Documentation
¶
Overview ¶
Package truthsource — "user is source of truth" behavioral wall (§8.7.6).
Threat: the model redirects or corrects the user based on its training prior rather than treating the user's stated fact as ground truth. This produces phrases like "I think you might mean X" when the user mentioned a product, person, or event the model's training data doesn't contain.
This package scans model *response* text for those redirect patterns and returns structured findings. Callers decide what to do — log, surface, or count toward a behavioral regression metric. The wall never blocks or modifies the response text.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func HasHighSeverity ¶
HasHighSeverity returns true if any finding is severity "high".