Evaluation cases
Questions, reference answers, and grading rules tied to a defined task. Keep answers out of the prompts you send to the model.
Groundtruth Data
Evaluation data with evidence behind every answer.
Use source-backed cases to find factual failures, inspect why an answer is wrong, and define the data needed to address it.
How it works
Four steps, one defined capability.
A proof evaluation measures the starting point. Remediation and validation are separate stages, not promises of automatic improvement.
Define the task
Choose the question your system needs to answer and the sources that can establish correctness.
Measure the baseline
Run a source-backed evaluation, retain model outputs, and inspect the evidence behind each grade.
Target the failures
Use verified corrections to scope remediation data. Your team controls changes to its model or system.
Evaluate the change
Test on separate cases. Report what changed, including regressions and unresolved errors.
What you receive
Data your team can inspect and use.
Provenance and results
Source references and retained evidence. Recorded model outputs and grades are included where the package specifies them.
A reproducible export
Structured files, a manifest, and explicit usage terms. Review the schema, source dates, and limitations before use.
Packaged dataset
500-case Drug Indication Proof Evaluation
A historical MED-RT source-grounding evaluation: 500 unique questions, verified reference answers, source provenance, and recorded model grades.
- JSONL evaluation and provenance files
- Historical model responses and grades
- SHA-256 manifest and internal-use license
This is separate from the retired 338-row diagnostic export. It is not clinical advice, an independently blinded holdout, or evidence of model improvement.
Browse the catalog
21 dataset and product-family entries.
Explore the existing inventory by task. The 500-case drug proof package above has automatic delivery. Other datasets require a reviewed scope and delivery agreement; larger remediation and validation assets remain pending publication.
Reference counts describe the listed export, not the combined size of every component in a product family. Component counts, source dates and limitations are in the dataset details.
Full catalog and component pricing →Custom projects
Build around a real failure in your system.
A scoped pilot can combine a baseline evaluation, evidence-backed failure examples, and targeted remediation data. Separate validation measures the result of an intervention your team makes.
Discuss a data pilotScope before scale
Larger remediation and held-out inventories require a reviewed package. They are not included in the 500-case checkout or automatically available for download.
Your data stays yours
Customer prompts, outputs, and failure records stay private by default. We do not silently train or modify your model.
Why the datasets stay separate →