0009. Add an opt-in semantic second gate over the verified claims

• Status: Accepted
• Date: 2026-09-11

Context

The verbatim gate (0002) proves a claim's quote occurs in the source and every number in the claim occurs in the quote. It is deterministic, model-free, and honest that it checks wording, not meaning: a quote can be present, and its numbers match, while the claim reads the opposite direction, states a hedge as settled, or is about something else. That semantic gap is the one thing the gate documents it cannot close.

Decision

Add an opt-in second pass (--second-gate / DRAFT_SECOND_GATE): after the verbatim gate, a local model judges whether each verified quote actually supports its claim, and unsupported claims are dropped before writing. It is off by default; the verbatim gate stays the primary, always-on check.

Alternatives considered

• Make it the default / replace the verbatim gate. Rejected: a model-based judgement is non-deterministic and can err; it must never be the thing that decides grounding. It is additive on top of the deterministic gate, not a substitute.
• A bundled NLI model (MiniCheck-class). Rejected for now as a heavy dependency; the pass runs through the configured engine seam instead, so a user points it at whatever model they already have.
• Fail-closed on model error. Rejected: a flaky or offline verifier would then silently thin a run's grounded material. The pass is fail-open — a model error, a cancelled context, or an unreadable verdict keeps the claim.

Consequences

• It costs one extra model call per verified claim when enabled, so it is a deliberate opt-in, not free.
• Its quality depends on the model behind the engine seam, which is why it is additive and fail-open rather than authoritative.
• The verbatim gate's guarantee is unchanged whether it is on or off.