Security and Safety

Knowledge artifacts influence scientific decisions, so integrity failures can be as consequential as unauthorized access. Protect both the confidentiality of source material and the semantics connecting claims, evidence, context, and recommendations.

Control ingestion and reference use

Use explicit reference packs and normalized evidence inputs. Validate their schemas before they enter a bundle, preserve rejection reasons, and reject duplicate identifiers rather than overwriting records. A source URI, bibliographic citation, accession, or accepted fingerprint is metadata; it does not prove that a source is trustworthy or authorize network access.

The knowledge package does not fetch remote references, manage credentials, scan attachments, or enforce filesystem and network policy. Applications that dereference locations must restrict allowed schemes and destinations, apply timeouts and size limits, authenticate where required, and store secrets outside knowledge models.

Treat statements, labels, aliases, ontology text, provenance fields, and review rationale as untrusted data. Escape them in the rendering or query system that consumes them. Do not turn reference text into commands, templates, or policy expressions.

Preserve epistemic safety

• Keep ambiguous protein, ortholog, pathway, and complex membership results explicit; do not force them into a unique identity.
• Retain contradictory and adverse evidence through resolution and export.
• Respect source context such as species, tissue, cell line, dose, timepoint, assay modality, and endpoint before combining records.
• Downgrade claims when annotation coverage, freshness, triangulation, or trust is insufficient.
• Carry caveats, disagreement, holds, and unresolved questions into every downstream decision brief.

Fingerprints detect deterministic content changes; they are not digital signatures, source authentication, or scientific endorsement. Where origin authenticity matters, the owning repository or application must provide signed manifests and access-controlled custody.

Respond to integrity incidents

Freeze affected decision briefs and downstream promotion. Preserve the source records, normalized bundle, ingestion report, graph validation, conflict clusters, and fingerprints. Determine whether the incident changed bytes, identity resolution, graph relationships, context, or trust policy, then find every brief and export derived from that state.

Publish corrected knowledge as a new attributable bundle with provenance to the superseded state. Never erase the conflicting record or rewrite a historical brief: reviewers need to know what evidence was available and why the earlier recommendation was made.