Runtime behavioral governance that steers model and agent trajectories against a written constitution

Triage works at inference time, the window where conventional security tooling has no visibility and model-level alignment has documented failure modes. Its architecture watches for behavioral divergence across multiple layers of an agent's execution: the prompt it received, the context it retrieved, the reasoning it produced, the tools it decided to call, and the output it was about to return. Each of those steps is measured against a set of guidelines the customer authors themselves, which the company describes as a constitution, held under version control so a change in policy is a reviewable event rather than a silent configuration edit.

The operating principle the company states is steering first and blocking last. When a trajectory starts to drift, Integrity applies the least intervention that returns it to policy, reserving refusal for cases where correction is not viable and escalating to a human where judgment is required. The result is centralized AI telemetry, replayable investigations with full execution context, and enforcement that works across frontier, open-source, and in-house models running hosted, on-premises, or at the edge. Triage raised a $1.5M pre-seed led by BoxGroup with Precursor Ventures and several operator angels participating.