NeutralIntelligence

NeutralIntelligence / Research & practice

Evaluate the task, not the category.

The same system can produce very different risks across different uses.

Research & enterprise search

Check retrieval access, source traceability, citation accuracy and behavior when evidence is missing.

Writing & customer service

Test factual consistency, policy boundaries, escalation paths and human approval for external content.

Coding & automation

Review permissions, executable actions, dependency handling, rollback and secure code review.

Data & document analysis

Check calculations, extraction errors, incomplete inputs and traceability to source material.

Image generation

Review suitability, provenance questions, rights, disclosure and approval for intended publication.

Knowledge management

Verify access boundaries, stale content handling, ownership and procedures for updating records.

Internal assistants and other uses

Describe exactly which actions are allowed and which require human review. Consider affected people as well as the direct user. A generic benchmark may not represent your work.

Create a use specific evaluation plan