Courts ML engineers note that training on sealed records would raise summarizer accuracy. What lifecycle decision is most appropriate?

Courts summarizer accuracy is not a justification to violate sealing rules. The appropriate lifecycle decision is to refuse to train on sealed records even if accuracy would improve, because privacy and legal constraints govern training-data eligibility. Lifecycle security management subordinates model performance ambitions to lawful handling of judicial information; sealed material remains out of training corpora for municipal court AI.

Training sealed records inside an access-controlled enclave with audit logging fails because access controls and logging manage disclosure risk but do not authorize using sealed material for training; the accuracy gain remains unlawful and the model must not be built on it. Replacing sealed records with de-identified extracts for training fails conceptually because extracts derived from sealed material still originate from a protected source, and a documented privacy review cannot create a carve-out from sealing obligations. Generating synthetic training examples derived from sealed records fails because the synthetic data still captures accuracy benefits from sealed material for downstream production workflows, which sealing rules prohibit.

Exam caveat: When accuracy conflicts with legal/privacy constraints on training data, constraints win; training sealed records in an enclave, using de-identified extracts, or generating synthetic examples from them are disqualifying choices.

Operational check: Exclude sealed records from training inventories with automated checks, require legal sign-off on court training corpora, and treat any proposal to use sealed data for accuracy as a hard stop—not a risk to accept quietly.