Where Structured Retrieval Stops Helping: Regulated Process Decisions Need a Quality-Unit-Gated Hybrid (Twin-Gate)

A companion study ("The Silent No") found that a build-time-validated structured layer beats answer-time vector retrieval on single-document citation faithfulness. Twin-Gate asks whether that holds for multi-document process decisions on a simulated bioreactor and batch-record world, where the governing rule may sit across equipment, deviation, environmental, and utility SOPs, raw-material specifications, the process description, the APQR, and the batch record, and the decision is continue, flag a deviation, hold, or abstain. The result reverses the single-document finding, honestly. On decisions that require judgment, pure structure is worse than retrieval (silent false-negative rate 15% vs 4%, paired McNemar exact p = 0.0007): determinism cannot perform the "in-spec but needs assessment" carry-forward, catching 0 of 24 carry-forward cases where retrieval caught 24 of 24. The quality-unit-gated hybrid wins decisively, with silent false negatives at 0% and correctness at 100% (p = 0.0039 versus retrieval). Deterministic hard-limit checks give reliability, a gated model judgment over the verified facts recovers the judgment cases, and abstention handles absent facts. The hybrid's false-alarm rate was tuned from 37% to 15%, back to the retrieval level, with no loss of catch rate, by invoking judgment only on real triggers and trusting the deterministic continue otherwise. The general lesson is to match the architecture to the task: serve facts deterministically, let the model judge, and gate the decision, which is the QU-Gate pattern, now validated in a second and harder domain. This deposit contains the preprint, the synthetic "document castle" in three shapes (human SOPs, structured cards, and AI-native cards, all generated from one fact model), the frozen gold scenario set (105 items, SHA-256 pinned), the system code, the two canonical runs, the figure, and the reference-integrity verification record. All manufacturing content is synthetic (GeneriCorp Therapeutics, a fictional company). The gold sign-off portal and its reviewer data are deliberately excluded. Code is licensed MIT; paper, data, and synthetic corpus CC-BY-4.0.