Appearance

Energy Language Lab ​

The control surface for HLM models, exposed as a browser UI. Survey basins, apply reversible surgery, verify diffs, replay history, and roll back — all against the bundled HLM3-Mix checkpoint.

Status

Demo-ready browser lab. Runs against a packaged HLM3-Mix 35M checkpoint by default; the same operations work on any entry in the bundle's catalog.

What the demo proves ​

A reviewer can directly manipulate the learned energy landscape and watch the model's behavior change in a controlled, reversible way:

  1. Survey the basin structure of the current checkpoint
  2. Inspect energies at the layer and trajectory level
  3. Inject a targeted change with Hebbian-style surgery
  4. Verify the diff against the prior state
  5. Replay history to see every operation applied
  6. Restore to a prior state with one operation

Available surfaces ​

SurfaceWhat it is
Browser lab (Gradio)Port 7861. Basin survey, all-layer comparison, surgery, strength sweeps, trajectories
HLM3-Mix audit scriptFast scripted pass: survey → energy → inject → verify → diff → history → restore
CLI REPLInteractive qriton_hlm shell against the HLM3-Mix checkpoint
Synthetic-weights sandboxNo-checkpoint surface for explaining basin surgery without loading a real model

The Gradio lab is the most visual; the scripted pass is the fastest demo; the CLI is the working surface for engineering reviewers.

What the reviewer sees ​

  • Basin survey output with cluster counts and energy summaries
  • A side-by-side diff after surgery, including verification signals
  • A history view of every operation in the session
  • A restore operation that brings the model back to a prior state — full reversibility, not approximate

Why this demo matters ​

A transformer attention map is not an operating surface. A Hopfield energy landscape is. The lab makes that distinction tangible: you can read the basin structure, change it, see the change, and undo it.

Caveats ​

  • The lab runs on the bundled HLM3-Mix 35M checkpoint by default. Operations generalize to other catalog entries but the demo path is tuned to this checkpoint.
  • Surgery primitives expose a Hebbian-style update; production deployments use the same primitives but with additional safety scaffolding.

Where it fits ​

The right demo for:

  • Engineering reviewers wanting to inspect "model editing, but reversible"
  • Researchers comparing attention vs Hopfield as operating surfaces
  • Buyers asking "what does control over an AI system actually look like?"