A broker opinion of value run is not one model call. In the run opened up below, the run started at 06:27:32 UTC on 16 September 2026 and the answer was delivered at 08:32:53 UTC, a wall clock of 2 hours 5 minutes 21 seconds. Beneath that one run sit 314 agent runs, and 120 evidence records are attached to it. Every figure on this page is read from that run's own records.
The request
The first message read: run a BOV at scale, with a source-backed value range, key comparables and a brief rationale. That is the whole input. No workbook, no template, no instructions about method. From that one line the run worked out for itself that the job was a valuation.
What the records show
Three numbers describe the shape of the work.
- The clock. Started 06:27:32 UTC, delivered 08:32:53 UTC. The delivery time is stamped on the run itself, so the 2 hours 5 minutes 21 seconds is a difference of two recorded timestamps, not an estimate.
- The agents. The run's records list 314 agent runs beneath it. Of those, 285 finished and 29 ended in an error state. We do not read the 29 as harmless. Which of them touched which part of the answer is exactly the kind of question the evidence records exist to answer.
- The evidence. 120 evidence records are attached to the run. A figure that appears in the answer should be traceable to one of them. A figure that cannot be traced to one should not be in the answer.
Why so many agents for one answer
Because the work splits. Classifying the request, researching comparables and writing up the result are separate jobs with separate inputs. Each is handed to its own agent, and one agent ending in an error does not by itself end the run. That is a strength and a risk at once: the run can finish while part of its work did not, which is why we publish the counts rather than only the answer.
What this run is not
It is an opinion of value, not an appraisal. A lender or an investment committee should treat it as a starting document: a sourced range with its comparables and its reasoning, which a person then tests.
It also does not guarantee a file. A run returns a written answer in the room. A workbook or a deck comes back on some jobs and not on others, so if a file is what you need, say so in the first message.
What to ask of any valuation run
- How long did it take, on a timestamp you can see? Not an estimate, a start and a finish.
- How many separate pieces of work stand behind the number, and how many of them failed?
- Which evidence does each figure in the answer rest on?
- What was left empty? A field with no source should stay visibly empty rather than carry a plausible number.
If a tool cannot answer those four questions about its own run, the number it returns has no lineage. If you want to see what a run of ours returns on a property you know, tell us which one.