DSPy adds ReAnchor and TypeSafe Jev support; two decision-model repos improved in minutes
dbreunig · x · 2026-10-24
DSPy's Drew Breunig reports that twice this week he cloned decision-model repos and significantly improved benchmark results simply by running DSPy's ReAnchor—while noting such benchmarks are meaningless without adjusting cuts/thresholds and reporting F1.
The accompanying article announces DSPy support for TypeSafe's Jev, a "System One" model:
- Jev isn't an LLM; it returns booleans, multiple-choice, and ranked-option answers—a fast, cheap decision engine built for software to call directly.
- It fits DSPy's task/implementation separation philosophy: install via pip install -U "dspy[typesafe]" and existing Predict programs can run and be optimized (GEPA/MIPROv2) on Jev without changing task definitions.
More from coding & agent
- Berkeley & AWS: agents return faster code with Lean proofs of equivalence, 1.85x mean speedup — xiuyu_l · 2026-10-24
- Bug Hunt Bench grades frontier coding models on planted bugs; Kimi's CLI cost 2.9x Claude Code — PawelHuryn · 2026-10-24
- banteg shares Kimi writing sample explaining exact analytic beam-shadow math — banteg · 2026-10-24
- Vocal harnesses for coding agents: a voice layer to keep you synced, liked by Jack — jaivinwylde · 2026-10-24
- OpenAI Launches dots: Agent Workflows That Hunt Bugs with Codex — OpenAI · 2026-10-24
- Creative dev demo: interactive birds with birdsong that fly away when you get close — jh3yy · 2026-10-24