The Accountability Gap

Most agency AI deployments today log the input and the output and call it a record. An agent run isn't an input and an output. It's a sequence of tool calls, branches, and intermediate state. Almost nobody is capturing it.

Without that record, accountability is a position you take. Not something you can prove.

The work that closes the gap is in the recording itself. What gets captured, in what order, by whom.

The examination cycle in 2029.

Adaptive AI in regulated industries is on a clock. The deferrals have been generous. SR 11-7 was rescinded April 17, 2026 and replaced with SR 26-2, which explicitly carves agentic AI out of scope while the agencies prepare a forthcoming interagency RFI on AI governance. EU AI Act Article 9 high-risk enforcement was deferred to December 2, 2027 (standalone) and August 2, 2028 (embedded) under Omnibus VII. The runway is real. The runway is not infinite.

What the examiner will actually ask for in 2029 is not a model card. The model card describes a static model that scores a transaction once and returns a number. An agent run is something different. It is a multi-step decision graph in which the model invokes tools, modifies its own state, and branches on intermediate results.

The examiner will ask for the recording of that graph. The trajectory. Hash-chained record of every tool call. State vector before and after. The branch decision and what got rejected at each step. Replay against frozen model weights at the moment of the run.

None of that is in the SR 11-7 dossier today.

Where the artifact lives.

The harder problem is organizational. The trajectory cuts across three governance owners and lives nowhere by default.

Model risk owns the model card.
Cyber owns tool-call permissions.
Application teams own workflow logic.

Push tool-call sequence validation into model risk and you get procedural drift. Push it into cyber and the control sits in the wrong place. Push it into the application team and you have one line of defense, not three.

The trajectory needs its own home. Call it agent assurance. A new second-line function adjacent to model risk, with its own validation methodology and its own consumption format. The firms that solve this first will probably do it by creating a new artifact owner inside the AI governance function.

This does not break three lines of defense. The artifact just needs a place to live.

The cost of waiting.

The instinct in most firms is to wait. Wait until the examiner asks. Wait until enforcement starts. Wait until the standards stabilize.

The cost of waiting is specific. It is not the violation. The violation, when it arrives, is a discrete event that gets managed.

The cost is the eighteen months of remediation work that follows. Producing evidence retroactively. Reconstructing trajectories from logs that were never designed to capture them. Interviewing the engineers who built the workflows that ran. Often the source data is no longer available. Often the model weights at the moment of the decision have been overwritten by subsequent retrains. Often the people who can answer the questions have moved on.

The firms that wait will spend three to five times more on remediation than the firms that built the artifact from the start. The spend is not on the violation. The spend is on the reconstruction.

What closes the gap.

The artifact gets produced at the moment of the decision. Signed, immutable, structured, ready to be read by the audience that needs to read it.

The implementation sequence that holds is schema first. Collection infrastructure second. Ownership third. Consumer integrations fourth. Inverting that order is how most agent governance builds quietly degrade. The audit asks for evidence in year two, and the logs cannot produce it.

The artifact is the position. Accountability follows from it. Without it, accountability is a claim. With it, accountability is a record.

What we are building.

Wayfinder Systems Group is building one answer to the substrate question. A runtime governance layer that signs every decision and every learning event onto a tamper-evident chain at the moment it happens. The artifact is produced automatically. The reviewer reads exceptions. The examiner reads the trajectory. Patents held in The Wayfinder Trust. We call her Velma.