| # Setup for a new user |
|
|
| This kit is local-first. No credential is included and no cloud service is required for |
| the Genesis Engine path. |
|
|
| ## Offline path |
|
|
| 1. Install Python 3.10 or newer and PyTorch. |
| 2. Open [`genesis_engine/README.md`](genesis_engine/README.md). |
| 3. Run `genesis.py`, then `serve.py` to open the local browser UI. |
| 4. Keep each user's `config.json`, memory, and ledger local to that user. |
|
|
| ## Quantum path |
|
|
| The IBM Quantum account is optional. To use your own hardware archive, set the token in |
| your local environment or configuration and never commit it: |
|
|
| ```text |
| IBM_QUANTUM_TOKEN=your_token_here |
| ``` |
|
|
| The trainer can also consume an archive produced by your own measurement scripts. Every |
| record used as hardware provenance should carry a backend name, job identifier, shot |
| count, and an explicit simulation flag. Simulator records must be labelled as simulator |
| data, never as IBM hardware shots. |
|
|
| ## Train and evaluate |
|
|
| Read [`TRAINING.md`](TRAINING.md) completely before starting a long run. Begin with a |
| small smoke test, keep a held-out split, compare the plain and CST arms with identical |
| seeds, and publish the result only after the run finishes and the checkpoint hash is |
| recorded. |
|
|
| ## Privacy checklist |
|
|
| - Never upload `.env`, API tokens, camera/audio captures, private experience logs, or |
| unreviewed dialogue memory. |
| - Review the file list before every Hub upload. |
| - Treat a model trained on private conversations as private until memorization and license |
| review are complete. |
|
|
