eq-layer: an open-source control layer that decides how an LLM should answer, with auditable contracts
Emergency-Tutor-7 · reddit · 2026-11-23
eq-layer 0.2.6 (pip/npm) is a control layer placed in front of a fixed model that decides how the LLM should answer before it answers, keeping decisions inspectable.
- Per-turn affect state tracks valence, arousal, escalationdelta, stance and subtext instead of classifying one message in isolation; an intent classifier keeps the full posterior over action types, and a one-step Bayes-risk router decides act vs. ask.
- Those states pick one policy family (mirror, validateredirect, direct, ask, deflate, repair, hold, boundary), factored into separate taskmove/socialmove/repairmove/realization controls so tone adaptation can't silently override the user's request.
- The action compiles into a system instruction and prompt prefix, then the response is audited against the contract: question budget, required clarifications, privacy promises, repeated repair, verbosity overrun.
The author admits the core claim is untested: no real-model A/B run yet, zero adjudicated labels, and steering is prompt-level only — logitbias exists but is never applied.
More from coding & agent
- wllama v4.0 ditches wasm mem64 for mem32 (up to 22% faster), 2.8x speedup on Chrome — ngxson · 2026-11-23
- Replit agents speed up NumPy 18.9x — but only Claude's team scales — amasad · 2026-11-23
- A better yardstick for 10x engineers: how much of their work gets reused — vboykis · 2026-11-23
- Microsoft rebrands Azure AI Search as Foundry IQ, a knowledge layer for agents — adnan_hashmi · 2026-11-23
- Gemini agent builds rigged Optimus robot in Blender with 145 parts from photos — majidmanzarpour · 2026-11-23
- Dev levels up Neon-built escape room game with an epic AI-generated intro video — jacalulu · 2026-11-23