OpenAI4S-generated codebases

This directory contains six independent implementations generated through OpenAI4S from the frozen prompts in ../queries/. The first two came from full openai4s run Agent trajectories; the remaining four use OpenAI4S's LLM client with the public Query and benchmark source as context after the Agent gateway proved unreliable for long code responses. All six received an explicit post-generation protocol-conformance repair. Filenames match the GT entry points in ../gt_codebases/, but generated implementations never import or read the GT runtime.

Generation interface, incomplete Agent-run state, repair disclosure, prompt and source hashes, and verification artifacts are frozen in generation_manifest.json.

generate.py --scenario NAME --overwrite regenerates only that Scenario and preserves the other provenance entries. Without --overwrite, an existing source and its provenance remain unchanged. Verification stages only the public fixture and reviewed benchmark modules, uses a clean environment and an OS sandbox, and denies access to the repository, GT outputs, and network. It requires usable macOS Seatbelt or Linux bubblewrap; unavailable isolation fails verification. Child output and machine-local error details are not persisted in the manifest.

Only dataset_profile=synthetic_protocol_smoke uses fixture string canonicalization. Other profiles require RDKit for chemical normalization. The offline tests execute every committed generated CLI, compare its artifact bytes with GT, and check the recorded artifact hash.

Files