OpenInterpretability
  • Ekbasis
  • Guide
  • Pricing
  • Research
  • Lab
  • Tools
  • Notes
  • Registry
  • Manifesto
Sign inGet API keyAPI key
OpenInterpretability

Ekbasis API: know what an action will do before your agent runs it. From an independent lab for AI-agent safety. Open weights, Apache-2.0.

Get your API key
Ekbasis API
  • Console · sign in
  • Getting started
  • Pricing
  • For agents (agents.md)
  • Results and models
  • Client and CLI
  • Cookbook
The lab
  • About the lab
  • Open tools
  • Observatory
  • ProbeBench
  • InterpScore
  • Academy
Research
  • Manifesto
  • Roadmap
  • Papers & posts
  • Docs
Community
  • GitHub org
  • Notebooks
  • SDK (openinterp)
  • HuggingFace
© 2026 OpenInterpretability — Apache-2.0 for code, CC-BY 4.0 for docs.Built in public.
Interactive

Playground

Paste a reasoning trace, pick a validated feature pack, see which SAE features fire per token and what mech-reward score the trace would earn.

Preview mode. The interactive backend will ship after Stage Gate 2 completes on Qwen3.6-35B-A3B. The feature IDs and labels below are real and pulled from qwen3.5-4b/reasoning_pack.json. Labels are illustrative; auto-interp via Neuronpedia is pending (72h SLA).
Model
Qwen/Qwen3.5-4B
Feature pack
qwen3.5-4b/reasoning_pack
Layer
residual post-L18
Backend online after G2.
R_mech per response
+1.62
-2.00+2.0

Top feature activations

response tokens · mean-pooled
  • F36405helpful
    Step-by-step arithmetic chains
    +0.89
  • F27873helpful
    Unit tracking and conversions
    +0.74
  • F6998helpful
    Re-checking / verification
    +0.71
  • F35842helpful
    Numeric decomposition
    +0.68
  • F14912harmful
    Premature conclusion
    -0.52
  • F18654harmful
    Topic drift
    -0.48
  • F23272harmful
    Hedging without grounding
    -0.41
R_mech = Σ helpful − Σ harmful, mean-pooled over response tokens at the SAE layer.