← Back to Science
Methods before claimsAn atomic, verifiable statement linked to evidence, a coverage envelope and separate human assessment dimensions.
Benchmarks
Synthetic control sets test the reference implementation only.
No-network hostile target preflight
Measure a capture target against static safety rules before opening a network connection.
One public HTTPS target and eleven negative controls cover local and cloud addresses, an unsafe port, credentials and parser ambiguity.
- exact agreement with the expected decision
- admitted and rejected control counts
- absence of network activity
Limitations
- Static preflight does not prove a later DNS resolution or redirect chain.
- Synthetic targets are not a complete production egress test.
Case-scoped trilingual search
Measure deterministic Russian, Kazakh and English query results in a temporary synthetic case.
Two positive and one negative query per language are executed and exact result record identifiers are compared with the expected set.
- exact record-set match rate
- positive-query pass rate
- negative-query pass rate
- zero-result rate
Limitations
- Nine synthetic queries do not represent the distribution of real user queries.
- The check does not measure morphology, transliterationA rule for rendering a name in another script; it aids discovery but does not prove identity. or semantic relevance.