Working draft — under active review. Numbers reproduce from frozen, hash-pinned artifacts; ablations and the scaling curve are in progress. This is a concrete instance of the Perslis thesis — neural proposes, symbolic disposes — proven on legacy COBOL.
We couple (i) symbolic, provenance-tracked knowledge retrieval; (ii) deterministic program composition; (iii) executable compiler-and-runtime verification; and (iv) explicit refusal. The system generates behaviourally correct GnuCOBOL across three structurally distinct program families, serves each answer in ≈486 ms on one CPU core with no GPU, and never emits unverified code. Because the verifier is an external deterministic oracle rather than a model grading itself, self-improvement cannot be reward-hacked.
Results (Wilson 95% CI)
| Measurement | rate | 95% CI |
|---|---|---|
| Deterministic system, self-test (3 families) | 13/13 | [77.2, 100]% |
| Adversarial dogfood (correct-or-refuse) | 27/27 | [87.5, 100]% |
| Compositional A∘B (non-trivial) | 40/40 | [91.2, 100]% |
| Neural baseline, unseen templates (motivates no-bake) | 0/299 | [0.0, 1.3]% |
Limitations are stated prominently in the paper: the working generator is deterministic composition (a from-scratch neural model did not generalise); composition and mutation spaces are hand-bounded; intent recognition is keyword-based; several headline rates are on small N with honest lower CI bounds. The broad scaling hypothesis is framed as open.