Here's the scene the last field note left you with, except this time you have the records.
It's a Tuesday in the back half of the quarter. Someone in finance pulls the monthly write-off report and the line is bent the wrong way. Not catastrophically. Just enough that the number doesn't match the forecast, and the kind of person who notices these things has noticed.
The question lands in your inbox by mid-afternoon, phrased the way these always are: can you tell us what changed?
In the world without records, you know exactly what comes next, because we walked through it last time. You pull cases. You read traces. You re-run a handful and they mostly land in the same place. Hours go by. You end up writing a doc with the word believe in it.
This time you open FoxCommand and you ask the records a question.
Step one: isolate what governed the decisions
Eleven weeks ago, auth_policy went from v1.3 to v1.4. Buried in that bump, the escalation threshold for one claim class moved from $500 to $750. It was a reasonable change. Someone approved it for a reasonable reason. Nobody connected it to a write-off report that didn't exist yet.
You don't have to reconstruct that history from memory, because every decision since carries the policy version that governed it. You filter the population by policy_version and the cohort separates cleanly: decisions made under v1.3, and decisions made under v1.4. No guessing which change is the culprit — each decision already names the regime it was decided under.
The trace tells you what the model did. The record tells you which rules it was doing it under. Only one of those lets you sort ten thousand decisions by cause.
Ten thousand decisions. The report says something moved. The dashboard can't tell you which ones.
The point: each decision already names the policy it was made under. No guessing which change is the culprit.
Step two: estimate the blast radius
Now the question finance actually wants answered: how big is this?
Because the records are governed artifacts and not log spray, this is arithmetic, not archaeology. You take the v1.4 cohort and re-check each captured decision against the old $500 boundary. Every decision where the claim amount sits between $500 and $750 is one that would have escalated under v1.3 and instead auto-resolved under v1.4. That set — exact, countable, dollar-weighted — is your blast radius.
Not “we believe the drop is driven by.” A count. With a dollar figure. And a list of case IDs you could hand to anyone.
2,860 decisions under v1.4
2,860 decisions were made under v1.4. How many actually moved?
The point: this is the sentence the verification-debt note said you couldn't write. Here it is, with a number in it.
Step three: simulate the rollback
The last question is the one that used to be impossible: if we put it back, what happens?
You simulate a rollback to v1.3 — and this is the part worth being precise about, because it's the whole reason the records exist. FoxCommand re-evaluates the preserved decision records against the old threshold. It does not rerun the model. The captured input shape, the retrieved documents, the evaluator output — all of it is frozen at the moment each decision was made, so re-checking it against $500 produces the exact counterfactual outcome, with zero probabilistic drift.
Each affected case flips from its v1.4 outcome (AUTO-RESOLVE) to its v1.3 outcome (ESCALATE), deterministically, because nothing about the decision is being regenerated — only the threshold it's measured against is being swapped back.
That's a real rollback estimate. Not a rerun that resembles the past. A re-measurement of the past as it was actually recorded.
A rerun asks the model to guess again. A simulation asks the record what it already knew. One drifts. The other doesn't.
Records frozen at decision time. Nothing here will be regenerated.
The point: a rollback you can defend, because nothing was regenerated to produce it.
What you send back
By end of day you reply to finance with three things you could defend in a deposition. The cause: the v1.4 threshold change on this claim class. The blast radius: N decisions, $X exposure, case list attached. The remedy and its cost: rolling back to v1.3 re-escalates N decisions, here's the operational load that creates, here's the net.
No believe. No educated guess assembled from partial evidence. The investigation that used to eat a week became a query that fit inside an afternoon, because the work of reconstruction was done eleven weeks ago, automatically, at the moment each decision was made.
That's the difference a decision record makes the day someone finally asks.
Building workflows you'll have to explain a quarter later? That's what this is for.