Experiment results

Click a card on the board to open the experiment’s detail. The left side holds the candidate list and experiment metadata; the right side holds the results and the reasoning trace. The ⤢ at the top left expands it to a full page.

​

Experiment results at a glance

Four numbers sit at the top.

• Candidates how many fix candidates were built
• Best gain the improvement of the best candidate
• Broke how many previously-working cases broke
• Cases tested how many cases were scored

Metadata follows.

• Cluster, Spend, Cost / candidate, Duration, and the completion time
• Safe on holdout N/N: cases protected when Hold-out regression was on
• Surfaces the places the experiment explored, shown as chips like prompt_replace, retrieval, guardrail, prompt, memory, and add agent

​

The candidate list

Each candidate shows its number, name, and gain, and the best one carries the Best badge. Click a candidate to open its detail (see Verify and publish).

​

Reasoning trace

A timeline where the agent narrates the experiment. Stage badges for diagnosis (looking through the cluster’s failures), plan (designing the candidates), and experiment n/N are followed by a time-ordered log of question, reasoning, reference, and answer entries as cases are replayed. It updates live while the experiment runs.

​

When the experiment finishes

The card moves to the column the results call for: Optimization when an improvement is confirmed, Regression when regressions remain or no improvement was confirmed. Either way the candidates and the record stay put, so you can verify a candidate or re-run the experiment with different options.

​

Learn more