🎛️ Operator Playbook — Tiny Recursive Model Demo

This playbook is a zero-to-hero checklist for non-technical stewards who want to wield the Tiny Recursive Model (TRM) demo as a production-grade intelligence core. Follow the steps in order; each stage includes verification cues so you can confirm that the AI, economic guardrails, and telemetry channels are behaving exactly as expected.

Architecture diagram · source preserved below

View original Mermaid source

graph TD
    A[Bootstrap Environment] --> B[Run `make demo`]
    B --> C[Review ROI Scoreboard]
    C --> D{Need Overrides?}
    D -- yes --> E[Issue Owner Command]
    D -- no --> F[Inspect Telemetry]
    E --> F
    F --> G[Launch Streamlit Control Centre]
    G --> H[Continuous Monitoring & Sentinel Audits]

1. Environment Bootstrap

Open a terminal at demo/Tiny-Recursive-Model-v0.
Provision an isolated toolchain (virtual environment + dependencies):
```
make install
```
Verification: A .venv/ folder is created and the command concludes with the Streamlit, PyTorch, and pytest packages installed.

2. One-Command Demo Run

Execute the full demo pipeline — TRM training, recursive inference, thermostat updates, sentinel enforcement, and ROI summarisation:

make demo

Verification checklist

Console banner “Launching Tiny Recursive Model Demo” appears.
Final panel reports GMV, total cost, and ROI (expect ROI far above 1.0).
File assets/telemetry.jsonl is created and populated with events.

If telemetry is missing, re-run make telemetry and rerun the demo; the writer will create the directory automatically.

3. Inspect ROI & Guardrails

Review the rich table printed by make demo; it shows the TRM ledger totals in real time.
Confirm sentinel supervision by searching the telemetry log for SentinelStatus entries:
```
jq 'select(.event_type == "SentinelStatus")' assets/telemetry.jsonl | head
```
You should see JSON records with healthy, paused, reason, and the instantaneous ROI, proving that the guardrail loop is active.
Examine thermostat adjustments to confirm adaptive tuning:
```
jq 'select(.event_type == "ThermostatUpdate")' assets/telemetry.jsonl | tail
```
Look for inner_cycles, outer_steps, and halt_threshold changes responding to ROI fluctuations.

4. Owner Overrides & Governance

Change TRM parameters live using the built-in owner console (no smart-contract coding required):

. .venv/bin/activate
python demo_runner.py owner trm halt_threshold 0.65

Verification: The CLI prints a yellow “Governance” panel describing the new value and persists the update back to config/trm_demo_config.yaml. Re-run make demo to observe the effect — sentinel and thermostat events in telemetry should reflect the new halt policy.

To pause TRM entirely, edit the config’s sentinel section or use the Streamlit sidebar pause controls (see next section).

5. Streamlit Control Centre

Launch the web UI for executives who prefer dashboards:

make ui

Verification cues

Browser (or forwarded port) displays the “🎖️ Tiny Recursive Model Demo Control Centre”.
Press Run Simulation to produce the same metrics rendered on the CLI.
Sidebar override pushes configuration updates and immediately surfaces success toasts.

6. Telemetry & Ethereum Hooks

Every demo run writes auditable JSONL events. Use python -m json.tool or the jq command above to inspect.
The orchestrator emits an EthereumLog entry with RPC URL, target contract, and ROI payload; this is the staging point for mainnet-grade event logging.
For compliance reporting, archive the telemetry file after each run (e.g., copy to cold storage or pin to IPFS).

7. Regression Testing

Guarantee that training, inference, thermostat, sentinel, and telemetry logic remain healthy:

make test

This command automatically disables third-party pytest auto-plugins and runs all unit tests (test_engine.py, test_thermostat.py, test_sentinel.py, and the integration-style test_simulation.py).

Verification: Expect “6 passed” on fresh environments. Warnings about NumPy initialisation are benign on CPU-only installs.

8. Troubleshooting Signals

Symptom	Diagnostic	Resolution
Telemetry log empty	Check write permissions on `assets/`	Run `make telemetry` and ensure the filesystem is writable
ROI below target repeatedly	Inspect `ThermostatUpdate` events to ensure recursion depth is being trimmed	Consider reducing `min_inner_cycles` or increasing `value_per_success` in the config
Sentinel pauses immediately	Open the latest `SentinelStatus` entry to read the `reason` (e.g., ROI floor violation)	Adjust sentinel thresholds or re-run training to improve accuracy
UI fails to launch	Verify the virtual environment is active and Streamlit installed	Run `make install` and re-launch `make ui`

9. Ready for Production Experiments

Once the dry run is satisfactory:

Update config/trm_demo_config.yaml with production RPC endpoints, owner addresses, and budgets.
Wire the Ethereum logger to a real logging contract (ABI integration hook in governance.py).
Feed real conversion data into the TRM training loop via demo_runner.py or by extending the orchestrator.

This playbook is designed to be looped daily: bootstrap, run, audit, iterate. With these controls in place, AGI Jobs v0/v2 transforms a tiny recursive network into a business-defining intelligence asset.