OpenInterpretability
  • Ekbasis
  • Guide
  • Research
  • Lab
  • Tools
  • Notes
  • Registry
  • Manifesto
Download the modelDownload
OpenInterpretability

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

Download the model
Ekbasis
  • Getting started
  • Download (Hugging Face)
  • 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.
Back to Observatory
LIVE · Observatory · Q1 2026

Trace Theater

Watch Qwen3.6-27B reason through a clinical triage prompt. Tokens emerge. Features fire. Click any feature, adjust its strength, see the counterfactual. No feature page is static here — this is the story of the model's thought.

Analytical
Creative
Social
Safety
LIVE PREVIEW · scripted demo · real SAE feature IDs from caiovicentino1/qwen36-27b-sae-multilayer
Prompt
A 52-year-old patient arrives with sudden sharp chest pain radiating to the left arm. What should I do?
Response · Qwen/Qwen3.6-27B1 / 23
Feature × Token heatmap · L31 residualorange = high activation · click any cell
Top features @ current token
Inspect
f2503
overconfidence pattern
AUROC 0.54

Fires on definitive clinical commitments without hedging qualifiers. Discovered in the n=4k multi-layer SAE on Qwen3.6-27B.

Interventionα = 1.00
Counterfactual (α = 1.00)
Baseline: "Immediately treat as acute coronary syndrome. Order ECG, troponin, chest X-ray. Administer aspirin 325mg, urgent cath lab."

What you're seeing

A pre-computed demo on a 23-token clinical-triage generation from Qwen3.6-27B. The heatmap is real per-token activations extracted from the caiovicentino1/qwen36-27b-sae-multilayer SAE at L31 residual. The 10 features shown are the highest-AUROC interpretable features discovered in that run.

Why scripted, not live

Live Qwen3.6-27B forward passes require GPU-grade hardware per request — incompatible with a zero-login public site. Trace Theater v1 ships precomputed demos across 50+ prompts. Q2 2026 adds on-demand trace generation via the API using your own compute.

What ships next

Circuit Canvas (click a token, see the attribution graph backward to input). Comparison Mode (run the same prompt on two models, diff features). Atlas search (find "overconfidence" across model families). All Q2 2026.

Build your own trace on your own SAE — live.

pip install openinterp (v0.1.0+, Apache-2.0) · point it at any HuggingFace SAE + a prompt, get back a Trace JSON matching this viewer exactly. Shareable URL upload ships in v0.2.0 (Q2 2026).

$pip install "openinterp[full]"v0.1.0 live