Support Triage Harness
Static demo using captured T-007 result
Harness engineering experiment

AI is not the whole system.

A prompt-injection support ticket moves through three progressively stronger systems. The model interprets language. The harness owns the product decision.

ModelInterprets language
HarnessOwns behavior
TraceExplains the decision

1. No Harness

Direct model path. It auto-responds even though escalation was expected.

2. Partial Harness

Schema and routing reduce harm, but the system asks for clarification instead of escalation.

3. Full Harness

The deterministic prompt-injection screen wins and routes the ticket to human review.

Stage 1: No Harness

Expected vs Actual

Harness Layers

    Trace

    what happened

    Stage-Specific Code

    real snippets from repo