Deploy Codex Symbolic Pipeline

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

Symptom	Cause	Fix
Non-deterministic outputs	Missing --seed flag	Add --seed 42
Missing corpus	Path error	Check data/ exists
CUDA OOM	GPU memory	Use --device cpu