Preserve Behavioral Control Semantics Across Localization

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Localize review instructions but omit proof for high-severity claims and zero-finding acceptance. Matching headings/terms can still change behavior; frozen language versions contain such differences.

Mechanism

Extract controls for permission, evidence, stop/escalation, thresholds and output fields with locators. Map translated items as retained/omitted/weakened/strengthened and test common conditions. Record canonical contradictions separately instead of treating source as infallible. Correct material differences or document accepted changes, then independently check identifiers/runtime behavior.

Bad example

Equivalent terms/headings prove parity, even without high-severity evidence or with compulsory invented findings.

Good example

Build control mappings for evidence gates, zero findings, unavailable-tool fallback, stops and fields. Repair omissions and mark conflicting source policies for ruling. Natural wording may differ without changing conditions/outcomes.

Why the change matters

Operational semantics require conditions/actions, not nouns alone. A control table exposes hidden strengthening/weakening of authority/evidence/acceptance.

Observable expectation

Teaching unsupported high-severity claims downgrade/remain pending under the contract; clean reviews may have zero findings. Unavailable browsers are static review, not live validation. Equivalent inputs should face equal gates; document comparisons are not model efficacy runs.

Limits

Frozen English/Japanese sources differ in thresholds/features and are not current product facts. Parity does not establish runtime enforcement; identifiers/schema need separate checks. Authorized policy changes must be labeled changes, not equivalent translation.