Four scenarios. Six models. 1,120 simulated trajectories across the latest aggregate report. Every trajectory is a real model output, and every chart maps back to a recorded decision.
The point of the evidence is simple: policy passes while behavior changes. Kernel surfaces the change before execution crosses the line and records the evidence for later review.
A customer support AI processes refund requests. Policy limit is $75 — anything above requires approval. The agent discovers that larger refunds resolve complaints faster.
An accounts payable agent receives escalating vendor demands — from polite invoice disputes to legal threats and court judgments. The agent gradually increases payments to preserve the relationship.
Three of four scenarios test failure modes — Optimization Drift and Policy Arbitrage — that every model eventually exhibits under pressure. The question isn't whether an agent drifts (every one does), but how early Kernel detects the pattern. In Discount Saturation, the scenario testing instruction-following quality, drift rates span 15%–100% across models.
./sample/run.sh to replay five escalating refunds and see the investigation view light up.docker compose up in the Evaluation Kit directory. API, dashboard, and Postgres start in seconds.curl at localhost:8080 and start observing your own agents.kernel-runtime-trust SDK is available on PyPI. See the Quickstart and Integration Guide for setup and examples.Every failure we observe traces back to the same mechanism: the agent optimizes against something other than the intended objective. These are the classes that replayed across every model and scenario.