Operator and researcher entry points. The top level holds shell scripts most
people run directly; subdirectories group the Python tooling by purpose so
unfamiliar readers can narrow their search quickly.
Top level — common entry points
| Directory |
Purpose |
| benchmarks/ |
Latency, throughput, and edge-profiling micro-benchmarks. |
| demos/ |
Standalone feature demos (TTS, PPG/HRV, EWC, federated, guardrails, multimodal, MCP, …). |
| experiments/ |
Research runs — ablation, closed-loop eval, DPO, LLM-as-judge, interpretability, fairness, sparse-autoencoder analysis. |
| export/ |
Model/data export tooling (ONNX, ExecuTorch, ExecuTorch-all-runtimes, diary→Parquet, GDPR user export). |
| security/ |
Red-team runner, Sigstore model signing, encryption-key generation. |
| training/ |
Training entry points (SAE, MLflow-tracked runs). |
| verification/ |
The 46 registered checks invoked by verify_all.py (code integrity, config, runtime, providers, infra, interview, security). |
- One script, one job. Every script exposes an argparse CLI with a
short --help. No script mutates repository state; all writes go under
reports/, checkpoints/, or paths you pass on the command line.
- Importable. Script bodies sit under if __name__ == "__main__": so
their helpers can be imported from tests without side-effects.
- Failing loudly. Non-zero exit codes for any precondition failure;
never swallow exceptions at the outer layer.
- Dependency-light at import time. Heavy deps (torch, transformers,
dspy, …) are imported lazily inside the relevant function so
--help never requires a full environment.
# Validate the whole repo (46 checks, strict mode)
python scripts/verify_all.py --strict
# Run the red-team harness on torch-free surfaces
python scripts/security/run_redteam_notorch.py --targets sanitizer,pddl,guardrails
# Export the encoder + SLM to ONNX
python scripts/export/onnx.py --output web/models/
# Run the closed-loop evaluation experiment
python scripts/experiments/closed_loop_eval.py --config configs/default.yaml