Fine-tuned GPT-2 Small (124M) action-ranker checkpoints from the proposer–verifier architecture for auditable regulated decisions, with commercial-loan pre-qualification as the testbed.

These models only propose the next step. They are meant to run behind the deterministic verifier harness in the code repo — the harness owns correctness (legal-move enumeration, arithmetic as policy bands, termination, the replayable audit log). A model here never computes a figure and never issues a decision. See the repo's whitepaper and docs/symbolic-line-grammar.md for the input format.

The decline-certified students from the latest run (3 seeds) — these back the decline-arm, 20,000-application verdict run, and red-team results:

Weights are fp32, preserving the SHA-256 lineage recorded in each checkpoint's FREEZE.json (clone of base gpt2, sha c7d00560…).

Demonstration on synthetic data; not validated on production loan data. For this rulebook a plain program reaches the same verdicts — the model earns its place on navigation cost and cheap adaptation as policy scales, not on correctness (which is structural). Full framing, results, and limitations: the code repository.