Pass your customer's security review. Ship your AI feature.

Keep your guardrails — they're good controls. But guardrails enforce; Gatekeeper proves. A security review doesn't ask whether you filter — it asks for evidence of what happened on every call: per-call audit records mapped to SOC 2 and HIPAA controls, policy and pipeline version lineage, and an exportable packet a reviewer reads in minutes. And two things guardrails can't give you by design: visibility into traffic that never invoked them — Gatekeeper flags untagged PHI calls as UNGOVERNED_CALL instead of letting them pass silently — and independence, because evidence produced by the same platform that runs your workload is the platform grading its own homework. Guardrails are a control. Gatekeeper is the evidence of control.

No product does — and you should be skeptical of any that claims to. HIPAA compliance is a program: policies, training, BAAs, risk analysis, and technical safeguards together. What Gatekeeper provides is the technical evidence layer for one hard part of it: every AI call through the gateway produces an audit record mapped to specific controls — HIPAA 164.312(b) and SOC 2 CC6.1, CC6.3, CC6.8 — showing what was detected, what was redacted, what was blocked, and under which policy version. When your auditor or a customer's security review asks "prove what happens to PHI in your AI features," these records are the answer you hand them. Gatekeeper is evidence for your audit, not a substitute for your compliance program.

No detector catches everything — and a vendor claiming 100% detection is telling you about their marketing, not their system. Gatekeeper is designed on the assumption that misses exist, and makes them bounded and visible instead of silent. Deterministic where possible: fields you configure — your MRN formats, member-ID schemas, known identifiers — are redacted by rule, not probability. Probabilistic detection, honestly measured: for identifiers in free text, we measure the hard way — strict span-level recall and precision, per entity type, on adversarial corpora — with per-entity floors locked in CI so performance can't silently regress. Misses are auditable, not invisible: every call is logged with its policy and pipeline lineage, so you can always identify which traffic was processed by which detector version. And untagged traffic is flagged loudly as UNGOVERNED_CALL, never silently passed. We'd rather show you how we measure than promise you a number that can't be true.

Good — many strong teams have. But when a hospital's security review examines it, the questions won't be "do you scrub PHI?" They'll be: what's its measured recall, and on what test set? Which detector version processed which historical calls? Can you export per-call evidence of what it did? In-house scrubbers almost always fail those three — not because the detection is bad, but because nobody built the measurement and evidence machinery around it. That machinery is most of what Gatekeeper is. Detection without evidence doesn't pass review; evidence is what we sell.

Yes. Our sample evidence packet (/sample-packet) is generated live by the deployed Gatekeeper pipeline using clearly-labeled synthetic data — the records are real rows the gateway persisted; only the data is fictional. It shows the three outcomes of a call: a governed call with detected entities redacted, a governed call blocked for prompt injection (with its block reason), and an untagged call loudly flagged as UNGOVERNED_CALL. We never use real patient or hospital data for demos or benchmarks — synthetic only, by policy.