Quality & reliability
Correctness is a product feature in regulated operations.
How we're different
- Deterministic core, AI at the edges.
- No citation, no field (flagged for review).
- Policy gates + audit logs for sensitive outcomes.
Trust rule: AI drafts; humans decide for sensitive outcomes.
Evaluation harness (overview)
Extraction accuracy
Field correctness on a set of “golden packets”.
Citation coverage
Every extracted field must reference evidence.
Regression tests
Prevent prompt/adapter changes from degrading outputs.
Safety rules
- No citation, no field: if the model cannot cite, the field is flagged for review.
- Safe mode on uncertainty: produce a checklist and route an exception instead of guessing.
- Policy gates: approvals required for external actions and sensitive outcomes.
Prompt injection & evidence integrity
Inputs are untrusted
Emails and documents can include instructions. KYBRIQ treats them as evidence only.
Bounded tasks
LLMs draft and extract under schemas; deterministic rules and approvals govern outcomes.
See the threat model
Short, practical threat model for internal engineering and security review.