An optional Jev decision backend in the Atomic coding Agent for bounded structured choices such as routing.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
Sends predefined questions to Jev and decodes answers for callers; regular models still generate code.
WHY IT HELPS
Separates structured decisions from text-generation interfaces.
SOURCE NOTE
The verified scope is an optional structured-output provider, not that Atomic uses Jev by default or for every feature. LICENSE includes an additional UI-attribution condition above stated business-size thresholds. It is not standard MIT; read the license text.
Turn schema-bound choices into typed workflow results while handling large candidate sets.
Material analysed:
From input to action
Compile the workflow's choice questions.
Use one request when within provider limits; otherwise partition and retain candidates.
Validate distributions and decode final selected keys into the result schema.
Implementation details that matter
Large choices become a tournament
inferJev can partition candidates, keep ranked survivors and run additional rounds. One high-level structured-output call may therefore involve multiple inference requests and a changed comparison context.
The parser checks probability keys, bounds, mass and whether choice is a highest-probability option. Structural validation cannot recover a relevant candidate eliminated in an earlier round.
Useful for explicit workflow control where schemas and interruption matter. For very large candidate sets, evaluate tournament recall and call cost against a retrieval-first alternative.
How to inspect it yourself
Read planRound, packRequests and inferJev together. Check cancellation between rounds and verify that retainForFinal preserves any required escape candidate.
Limits of this review
No Atomic workflow or provider call was run.
Tournament ranking is not equivalent to evaluating every option together in one model context.