Baseline: Rev 2.0-local · Alex Blythe · September 23, 2026
A local model chooses an engineering workflow. Every choice and alternative is logged. No solver launch, approval, qualification change, or automatic handoff occurs here.
Experimental—not approved for automatic routing. On the frozen 21-request pilot, AnyJev selected 12 expected routes; the keyword baseline selected 20. Reversing the options changed 12 AnyJev selections. Review every choice; the scores do not establish reliability.
CPU inference is bounded to 180 seconds. Scores are uncalibrated—not engineering confidence.
No choice yet.
The log retains all options, ranking scores, selected workflow, selection explanation, model identity, input, and chained record hashes. Hashes detect accidental changes; they are not protection against a privileged administrator rewriting the log.
No decisions loaded.
AnyJev L0 versus raw scores from the same model and a fixed keyword heuristic. These 21 developer-authored examples are a pilot, not an independently qualified routing benchmark. Existing chat routing is unchanged.
Results have not been loaded.