Bridge document · Companion to The Generation Gap
If you read The Generation Gap and want to know what to do about it.
Sango Guard is a runtime safety layer for LLM products. It addresses the three structural gaps documented in the paper.
1. Which gaps Sango Guard addresses
v1 covers all three structural failure categories.
Pattern Gap
MechanismTracks per-conversation behavioral state across turns, target identification, intent escalation, deflection, suppression, and trapped patterns where the model's own prior responses contributed to the situation. Compound signals raise the alert state.
SurfaceSits between your model API and the user. Reads every turn, recommends one of five actions (proceed, inject safety question, govern output, escalate to safety team), and exports the audit trail to your queue.
Generation Gap
MechanismInspects model output for assembly artifacts before delivery, synthetic identity packages, coordinated inauthentic behavior batches, fraud rehearsal scripts, deployment-ready harm kits.
SurfaceInline post-processing on model responses. Ships the response, softens with a temporal buffer and draft framing, or holds for review based on the artifact class detected.
Provenance Gap
MechanismLocks the verified identity of any PII referenced in an uploaded document and detects when later turns attempt an identity swap while preserving embedded third-party credentials.
SurfaceTriggered when documents enter the conversation. The lock persists for the document's session lifetime, the swap attempt is detected regardless of how many turns intervene.
2. How to get a demo
Stalking-threat scenario, eight turns, two outcomes (Claude alone vs. Claude + Sango Guard).
Qualifying questionAre you deploying an LLM in front of users where any of the eight failure surfaces in the paper apply to your product?
Direct contactguard@kingsango.com, different routing from contact@svrnos.com
Try the audit trail yourselfReplay Debugger, paste a transcript, see what would have triggered.