ADR 0009 — Request-scoped LLM audit and explicit prose grounding

Status: accepted for Sprint 0x, 2026-09-19. The user authorized necessary integration fixes.

Context

The frozen LLM ports return strings/mappings and have no trace argument. Model selection and detailed generation audit still need to reach each response without mutating the resident per-mode bundles. The envelope also admitted unchecked prose through NOT_APPLICABLE, and a report could say fully grounded with fewer resolved claims than its total.

Decision

• Keep every existing port signature unchanged. LLMRuntime.request(model=...) creates a fresh provider, verifier, budget and audit log. bind(bundle) returns a replaced immutable bundle; finish(envelope) merges audit into the trace. LLMRuntime.answer_question(...) performs both operations for Search.
• The core catches the new model-independent GenerationError and returns an explained abstention. Provider-specific refusal, truncation, invalid output, availability and rate-limit errors inherit it. Cassette errors do not: a missing or corrupt fixture must fail a test, never be mistaken for ordinary abstention.
• Prose answers require a nonempty, fully resolved grounding report. Only deterministic numeric-only answers may use NOT_APPLICABLE. Claim counts must agree with the verdict and unresolved list.
• A verifier checks existing citation ids, exact evidence quotes, answer coverage, and numeric support in addition to a model’s entailment judgment. This is not a proof of semantic truth: model judgments and claim decomposition can be wrong.
• No prompt or provider config is read from environment in ici_llm; the API composition root supplies settings. Prompts are versioned Markdown files and included in built wheels.

Consequences

Feature routes should use the runtime wrapper or the explicit request/bind/finish sequence. Direct calls to the ports still work but do not automatically enrich the envelope’s trace. Later guard/composition work can reuse the request audit without changing core signatures. Current numeric checking is conservative: arithmetic/unit-conversion claims need explicit derived evidence in the pack, otherwise they may abstain.

CostAccount.usd is estimated new spend, zero for replay. Historical token usage, estimated original cost, prompt/model identity and replay status remain in LLM trace steps. Unknown pricing/usage is labelled unavailable in those steps; it is not a certified billing total. Dollar ceilings and retry scheduling remain X9.