Skip to main content
Bench collects repository evidence during recognition. You can add more information later from Understanding inside an AI system.

​
What should Bench know?

Describe the intended outcome, a rule, or an important exception. For example:
This training agent plays the customer. It must never take the trainee’s role, even when asked to switch. End the exercise only after explicit confirmation.
Use Add more info to attach a document or upload test cases. Cases belong in the test library when their inputs and expected answers need to remain structured.

​
Evidence is not policy

Traces show what happened. User rules define what should happen. An imported model score is not automatically ground truth, and an observed behavior cannot silently override a written policy. A case-specific correction applies to that case, not every conversation. Bench retrieves relevant evidence rather than putting the entire source archive in every prompt. Explicitly selected library cases and criteria use their own pinned input path so summary truncation does not remove them.

​
Manage sources

Open Manage sources to search, inspect or remove saved evidence. Sources are paginated. Saving new evidence queues a summary refresh in the background. You can leave the page while it runs. A failure keeps your sources saved and offers Retry. The worker uses the exact context revision. An older response cannot replace a newer summary or restore removed evidence. Repeated requests for the same revision do not create duplicate jobs.

​
If understanding could not update

This means the background summary did not complete, not that the source upload was lost. A failed job remains visible until it is retried or newer evidence creates another revision. Reloading the page does not start another model request. Failures can come from a model request, a timeout or a response that does not pass Bench’s structure and citation checks. Saved evidence remains available in Manage sources. Missing business purpose is a context gap, not automatically a summary-generation error. Evaluations use the pinned, attributed source evidence directly, not just this display summary. A failed summary therefore does not by itself remove context from the evaluation or block it. The evaluation still needs working model access and eligible inputs. Repository rescans reuse the existing scan payload: README and package metadata are shared background, while prompt, model, tool and recognition evidence belong to the relevant system. Unchanged rescans do not duplicate sources. Removing an agent retires its repository evidence without deleting its run history or your corrections. Removing a source also erases historical context derived from it. Previously computed scores remain historical records and are not recomputed under a new policy. A new bench creates a new evaluation against the remaining evidence.