Absolute Zero Reasoner v0 Runbook
This runbook equips a non-technical owner to operate the Absolute Zero Reasoner demo with confidence.
1. Launch the demo
cd demo/Absolute-Zero-Reasoner-v0
python -m absolute_zero_reasoner_demo.run_demo --iterations 25 --tasks 6Outcome:
- Console JSON summary (ROI, GMV, cost, TRR++ baselines).
reports/absolute_zero_reasoner_report.mdwith Mermaid timeline.reports/absolute_zero_reasoner_metrics.jsonfor downstream analytics.
2. Interpret the telemetry
Open reports/absolute_zero_reasoner_report.md in any Markdown viewer. It contains:
- Mermaid line chart of solver success rate versus iteration.
- ROI table listing GMV, compute cost, and guardrail notes per iteration.
- Alerts column explaining thermostat or sentinel interventions.
Positive ROI (GMV > cost) indicates the agent is delivering more value than it consumes.
3. Adjust behaviour without coding
Modify absolute_zero_reasoner_demo/config/default_config.yaml:
iterations/tasks_per_iteration– training budget.proposer.max_program_loc/proposer.difficulty_step– curriculum difficulty ramp.solver.accuracy_floor/accuracy_ceiling– exploration vs exploitation.rewards.econ_weight– importance of market utility.guardrails.target_success_rate– thermostat sweet spot.
Re-run the demo to apply changes. All knobs are safe to tweak; the guardrail centre prevents runaway behaviour.
4. Integrate with AGI Jobs v0 (v2)
- Connect to live LLMs: Swap the stochastic solver for the fm.chat adapter (
solver.pyis isolated for easy replacement). - Plug into the marketplace: Replace
market.pywith the production economic simulator that reads on-chain job data. - Stream telemetry: Send
TelemetryStreamrecords to the global observability bus for 24/7 monitoring. - Automate with CI: Add
python -m absolute_zero_reasoner_demo.run_demo --iterations 5to nightly workflows to validate sandboxing and telemetry health.
5. Troubleshooting
| Symptom | Resolution |
|---|---|
SandboxViolation error |
Task attempted to import a forbidden module. Reduce proposer difficulty or inspect generated program. |
diversity-floor-breached alert |
Curriculum became repetitive. Increase proposer.difficulty_step or flush buffers by deleting reports/*.json. |
| ROI negative | Increase iterations (more self-play) or raise rewards.econ_weight to prioritise valuable skills. |
6. Reset state
Delete reports/absolute_zero_reasoner_metrics.json and rerun the demo to start fresh.
7. Automate via Makefile
After pulling the latest repository root:
make absolute-zero-demoThis target runs the demo with production defaults and prints the summary JSON, guaranteeing consistency across environments.
Stay within this runbook and you can operate the Absolute Zero Reasoner without touching code.