This is what a Margin Review actually found.

Each one points at a record you can open.

Three different kinds of finding, taken from the ledger as written. The dollar figure is what Reeve priced the finding at. The trace is where it came from. That trace is the whole point, because a finding nobody can check is just a claim.

The money at risk was bigger than the money owed.

Inside Medicaid, this agency's collections are clean. Nearly everything it billed, it was paid, and only six invoices were still outstanding. So this is not a report about lost revenue. The dollars worth talking about were going the other way: care delivered past what the state had approved, billed, and now exposed to recoupment. Nothing inside the billing system was wrong, which is why nothing inside the billing system caught it.

• Care delivered past the authorization, and billed: $3,832.06. Across 112 authorization weeks, delivery ran past the cap. In 85 of them the extra hours were billed too, and that is the set at risk, 136.11 hours, each week priced at the schedule in force on its own dates of service. The other 27 weeks ran over but were never billed, which costs nothing and recoups nothing.
• Four Medicaid visits that failed EVV to the aggregator. They deny until the aggregator record is fixed. Four, out of thousands. Whatever else is happening here, EVV capture is not it. An earlier version of this page put a dollar figure on those four. It has been removed: these exports carry no remit data, so whether those visits were ultimately paid is not knowable from this book, and the software refuses to price them for exactly that reason. The page now says what the software says.
• Five past-due Medicaid invoices, $2,446. Submitted and unpaid, aged 22 to 43 days. Ordinary receivables, listed for completeness, not as a finding. Any aging report shows these.
• 284 Medicaid visits missing either a client or a caregiver signature, roughly $25,000 of care. That dollar figure is an estimate and deliberately rounded: the signature file and the billing file share no visit-level key, so it is scaled, not joined. Not a loss today either way. It is the document an audit samples first, and it sits outside the $3,832 because no dollar of it has gone wrong yet.

Four corrections, and how they were caught. An earlier version of this page put the over-cap figure at $1,373. The authorization file measures in hours, and the rate that priced it was a fifteen-minute rate, so the finding came out roughly three times too small. Separately, this page said six past-due invoices totalling $2,696 and 972 invoices reviewed. Both were all-payer counts on a page that is a Medicaid-only read: one of those six invoices, $250 of it, was private pay, and 429 of those 972 invoices were not Medicaid at all. The Medicaid figures are five invoices, $2,446, out of 543. Then the over-cap figure moved once more, from $3,834 to $3,832.06. The first repair had priced all 136.11 hours at a single blended rate; the state publishes a new schedule on set effective dates, so each of the 85 weeks is now priced at the schedule in force on its own dates of service. Two dollars is not the point. The point is that a hand-built number and a re-runnable one disagreed, and the re-runnable one won. Each error was found the same way, by re-reading the source files instead of the summary, which is the only method this page argues for. The second was found by software written to check the first, the third by software written to check the second, and the fourth by software written to check the third.

The fourth is the overlap finding further down this page, and the software that caught it was written to ship it. Turning that finding into a detector meant stating the rule in code, and the code disagreed with the page. The original measured an overlap as the previous visit's end minus the next visit's start. That is how far into the previous visit the next one began, and it equals the real overlap only when the second visit ends after the first one does. When a short visit sits inside a long one, it overstates. One pair on this book was published at 570 minutes when the two clocks actually overlapped for 255. The total was 38.75 hours and is 28.00. The worst single overlap was 9.5 hours and is 4.25. The dollar figure was roughly $1,097 and is roughly $793. Two more errors sat underneath that one. The comparison only ever tested a visit against the one immediately before it, so it never saw an overlap between two visits that were not adjacent, and it undercounted: 22 is 23. And the payer test read one side of each pair instead of both, so “18 of 22 with state Medicaid on both sides” is 12 of 23. The correction cuts the hours by more than a quarter, cuts the worst case by more than half, and takes six pairs out of the Medicaid-on-both-sides claim. Only the instance count moved the other way, and one more conflict in a caregiver's day is not a better number for the agency that has to answer for it. The published version was the stronger finding. This one is the correct one.

Then they handed over several more months.

Several consecutive months of the same book, read the same way. Medicaid EVV visits lined up against Medicaid billing rows. A single month tells you what a book looks like on one day. Several consecutive months tell you whether what you found repeats, and two things came out that a single month could not show.

One thing to note before the numbers: the agency grew over the period, and its monthly visit counts rose as it did. That matters for the rest of this section, because the volume rose while the rates below did not move.

About a third of visits carry a manual edit, every single month.

An EVV record carries a manual edit reason code when someone corrected it after the fact. A phone died mid-shift. A caregiver forgot to clock out. Somebody fixed the record so the visit could be billed. Reeve counted those records as a share of each month's Medicaid visits, and the share barely moved across eleven months.

• The share of visits carrying a manual edit held near 32% every month, in a narrow band. Visits with no client signature ran between roughly 8% and 15% over the same period.
• None of this is money at risk. These visits cleared the aggregator and were paid. Reeve did not price them and will not. A manual edit means a record was corrected, and corrections are a normal part of running EVV.
• We do not know whether 32% is high, normal, or low here. Manual-edit rates in the 20% to 40% range are commonly reported across the EVV industry, and what we have is one agency's book with no comparison set behind it.
• What the read shows is that the population never empties. It regenerates at close to the same rate every month, which is what separates a cleanup that finishes from something that has to be watched continuously.
• That population is what an auditor samples from. Roughly one visit in three carries an edit, so roughly one visit in three is a candidate for a records request. That is the honest claim here and the only one this data supports.

The records say one caregiver was in two places at once.

Across the period, roughly two dozen times, the same caregiver ID is clocked in to two different clients at two different addresses with overlapping visit times. Both sides were billed. Nothing in the billing workflow objects, because each visit is individually well formed. The conflict only appears when the caregiver's whole day is read against itself.

Instances occurred in most months across the period, with no month carrying more than a handful.

• The cause is unknown. Reeve read records, not intent. Ordinary explanations fit this pattern: an administrator entering visits on a caregiver's behalf, a shared device or a shared login, a scheduling system writing records on its own. None were ruled in and none were ruled out. What the files say is narrower than any of those. One caregiver ID is recorded at two addresses at the same time, and both were billed. The records contradict each other. Which side is wrong, or whether the record is simply wrong about who was there, is not a question a file can answer.
• Six of the 23 ran over an hour, totaling 17.75 hours. These are not rounding artifacts at the edge of a shift. The largest single overlap is 4.25 hours.
• Two caregivers account for 13 of the 23. The pattern concentrates. Of the nine others, one appears twice and eight appear exactly once.
• The money here is trivial, and the engine would not compute it. Reeve prices only against a published fee schedule, and an observed average is not one, so it refuses this finding outright. The figure that follows is therefore a hand calculation, and it is weaker for it: 28.00 hours at the agency's own blended billed rate for that period comes to roughly $793 across the read. Against the year's billings, that is a rounding error. What matters is what a records request does when it lands on a visit whose own timestamps contradict another visit on the same caregiver's day.
• This agency's billing team is good. Collections are clean and nothing went unbilled for eleven months. They caught none of these 23, and that is the finding. It is invisible from inside a billing workflow, because a workflow checks each claim against its own authorization and never checks a caregiver's day against itself. This is not carelessness.
• Knowing first is the whole value. The agency is the exposed party here. Every one of these is answerable in a quiet week with the records still in hand. The same 23 are a much harder conversation when a payer opens with them.

Four ways the overlaps could have been nothing.

A finding like this is worth more for the explanations it survives than for the number it lands on. Reeve went at the 23 four separate times looking for a boring reason they existed. All four came back empty. The attempts are here so you can judge whether they were the right ones.

Four read-only exports and about twenty minutes.

The agency pulled its own exports from its own system and removed the identifiers before any of it reached Reeve. Reeve read the exports and wrote nothing back. There was no integration and no access to the agency's software. There was also no contract, because this was not a client engagement. It was an operator handing over a stripped book to see what a reconciliation would say about it. A paid engagement works differently: Reeve is a business associate under HIPAA, and a business associate agreement gets signed before a file moves. The eleven-month read went the same way, on exports the agency pulled itself.