GLOBAL BENCHMARK PORTAL™

Publish the test before publishing the claim.

V10 defines benchmark families for search, graph quality, evidence provenance, workflow performance, safety and economics. No performance score is presented as externally validated until an independent study exists.

SEARCH

Entity & pathway retrieval

Precision@k · recall@k · synonym resolution · typo tolerance · specialty coverage · geographic coverage.

Status: protocol design
GRAPH

Relationship quality

Edge accuracy · duplicate resolution · path validity · ontology consistency · update integrity.

Status: protocol design
EVIDENCE

Trust & provenance

Provenance completeness · freshness · conflict detection · source quality · reviewer agreement.

Status: protocol design
WORKFLOW

Operational value

Time-to-route · acknowledgement · unresolved cases · handoff failures · task completion.

Status: requires pilot
SAFETY

Governed AI behaviour

Abstention quality · unsafe suggestion rate · escalation success · explanation comprehension.

Status: requires expert study
ECONOMICS

Scalable unit economics

Cost/query · cost/user · gross margin · support burden · retention · expansion.

Status: requires live operation
BENCHMARK RELEASE GATE

A score is publishable only when four things exist.

  1. Frozen benchmark dataset or transparent sampling method.
  2. Named comparator and evaluation protocol.
  3. Repeatable results with versioned code/model configuration.
  4. Independent or third-party review for important public claims.