AI model evaluation
Runs one question set against multiple AI models, scores the answers, and measures behavior under pressure: agreement bias, instruction adherence, refusal calibration, and consistency.
Capabilities
Benchmarks measure capability. Prism measures behavior, using the customer's own questions.
Parallel comparison
One prompt set across hosted, private, and local models, with answers shown side by side.
Scoring
Rubrics, reference answers, judge models, and human review.
Behavioral fingerprints
Sixteen behavioral dimensions, including sycophancy resistance and safety calibration, measured under neutral and adversarial conditions.
Adversarial escalation
Graduated pressure that identifies the point at which a model fails.
Drift tracking
Repeat suites after model updates and compare results, including over long conversations.
Reports
Findings mapped to ICD 203 and the NIST AI Risk Management Framework.
Vendor-agnostic
Presba does not sell models. All providers are evaluated the same way, evidence is retained, and decisions can be re-checked after any model change.
- Any provider, private endpoint, or local model
- Prompts and answers stay in the customer environment
- Cost and latency measured with quality
- Results stored in Pallette with history and access control
- Methods based on Presba's published research
Evaluation demo
Customer-supplied questions, run on the models under consideration.