Workflow status

The audit question is almost always by clause or by responsible person — both are filters here. The result comes out in CSV through the page's export button.

Each action step of a routine shows here with the status the policy gave it: Applied (done, with the system’s reply), Awaiting review (proposed and waiting for CONFIRM, with the code), Suggested (only noted what it would do) or Blocked by cap (over the day’s action limit). The execution summary adds up all four — proof, for the owner and the auditor, that the agent did only what it was allowed to.

The trail keeps the clause's code and also its text as it was on that day. Keeping only the link would answer tomorrow's question with tomorrow's rule, which is not the one that decided.

The one responsible isn't whoever clicked — in an autonomous decision there is no click. It is the human configured on the Autonomy screen as responsible for that agent. An agent with no one responsible is an orphan decision, and the screen highlights it.

The line "I decided not to collect because there was an open agreement" is as auditable as the one that acted. A trail that only records actions hides exactly the case where the agent failed to do what it should have.

The evidence is what the agent READ before deciding — the facts, with the source of each one — plus the autonomy level of that step, the amount involved and the policy clause that applied. It is the answer to "based on what?", which is different from "what it did" (the step by step) and from "under which rule" (the policy). A run before September 2026 has no attached evidence, and the screen says so instead of showing an empty space. Whoever doesn't have the "View Values" permission gets the amount hidden: it is masked on the server, before it goes out — hiding it only on screen would leave the number accessible in other ways.

The Center audits the NUMBER (the scoreboard, the calculation record) and what CHANGED. This one audits the DECISION: what the agent saw, what it decided, under which clause and who answers for it. One run usually contains many decisions.

Because until September 2026 only two of the six engines that act on their own wrote to it. The Buyer, the Seller and the Process Analyst acted without recording decisions, and the Reconciler only recorded when someone undid something. As of this version all of them go through the same gate — autonomy level, written policy and record — including when the decision is to do NOTHING. That case is the most important of all: without it, "why did nobody collect this month?" had no answer anywhere.

It is what the agent DID to reach that decision, in the order it did it, read from the record and not from the summary it wrote. The trail always answered "what it decided and under which rule"; what was missing was "what it did to get there", and that is the question of whoever is evaluating whether to move that agent up from confirm to execute. A run before September 2026 has no recorded steps, and the screen says so instead of showing an empty list — an empty list would read as "the agent did nothing", which is the opposite of the truth.