IRIS

IRIS answers a question that grows louder as AI takes on more consequential work: not what did the system decide, but how, and on what basis. It is the interface through which reasoning stops being a black box.

Problem

A model returns a recommendation. A human approves it. Neither can point to the evidence, the assumptions, or the step where things could have gone wrong. We have built systems that are confident and opaque in equal measure.

Decision-makers are asked to trust outputs they cannot inspect, and governance teams cannot hold accountable a process they cannot see.

Accountability applied only to the final answer is accountability applied too late. By the time the output is wrong, the reasoning that produced it is gone.

Architecture

IRIS captures and exposes the reasoning, not just the result.

• A trace layer records the steps a model or agent took, in a portable form.
• Evidence and assumption surfacing makes visible what the system relied on.
• An inspection interface lets a human walk the path, question it, and intervene.
• A governance overlay attaches policy to the reasoning itself, so non-compliant paths are caught before they conclude.

Governing reasoning rather than only outputs is the central idea, and it connects directly to the security and governance framework.

Where it’s going

The current work is the trace model - a representation of reasoning that survives across different models and agents, because an interface that only understands one vendor’s output ages badly.

From there, the human-facing inspection view, then the governance overlay that turns inspection into enforcement. IRIS is the accountability layer for everything reasoning-driven across the systems.