NeurionReference
About the benchmark results
Read reported results with their source, task configuration, and evidence limits.
HydraLabs · Published
Identify the source of a score
HydraLabs-reported internal results and externally published results are different sources. The model pages label the benchmark and source for each displayed result. HydraLabs internal scores are reported by HydraLabs and are not independently verified.
A score remains attached to its stated model, benchmark, and configuration. A result from another model revision, tool setup, or benchmark variant is not a substitute for that exact run.
Understand the evidence behind a comparison
A reproducible evaluation needs the exact model revision, dataset version, harness, tool configuration, sample size, scoring rules, attempt count, and relevant time or resource budget. Cost and latency also require their own measurements.
The HydraLabs-reported scores do not include a complete set of reproducible run reports. Read them as internal reported results, not as independently reproduced evaluations or verified duration and cost measurements.
- Compare the same task set and benchmark version.
- Keep tool-assisted and model-only configurations separate.
- Distinguish reasoning settings from different agent harnesses.
- Do not interpret chart placement as measured cost or time unless the chart explicitly defines that measurement.
Compare tool configurations carefully
Runs with CyberBot and runs without CyberBot describe different configurations. A difference between their reported scores is not, by itself, a controlled measurement of CyberBot’s causal contribution.
The same principle applies to coding harnesses and managed-agent workflows. Review what was supplied to the model and what tools it could use before comparing the percentages.
Evaluate representative work
A benchmark covers a particular task set. Assess Neurion on representative work and inspect correctness, tool use, evidence, recovery, duration, and usage separately.
A passing local integration test confirms the behavior checked by that test. It does not establish live model quality, a public benchmark score, or a guaranteed production outcome.
Keep the source version and task configuration with any results you share so another reader can understand the comparison.