Deploy Codex Symbolic Pipeline¶

Last Updated: 2026-06-22

Run the deterministic symbolic training pipeline and verify reproducibility across environments.

Overview¶

The Codex symbolic pipeline trains a model using a corpus, preference data, and demonstration examples. The pipeline is fully deterministic: given identical inputs and a fixed random seed it produces bit-for-bit identical outputs.

Prerequisites¶

pip install -r requirements/lock.txt

Run¶

python deploy/deploy_codex_pipeline.py \
  --corpus data/corpus.jsonl \
  --demos data/demos.jsonl \
  --prefs data/prefs.jsonl \
  --output-dir runs/exp1

Reproducibility Validation¶

After two independent runs, confirm canonical manifest equality:

pytest tests/test_deploy_codex_pipeline.py -q
python scripts/audit/build_integrity_chain.py

The audit chain hashes every output artifact and records them in audit_artifacts/canonical_manifest.json. Two runs producing identical manifests confirms full reproducibility.

Troubleshooting¶