Recall · Studio

Evaluation log

Recall's models propose and its code decides. This page shows what the code refused, read from the records each kit keeps of its own workflow. Text model: nvidia/nemotron-3-super-120b-a12b. Offline evaluation with method and every number: docs/EVALUATION.md. Generated 2026-10-31 15:18 UTC.

Live benchmark runs

Any logged-in judge can start a run from the studio (Tools, Live benchmark). Each run takes 5 fresh tasks at random from the held-out test set in eval/technique-samples.json and puts them through the full workflow. Code scores them: was the technique one a careful person would accept, did the storyboard meet its contract, and on which attempt. Every run is listed, including failures.

The live technique score also counts picks handed to the Memory Expert, which the published selection figure leaves out.

No live runs yet.

Across every account

Counts only. They come from the records each kit keeps of its own workflow; no wording, cue or answer is shown.

Who chose the technique: split into parts, each chosen separately 2 · Memory Expert (Nemotron) 5 · decision model 2 · the learner 1.

Published records

Every refusal in the kits of demo, newest first, with the reason recorded at the time.