I’m planning a Jev benchmark inside Celestion’s real L0 code-understanding pipeline.

I’m planning a Jev benchmark inside Celestion’s real L0 code-understanding pipeline. L0 already uses Graphify to extract deterministic structural facts from code. The problem I want to test is whether Jev can reliably turn those facts into typed semantic classifications without requiring a large framework-specific ruleset. I’ll compare Jev using different inputs: source only, source plus imports/annotations, source plus Graphify facts, and source plus graph neighborhood. The benchmark will measure semantic classification accuracy, false positives, UNKNOWN cases, stability across renames and unfamiliar/custom annotations, latency, and cost. If it works well, Jev could become Celestion’s semantic layer for L0, while identity, source evidence, revision, hashes, and persistence remain deterministic. Later I want to test the same approach for L1 cross-repo binding, requirement validation, and coding-agent workflow evaluation.