Investigation setup

Do these results support the next use of compute?

Recorded scores are evidence, not a verdict. Select the runs, then audit whether their evaluation conditions support a direct comparison.

Comparing Run A and Run BLive page scope exposed to the browser agent

Decision questionDoes this apparent improvement justify another training run?

Run A

New checkpoint · sanity evaluation

Selected

Recorded Recall@584%

Candidate manifest200

Manifestpool-sanity-200

Run B

Baseline · full evaluation

Selected

Recorded Recall@576%

Candidate manifest1,000

Manifestpool-baseline-1000

Neutral audit

Comparability, before ranking

The higher recorded score is visible. Whether it establishes a better model is still unchecked.