The situation
A model reads the document, scores the case, drafts the response or recommends the next step. It is useful, it is fast, and it is already in the loop for records that matter: a claim, an application, a customer's account, a person's employment. Someone will eventually ask what it was allowed to do and whether it did only that.
What goes wrong today
The model's output goes straight into the process. A workflow step reads it and acts. The same case can get two different answers on two days. The explanation is whatever the model says when asked, which is a story rather than a reason. The log, if there is one, records that a step ran, not what the model proposed or what rule allowed it.
Compliance tooling then documents all of this from the outside: a register says a model is in use, an assessment says the risk is managed. None of that is in the path of the decision.
How Amon Sul changes the failure mode
The model is moved from inside the decision to in front of it. Its output arrives as a fact, like any other fact a sequence produces. The policy decides what that fact means, and the only moves available are the routes the policy names. A model cannot approve, reject, pay, close or escalate. It can propose, and the proposal is logged next to what the policy did with it.
There is no AI inside Amon Sul itself. Every transition is deterministic and atomic. If the model is wrong, the policy still only allowed what it allows, and the record shows exactly where the wrong proposal came in.
What this enables
You can use a model in a process you have to defend. You can replace the model, or change its prompt, without changing what the process is allowed to do. You can answer "what did the AI do here" with the entry that was written at the time, under the policy version that applied, reproducible today.
The regulatory angle
Rules for AI in consequential decisions converge on the same demands: record-keeping, human oversight, a decision that can be reconstructed, a model that stays within a defined scope. The tooling built for those rules produces the documentation. It does not produce the runtime. Amon Sul is the runtime: the gate the model cannot pass on its own, and the log that proves it did not.