Privacy and exposure

Use this guide for extraction, breach-risk, exposure, broker-lifecycle, and offline search-provider evaluation. SynthWorld supplies safely fictional public inputs and separate evaluator truth; it does not anonymize supplied real data or predict real-world harm.

Exact-span extraction. Create product-safe pages with:

synthworld generate-public-extraction \
  --seed 20260719 \
  --persona-count 10 \
  --output extraction-input.json

Adapters return the versioned extraction prediction contract. The scorer reports span precision, recall, F1, overlap, and misses by data class. See examples/evaluate_extraction.py and DATA_DICTIONARY.md.

Breach-risk calibration. Create product-safe breach observations with:

synthworld generate-risk-public \
  --seed 20260719 \
  --persona-count 10 \
  --output risk-input.json

Every case receives a predicted band. Optional score or probability fields are scored only when supplied by the task contract; the expected score is a deterministic reference index, not a probability or forecast.

Exposure and broker lifecycle. synthworld generate-corpus creates deterministic breach, broker, search, and social scenarios. Broker-removal lifecycle evaluation is available through synthworld evaluate broker; do not infer broader longitudinal product behavior from the static corpus.

Offline search-provider projection. generate_search_projection exercises search-consumer logic without a network dependency. It includes collisions, insufficient-evidence cases, syndicated copies, stale observations, missing snippets, and seed-dependent result order while keeping evaluator truth separate. The report keeps false accepts, false rejects, unwarranted decisions, coverage, distinct findings, stale acceptances, and difficulty support independent.

Boundary. Only artifacts explicitly documented as public belong on the product side. Annotated or evaluator bundles may contain expected results even when the same benchmark also provides a product-safe projection.

Use the documentation site’s evaluation guide for the general integration sequence. For schemas, formulas, benchmark identities, and frozen reference values use DATA_DICTIONARY.md and BENCHMARKS.md.