Camera-ready correctness + adversarial-review pass for the IEEE SMC-IT/SCC 2026 paper.
Fixed two latent OPA policy bugs found by review:
- Rule 4 (accelerator fallback): now fires on any gpu/fpga step lacking a fallback, independent of the optional needs_acceleration flag (closes the omit-flag bypass; covers FPGA).
- Rule 10 (landscape_type): guards on is_string() so a Pydantic-normalized null/absent value is permitted (the field is optional), with a defense-in-depth clause for a present non-string value.
Added the experiments the camera-ready describes:
- baseline_validator.py: an in-process Python re-implementation of the ten rules, proven to reproduce OPA's decision on the ablation corpus (Section V-B baseline).
- scripts/mcp_agent_demo.py + fixtures: end-to-end MCP tool-surface demonstration (Section IV).
- docs/experiments/: OPA-vs-baseline, MCP demo, and 30-iteration performance-scaling data (Table V).
See CHANGELOG.md. This is the paper artifact for the camera-ready.