Energy Language
Energy Language is a complete programming language for editing neural network behavior through basin surgery. It provides 48 operations across three interfaces.
Interfaces
| Interface | Use case |
|---|---|
| CLI / REPL | Interactive exploration and scripting |
| Python API | Integration into pipelines and applications |
| Jupyter | Research with visualizations |
Operation Categories
Observe — read the landscape
survey, survey-all, verify, energy, probe, landscape
Modify — change the landscape
inject, remove, move, strengthen, weaken
Concept — semantic operations
capture, inject-concept, remove-concept, blend, concepts, export-concept, import-concept
Control — flow and persistence
load, apply, restore, restore-all, save, status, set, info
Concept Algebra — vector arithmetic on concepts v0.10
similarity, add, subtract, analogy, compose, interpolate
Consolidation — memory sleep cycle v0.10
consolidate, dream
Watermark — IP protection v0.10
watermark inject, watermark verify, watermark strip-test
Multi-Model — cross-model operations v0.10
compare, transplant
Causal — discovery and intervention
causal scan, causal intervene, causal counterfactual
Safety — guards and audit
guard, guards, diff, benchmark, history
Verify — generate and test
generate
See the full Operations Reference for details.
How It Works
Every operation maps to a mathematical primitive on the weight matrix W:
- Inject = Hebbian outer-product update:
W += α · outer(basin, f(basin)) - Remove = Anti-Hebbian:
W -= α · outer(basin, f(basin)) - Move = Remove at source + Inject at destination
- Verify = Perturb by 10%, re-converge, check cosine similarity > 0.9
Strength is normalized relative to ‖W‖_F to preserve scale.