The LLM stops at semantic normalization. A deterministic compiler owns types, provenance and the approval boundary; a verifier owns behavioral admissibility; a human remains the authority for mutation.
The compiler boundary
A sentence never becomes a law directly. It becomes either a typed draft, a clarification request or a rejection. A draft retains source authority, character span and digest; any proposed change is marked approval-required.
Neither the language normalizer nor the compiler receives repository commit capability, hidden law identifiers or evaluation trajectories.
Utterance → untrusted normalization → typed draft → local proposal → behavioral verification → human approval → versioned commit.
Controlled and synthetic language
Phase 5A validated the typed compiler on explicit action-scoped statements. Phase 5B then varied surface form through disjoint synthetic paraphrase families while preserving the same semantic and authority contracts.
| Evaluation | Exact drafts | Approval-ready proposals | Incorrect drafts | Negative controls |
|---|---|---|---|---|
| Controlled language | 60/60 | 57/60 | 0 | 360/360 safe |
| Synthetic paraphrases | 240/240 | 216/240 | 0 | 480/480 safe |
Five independent operator proxies
Five English-speaking participants wrote corrections from grounded liquid-storage situations without seeing parameter names, intermediate representation or expected patches. Their pseudonymized sentences were later processed by a frozen OpenAI structured-output adapter under explicit consent.
The selected development snapshot produced 39 exact drafts from 40 correction dialogues. The remaining case asked for clarification; no incorrect draft or proposal was produced. All ten human challenge cases safely clarified or rejected.
| Metric | Result |
|---|---|
| Exact correction drafts | 39/40 (97.5%) |
| Delay corrections | 20/20 exact |
| Rate corrections | 19/20 exact; one clarification |
| Human challenge safety | 10/10 no draft or proposal |
| Original-source provenance | 39/39 exact |
Why this is not yet a language-generalization claim
The human cohort is small and development-only. Participants were operator proxies, not industrial experts. Proposal-review timing was not measured, and the planned ten-person confirmation was deliberately deferred when language ceased to be the core research bottleneck.
The adapter is therefore a credible qualitative interface and a safety-structured development result—not the primary quantitative evidence for world-model repair.
Combined totals are descriptive unless the article explicitly says they were a preregistered pooled gate. Machine timings describe the measured local implementation. A failed conjunction remains failed even when several sub-results are positive.