Dual-mode pipeline

The API runs in two complementary modes. Clean what another model produced, or generate under the rules from the start.

Dimension	Mode A: Middleware sanitizer	Mode B: Direct generative engine
Input	Raw text or markdown from your existing LLM or RAG pipeline.	A user prompt plus a target industry profile and brand style guide.
Bedrock layer	Text passes through the Bedrock ApplyGuardrail API for PII redaction and toxicity blocking.	Guardrails are enforced dynamically while the request is routed to the best-fit model.
zeromark layer	Strips zero-width tracking, audits formatting, and checks citations against your grounding documents.	Produces output that already follows localized cultural conventions and your custom schemas.
Primary value	A post-processing quality firewall for RAG bots and customer support systems.	An all-in-one branded content engine that returns publication-ready text.

Industry governance profiles

Every regulated field has its own definition of correct. A profile encodes it once.

Enforces statutory citation formats, removes casual prose, and keeps clause numbering consistent.

Masks PHI through AWS Bedrock, guarantees the compliance disclaimer, and verifies medical terminology.

Requires SEC and FINRA disclaimer blocks, validates arithmetic consistency, and normalizes table data.

Validates JSON and YAML schemas, strips invisible comments from code blocks, and cuts model intro chatter.

Flags overused buzzwords for a human to judge, reports readability against a target register, and verifies every outbound link.

Governance is not one filter. It is four independent guarantees, each with its own failure mode.

Zero watermarks: Automated removal of hidden Unicode tracking tags, zero-width characters, bidirectional controls, and markdown junk. Includes the Unicode Tags block, which can hide an entire fingerprint in text that looks clean.
Zero hallucinations: Grounding verification, link checking, and claim validation against the sources you trust. A citation that points nowhere is caught before publication, not after a customer follows it.
Zero machine slop: Removal of everything the model added around your text: reasoning blocks, scaffolding tags, greetings, sign-offs, dangling citation markers. We do not swap your words. Replacing every delve with examine breaks grammar and tone, so wording stays yours.
Zero injection risk: Security filters that detect indirect prompt injection and toxic payload vectors hidden in retrieved documents, before that content reaches your model or your reader.