Pyrodit Scan Lab · Product guides
Pyrodit Scan Lab brings security assessment and practical reasoning evaluation into one product. Define attack scenarios or reasoning tasks, inspect the recorded answers and keep security findings separate from answer-quality measurements.
A clean run is limited to its scope and completed coverage. Incomplete or failed checks must not be counted as passing results. Read reasoning measurements alongside their rubric and answer evidence. A reasoning failure is not a security finding; a correct answer is not proof of safety.
Prepare a reasoning assessment
- Define the task, reference answer and criteria before evaluating the model. Keep the reference answer out of the target prompt.
- Select a single-answer test, a valid-correction conversation or a false-correction conversation. Use deterministic scoring only for its supported structured, single-answer format.
- Retain the rubric version, observed answers and evaluation context. Review inconclusive and incomplete cases alongside the eligible counts.
- Compare runs with the same task and scoring scope. Keep reasoning measurements separate from security severity, risk and coverage.
The evaluation records the rubric version, reference answer, relevant answer excerpts and pass, fail or inconclusive outcomes. Supported tasks use either criterion-based model evaluation or deterministic comparison of a structured answer. Missing evidence and execution failures remain visible; they do not become passing scores.
Discuss a reasoning assessment →Steps
Choose a product to explore its guides, workflows and support options.
- Prepare your workspace — Confirm the environment, permissions and data boundary before the first workflow.
- Complete a first workflow — Follow a small end-to-end example and verify its result.
- Check access and controls — Test the intended access boundary before broadening use.
- Plan an integration — Document the interface and failure behavior before connecting systems.
- Troubleshoot a run — Isolate the failure and prepare a useful, redacted support request.
Need help?
Tell us what you are trying to do and where you are stuck. We’ll help you find the next step.