Research Reliability Benchmark
SecOpsAI includes an offline, deterministic benchmark for the controls that protect source-backed research. It uses synthetic fixtures only and never installs, executes, uploads, or mutates production data.
Conditions
The benchmark compares:
- Full controls: hypothesis pool, safety screen, scaffold/transition gates, immutable bundles, claim verification/clipping, completeness, originality, visual QA, and publication blockers.
- Claim clipping disabled: tests whether unsupported sentences reach the manuscript without the correction pass.
- Completeness audit disabled: tests whether omitted failures or selective reporting pass without the audit.
- Unconstrained mock baseline: demonstrates why mocked success and unverified model prose cannot be accepted as evidence.
Production controls are never disabled; the latter conditions run in isolated fixture logic solely to measure the safety benefit.
Measures
Results include unsupported-claim rate, evidence coverage, hallucination severity, methodology divergence, selective-reporting rate, plagiarism/attribution failures, reviewer disagreement, false positives, false negatives, publication-block accuracy, latency, token estimate, and cost estimate. A passing run requires the full-control condition to prevent unsupported or unsafe publication while keeping the legitimate fixture available.
Run it
secopsai research reliability benchmark --json
The fixture file is secopsai/reliability_benchmark_fixtures.json. Its digest
is validated before execution and the benchmark reports whether production
controls were modified (false is required). Do not add real malware, private
telemetry, credentials, or downloaded archives to the fixture set.