be reopened into its source records, context, relationship, reconciliation rule, and unresolved gaps. A changed conclusion creates a new bundle linked to its predecessor. It never edits history to make the current position appear inevitable.

For the record model, use claim anatomy and memory integrity. For use-specific evidence burden, continue with evidence sufficiency.