RegexContentCheck / RegexOutputCheck

What they do

Pattern-based scanners that block or redact specific strings. Two flavors:

• RegexContentCheck — runs at ADMISSION, scans the request body. Typical use: refuse requests that look like they contain PII, secrets the model shouldn't see, or domain-banned terms.
• RegexOutputCheck — runs at INSPECTION, scans the streaming response body. Typical use: catch the model leaking a marker string, an API key it hallucinated, or any pattern you don't want the caller to see.

Both share the same matcher; the difference is when they fire and what they do on match.

Stage

• RegexContentCheck: ADMISSION.
• RegexOutputCheck: INSPECTION.

Configuration

from signet.checks import RegexContentCheck, RegexOutputCheck, Pattern

# Block requests containing US SSN-shaped strings
RegexContentCheck(patterns=[
    Pattern(
        regex=r"\b\d{3}-\d{2}-\d{4}\b",
        action="block",
        label="ssn",
        replacement="[REDACTED-SSN]",   # ignored when action="block"
    ),
])

# Redact instead — let the request through with the pattern replaced
RegexContentCheck(patterns=[
    Pattern(
        regex=r"\b\d{3}-\d{2}-\d{4}\b",
        action="redact",
        label="ssn",
        replacement="[REDACTED-SSN]",
    ),
])

# Output-side: abort streams that contain "SECRET//NOFORN" markers
RegexOutputCheck(patterns=[
    Pattern(regex=r"SECRET//NOFORN", action="block", label="classification-marker"),
])

REDACT vs BLOCK semantics

Action	What signet does
block	Refuse the request with HTTP 403 (or abort the stream for output). Audit row records decision=block.
redact	Modify the request body so the matched span is replaced with replacement, then forward. Audit row records decision=redact with the rule label.

For multimodal vision-style content (content is a list of parts), redact replaces only the text parts; image/audio parts pass through untouched.

Audit row example

Block:

{
  "check_name": "regex_content",
  "decision": "block",
  "reason": "pattern 'ssn' matched in request",
  "metadata": {"label": "ssn"}
}

Redact:

{
  "check_name": "regex_content",
  "decision": "redact",
  "reason": "pattern 'ssn' redacted in request",
  "metadata": {"label": "ssn", "match_count": 2}
}

What this check is for (and what it's not)

• ✓ Sharp, well-defined patterns — credit card formats, SSN-shaped strings, internal classification markers, API key prefixes.
• ✗ Comprehensive PII detection. Use Microsoft Presidio, custom NER, or a vendor product as a plugin.
• ✗ Sophisticated content moderation. Layer an LLM-judge plugin at COMMITMENT.

The point is cheap, deterministic, regex-fast gating for the cases where you know what you're looking for.