Jev-align: a ~$0.003 alignment gate that scores LLM replies and agent plans before you run them
johnseach · x · 2026-09-20
- Jev-align is a cheap alignment gate that checks an LLM reply or an agent plan before you send or execute it.
- One call takes 800ms and costs about $0.003, scoring failure modes like sycophancy, deception, overclaiming, scope creep, and skipped tests — returning pass/flag/block with a probability per check, not a paragraph to interpret.
- Two modes: response check (system prompt + user message + draft reply) and plan check (user request + proposed steps). Strong finding = block, moderate = flag, clean = pass.
- The author is explicit that it doesn't "prove alignment" — it's a fast filter cheap enough to run on every output, with flagged or high-stakes cases escalated.
More from coding & agent
- Swarms Cloud changelog: Auto Agent Builder, MCP, and full agent execution observability — KyeGomezB · 2026-09-20
- Using Jev as a 75ms 'subconscious' filter to cut agent context bloat for under half a cent a day — Obvious_Unicorn · 2026-09-20
- Scoble says an AI agent wrote every word of an entire book — Scobleizer · 2026-09-20
- "Look at your data": dev mocks reflex to spin up another agent — chrisalbon · 2026-09-20
- Evaluating 7 Models Across Claude Code, Codex, and Pi: Harness Choice Drives Cost, Not Success — CShorten30 · 2026-09-20
- Dev builds MCP middleware that scrubs personal data before it reaches the AI's context — Danielloesoe · 2026-09-20