Fractus Sandbox — CTE Cockpit
A real-time dashboard to see every metric of the Continuous Thought Engine and control/correct it on the fly. Not a chat — a cockpit.
What it shows (8 panels)
| Panel | What you see |
|---|---|
| Live Monitor | ‖thought‖, confidence gauge, tick count, lb_loss, salience_loss, memory count, μ/σ of thought |
| Oscilloscope | Kuramoto phases per block, animated on the unit circle (one cell per block) |
| Expert Routing | Bar chart of expert_hits — which experts fire, dominance % |
| Training Curve | Loss + perplexion over time (live, last 100 steps) |
| Memory Bank | All persistent memories: context, importance, vector norm — delete individually |
| Cognitive Mode | Current mode (focused/creative/exploratory/procedural) + probability bars |
| Chat + Controls | Talk to Fractus, step a single tick, reset, grow, inject tokens |
| Config + Params | Toggle memory, salience bias slider, full param breakdown (d_model, layers, experts, rank) |
What you can do
- See every internal metric update live over WebSocket (push on every tick).
- Control: chat, single-tick step-through (debugging), reset thought state, force
maybe_grow, inject arbitrary tokens. - Correct: delete a bad memory, add a manual one, or retrain on text (gradient correction via
tick_chunk_train). - Debug: tick-by-tick inspection of the residual stream, routing distribution, and oscillator phases.
Run
pip install fastapi "uvicorn[standard]"
cd sandbox
uvicorn app:app --host 0.0.0.0 --port 7860
# open http://localhost:7860
Endpoints
Monitoring (GET): /api/state · /api/phases · /api/routing · /api/mode · /api/memory · /api/training · /api/params
Control (POST): /api/chat · /api/tick · /api/reset · /api/inject · /api/train · /api/grow · /api/config · /api/memory/add · /api/memory/delete/{idx}
Live: WS /ws/live — pushes full engine state on every tick.
Config
The sandbox boots a small CTE (d_model=128, 2 blocks, 8 experts, ~6.8M params) so it
runs on CPU in real-time. To drive a bigger/loaded engine, edit get_engine() in
app.py to load a checkpoint via grow.py or torch.load.