AI & Detection Engine Disclosure

Lacuna uses two detection passes in parallel on every job. Both run by default for every tier — including the public hero demo — because the recall gap between the pattern pass alone and the pattern + AI combination is the difference between "missed name on a resume" and "caught it." Both passes run on Lacuna-operated infrastructure; page text is not transmitted to any third-party AI service.

Pattern engine. Built on Microsoft Presidio with custom recognizers (legal court case numbers, FOIA exemption codes, DEA registrations, US NPI / MBI). Deterministic and rule-auditable: the same input always yields the same spans.

AI model. A compact, open-weights named-entity recognition model purpose-built for PII detection — not a generative chatbot. We host it on our own workers: the weights are pinned by cryptographic hash and vendored into the worker image, so no model API is called at runtime and no inference traffic leaves our infrastructure. It does not generate text; it labels spans. Nothing is trained or fine-tuned on customer documents.

Opt-out. Organisations that prefer fully deterministic detection can disable the AI pass in Settings → Organization. When disabled, every job for that org goes through the pattern engine only.

Enterprise BYO-LLM. Enterprise customers can optionally bring their own LLM key (Anthropic, Azure OpenAI, or on-prem) as an additional second-opinion detection layer. That traffic is governed by the customer's own agreement with their model provider and is itemised in their DPA; it never applies to other tiers.

The engine that processed each redaction is recorded in the audit log alongside the engine version, span counts, and a SHA-256 of the input file.