Scott Molluso · for Safeguard Properties
An agent layer for mortgage field services. Five agents in production against real guidance, every decision attributable to the exact rules in force when it was made.
Arrow keys or space to advance
The problem, in money
Over-allowable approval is decided after dispatch, so the crew is paid and HUD is not billed.
Two appeals exist per decision and the second is final. Rebuilding the chronology by hand means most are abandoned.
SCRA and PTFA routing depends on a contractor remembering a banner and dialling a number instead of typing a note.
Mortgagee Neglect runs from when a property should have been determined vacant, not when someone said so.
What exists today
Every figure on this slide is read from the database on load. Nothing here is typed into the deck, which is the same property the system claims for its own decisions.
The fleet
Coverage
Policy condition capture and exception routing.
Uniform Property Dataset validation at the point of capture, so submissions do not bounce on first-pass review.
Marketing readiness and disposition condition tracking.
This slide is on purpose. A coverage map with no gaps in it is not a map, it is marketing, and this is addressed to the people who would find them in the first meeting anyway. The gaps are the first ninety days.
How it is governed
System prompt
Task instructions
Behavioral constraints
Tool-use policy
Escalation paths
Every run stamps the exact versions in force when it happened. Six months later the question is never what the prompt says today, it is what it said that day. Without the stamp a trace is a log; with it, a trace is an audit record. It cannot be backfilled.
Changing the rules
Saving writes a draft. The live document is untouched.
The suite runs against the draft, in place of the live document of the same type.
The server refuses unless that exact version scored at or above the standing baseline.
Editing in place leaves a document live and untested for as long as the suite takes, and nobody can say afterwards which runs happened inside that window.
The action log
“Our log says we did it” is our word for it. “Reference AB-88431, check your system” is corroboration from a party with no stake in the story.
Three irreversible actions and one approval in a run left no way to prove which one was approved. Now each action names its own.
An agent that failed twice and succeeded on the third attempt used to look identical to one that succeeded immediately.
Whether an irreversible effect was later compensated, by whom, and when. A reversal that lives only in the other system is one this log cannot see.
Evidence
Some of those failures were deliberate. A gate that has never fired is a gate nobody has tested, so documents were degraded on purpose to confirm the suite catches it. Escalation is scored as a success outcome: an agent whose evals punish stopping learns not to stop.
What it costs to run
Cost per agent, per run, is on the agent detail screen. The point is not that it is cheap. It is that the number exists at all, attributable to a specific agent and a specific version of its rules.
Ownership
Molluso AI owns
The governance framework. Versioned documents, the eval harness, the action log, the control plane. It predates this conversation and knows nothing about mortgage field services, which is verifiable by grep.
Safeguard gets it licensed, free, day one.
Safeguard owns
The agents, the decision tables, the peril table, the allowables logic, the corpus, the screens. All of it domain work, all of it yours.
The fork
What the role describes, and what I would do first regardless. Integrate through Connect, one module in production, no platform decision required from anyone.
Used to be a three-year bet. What you are looking at was built in days. That estimate is from a different era, and it is the only thing I would actually ask you to reconsider.
Either way the first ninety days are identical: one service line, in production, governed, with the regression gate wired before anything touches a live order.