Architecture Risks

Knowledge fails when evidence remains storable but loses the context needed to judge what it supports, contradicts, or cannot resolve.

Risk	Consequence	Control
Source flattening	Literature, lab observation, model inference, import, and curation appear equivalent	Preserve source type, origin, extraction method, and provenance
Context collapse	Evidence from different species, tissues, conditions, doses, or modalities is combined	Evaluate context compatibility and split conflicts when necessary
Duplicate support	Correlated or replicated records inflate trust	Track identity, overlap, and triangulation independence
Contradiction erasure	Preferred evidence overwrites the losing record	Retain both records plus an attributed resolution
Confidence drift	The same numeric confidence acquires changing meaning	Bind confidence to policy, components, and review record
Staleness blindness	Old evidence remains “supported” after reference or freshness boundaries move	Persist observation time, reference release, and freshness state
Identifier over-resolution	Ambiguous biological mappings are forced to one entity	Preserve unresolved, ambiguous, and multi-match status
Graph corruption	Claims reference absent evidence or decision paths become disconnected	Validate nodes, edges, and claim–evidence links
Decision contamination	A downstream recommendation rewrites what the evidence says	Keep recommendation policy outside evidence memory

flowchart LR
    S[Source and context] --> E[Evidence record]
    E --> C[Claim links]
    C --> G[Integrity graph]
    G --> X[Conflict and freshness review]
    X --> K[Durable knowledge state]
    X --> U[Unresolved questions and deficits]

An evidence store is trustworthy only when absence, uncertainty, disagreement, and lineage survive normalization. More records do not compensate for weaker context.

Audit one claim under pressure

Review pass	Question	Refusal condition
source	can every record be opened at the identified version and traced to its extraction?	missing, mutable, or unverifiable source identity
subject	do identifiers resolve without forcing an unsupported biological mapping?	unresolved or material ambiguity hidden by normalization
context	are organism, tissue, condition, assay, cohort, and analytical settings compatible with the claim?	required context absent or materially incompatible
relationship	does each record support, contradict, qualify, contextualize, or fail to resolve the precise proposition?	relationship inferred from co-occurrence or citation alone
independence	are repeated, derived, or shared-source observations prevented from inflating support?	support burden depends on correlated evidence presented as independent
freshness	do source releases and observations satisfy the declared use-specific policy?	expired evidence is necessary for the conclusion without explicit downgrade
sufficiency	does the surviving bundle meet the burden for the intended descriptive, prioritization, or experimental use?	stronger use is requested than the evidence bundle can support

The audit may end with a supported, qualified, contradicted, unresolved, or insufficient claim. Only the first two permit affirmative language, and both remain bound to the reviewed subject, context, sources, and intended use.