So I spent the evening experimenting with Jev, adding it to my working environments across several product surfaces I am developing. Here are the results from the first evening working on

So I spent the evening experimenting with Jev, adding it to my working environments across several product surfaces I am developing. Here are the results from the first evening working on auteurintelligence.com. Algernon × Jev: 48-hour integration report (Sep 18–19). All offline replays on our own labeled corpora; Jev sits behind a provider-neutral evaluation layer (named profiles, local fallback, calibration keyed by provider/model). Wins • Intent pick from a resolver shortlist: top-1 0.835 vs 0.67 baseline; wrong mutating actions 11% → 0%. ~0.5 s/call. • Agent routing (Jev + semantic router): 0.831 vs 0.814 holdout; they agree ~93%, and on disagreements Jev was right 14/15. ~0.7 s vs 5.6 s p50 for our embed+local-LLM path. • Calibration diagnosis: same accuracy, ~18× faster with a single-choice question. • Retrieval relevance: recall@10 0.807 vs 0.636 baseline (small set). • Small-sample: niche scoring 0.967, research screening 0.833, manuscript checks 0.861; p50 ~0.5–0.6 s. Weak spot • Build-failure triage (full run, 627 real cases): 59.2% vs 64.3% for our local-model baseline, with 75% abstention. p50 726 ms, p95 1.08 s. Staying in shadow. Questions Best practice to cut abstention on multi-option triage: split into atomic booleans, or a different question shape? , Any guidance on calibrating confidence per question set? We only let answers decide above per-surface calibrated thresholds. , Is there a long-lived / streaming mode? Our one-shot sidecar costs ~0.5 s