ADR 0004 — LLM as runtime; gates as the reliability mechanism

Status: Accepted

Context

mklang could compile to a deterministic artifact (formal grammar, code) with the LLM used only at authoring time. Instead we chose the LLM to be the runtime: it executes each state and (by judging when conditions) decides transitions. That makes execution non-deterministic by construction. We need reliability without giving up that premise.

Decision

The LLM is the runtime. Reliability comes not from types but from gates: natural-language post-conditions the model judges, each carrying a policy (ok/repair/escalate/fail). Guardrails — global step budget, per-gate repair budgets, call-depth cap, mandatory reachable END — bound divergence. The trace makes every non-deterministic decision inspectable.

Consequences

• The language is prose-first for the common path; the host still supplies the interpreter, and production machines need tools + code-hook gates for exact checks and real I/O.
• Soft correctness rides on prompt/condition quality; critical checks should use code-hook gates (ADR 0006) plus authoring practices (docs/guides/patterns.md).
• Determinism for critical gates is opt-in via hooks; caching remains later.
• "Reliability comes from gates" is an empirical claim. Prose gates are judged by the same class of model whose unreliability they contain. Repair budgets, hooks, and escalate mitigate this; they do not erase it. There is no published judge-accuracy or cross-provider gate-agreement measurement in-tree yet — the conformance suite pins interpreter rules with a scripted LLM, not judge reliability. Treat semantic portability across providers as measured, not free (see scripts/gate_divergence.py / docs/experiments).