The laptop runs warmer than her office machine, which she's mentioned to David twice and which he keeps meaning to look at. She has it propped on the pull-out shelf of the spare room desk — a room that's technically a home office but doubles as storage for the stroller they haven't returned yet and a row of boxes from the last move. The A/C holds at sixty-eight degrees, same as the processing center in north Phoenix, which means she's wearing a cardigan at her own desk on a March evening with her daughter asleep down the hall and David in the living room with something on low — dialogue she can hear through the wall but can't parse, a laugh track at intervals. She has two hours before her focus runs out.
She opens the work portal with her HomeFirst credentials and pulls Q4 archived denials, sorted by denial code. D-42: forty-one files. She creates a spreadsheet on her personal drive — not the work system — and starts building columns: denial code, property address, neighborhood index score, applicant profile notes, apparent property risk. A Y or N at the end of each row.
The first file is a property in a recorded flood zone off Cave Creek Road — FEMA designation public record, flood premium alone making the carrying cost unworkable. Y. Incomplete foundation documentation in Chandler, dispute pending at the county assessor's office. Y. Unresolved HOA special assessment. Y. She works through the queue — flood zones, title complications, environmental reports pending — the denial code doing what it was built to do, one after another.
She gets through twenty-six before her eyes stop cooperating. She saves the file and closes the laptop.
The second evening she finishes the remaining fifteen. Thirty of the forty-one D-42s have Ys. Eleven have Ns. She goes back through the eleven, cross-referencing county records in a second tab. Three give up their explanations on closer review — risks she'd weighted too lightly, problems she'd scrolled past. She moves them to Y. Eight remain.
She rereads them and pulls six more from the original thirty — files where the cited risk existed but wasn't clearly dispositive, where she'd given benefit of the doubt that scrutiny doesn't sustain. She moves them. Fourteen total. Fourteen D-42 denials with no property risk she can find in the available data.
She's kept the Hernandez file in row two since the beginning. She looks at the column of Ns, and on the third evening she maps them.
She downloads a GIS shapefile for Phoenix metro school district boundaries from the state Department of Education data portal — public record, available for anyone who looks for it — and opens it in a browser mapping tool she used at ASU for a spatial statistics elective. She imports the fourteen addresses. They appear as orange dots distributed across the metro, and the distribution isn't random; she can see the shape before she overlays anything. Three rough clusters, each settled into a different quadrant of the valley. She adds the school district layer and the map updates: the clusters resolve onto boundaries.
Eleven of the fourteen properties sit within a quarter mile of three districts. Two run magnet programs — one focused on STEM, one on performing arts. The third has a dual-language immersion curriculum with a documented waitlist. She tabs to the applicant profiles and checks them in order, looking for one thing: children. Every household has at least one child of school age. All fourteen.
She does the count before reaching for any tool. Fourteen applications, fourteen families with children, fourteen properties near specialized school programs. Fourteen D-42s with no apparent property risk. She opens a chi-square calculator in another tab. She uses the quarter's full D-42 dataset as her base population — forty-one denials — and tests proximity to magnet school districts against denial with no visible property risk. The p-value comes back at 0.031. She writes it on a sticky note beside the laptop and sets the pen down.
Fourteen is a small sample and 0.031 is not conclusive. She knows this. But if the distribution were random, she'd expect to see this clustering by chance less than four times in a hundred, and 0.031 is not the number she'd have predicted on the first evening when she pulled forty-one files because she couldn't let the Hernandez application go.
Fourteen families. The orange dots map to addresses and the addresses map to files and the files map to people: a teacher couple applying for a property in Tempe near the STEM magnet, a nurse and an IT administrator in Mesa, the Hernandez family in their south Phoenix rental applying for the house in Gilbert with its cul-de-sac and its science program. All of them qualified on paper. All denied for property risk the property records don't support.
She knows what algorithmic bias looks like. She attended the company's fair lending elective in 2026, after the California settlement, and read enough of the supplemental materials to understand the mechanism: the model trains on historical lending data, and historical lending data reflects decades of decisions that systematically favored certain borrowers and penalized certain zip codes. The bias compounds without announcing itself. A family with clean financials gets denied because the model learned from old data that their neighborhood's default correlation is elevated — even if that correlation traces back to the deliberate withdrawal of capital from that neighborhood two generations earlier. Disparate impact. Not intent. Effect.
This isn't that.
Disparate impact in lending denies access based on where you are. These fourteen families aren't being penalized for their current neighborhoods, which are working-class to middle-income, nothing in the data that should flag as risk. They're being denied the ability to move to neighborhoods with specific school programs. The direction runs the other way. Whatever the model is seeing, it isn't seeing these applicants as risks based on their origin. It's seeing their destination.
She sits with this and works through the other explanations. Maybe there's property data feeding UnderwRite that she can't see from the case files — development zoning disputes, infrastructure assessments, speculative investment flags ingested through a variable pipeline she's never seen documented. That's possible. UnderwRite Pro's full variable set is proprietary; what populates in the verdict field is not a window into what produced the decision. Property risk in the model's terms might mean something narrower, or wider, or different from property risk in hers.
Or it's fourteen coincidences. A cluster that formed around school districts the way any scatter of fourteen points might cluster if you overlay enough GIS layers with a willing eye. The p-value says it's probably not. Probably is not the same as certainly.
She knows her own failure mode. Pattern recognition is her skill and her liability, and the two are the same capacity. She can find patterns that aren't there; she's done it before. Early in her career she'd convinced herself she'd found systemic undervaluation in a specific zip code corridor — built an explanation around a real cluster that turned out to be a coincidence of the appraisal cycle. Her manager had been kind about it. The pattern was real. The story she told about the pattern wasn't. She looks at the map, the orange dots sitting near their school district lines, and can't unknow what the p-value said.
She is on application 28 of 44 when the next D-42 comes through on Wednesday. The counter in the upper right of the UnderwRite dashboard resets on Wednesdays; she's been watching it tick upward since nine AM without registering it consciously — background, automatic. Application 28 is a family in Scottsdale. Two adults, two children under ten. They're applying for a four-bedroom near the Arcadia district's K-12 performing arts school. DTI of 36 percent. Credit scores: 724 and 709. Down payment at 22 percent LTV. One borrower is a pediatrician; the other is a freelance graphic designer with three years of stable self-employment income documented on the Schedule Cs. The file is meticulous — tabbed, chronological, nothing missing. UnderwRite Pro processes for eleven seconds. The verdict field populates: D-42. Property risk.
She clicks through to the property data. 1998 construction, updated in 2019. Clear title. No flood zone designation; she checks the FEMA map anyway, out of habit now. The county environmental records show no hazard flags. She runs the assessor's record in another tab: appraised value consistent with comparable sales in the corridor for the past eighteen months. The appraisal notes in the file are two pages, routine, no exceptions noted. Nothing.
She processes the denial. Generates the adverse action notice, queues the file for document retention. Two minutes at her standard pace. That evening she adds the row to the spreadsheet. Scottsdale, 36% DTI, two kids, performing arts district, D-42, no visible risk. N in the final column.
Fifteen.
She reruns the chi-square with the updated count. The p-value drops to 0.024. On Friday afternoon Ray Soto stops at her desk.
He stands just past her right shoulder — his habit for performance conversations. Close enough that she knows he's there; far enough that turning to look at him requires a choice. She closes the compliance review she's been working and turns.
"Your output this week." He glances at his tablet, though she's certain he knows the number without it. "Running about eight percent below your quarterly average."
"I know." She keeps her voice level. "A couple of complex files slowed me down. I'll make it up before month-end."
Ray nods. He's not flagging her for discipline — she can read that in what he doesn't say and in the pace of the conversation, which is brisk rather than deliberate. This is Ray operating from the dashboard: throughput below average, manager reflects the metric, employee acknowledges, conversation ends. He asks if everything is okay. She says fine, just those files. He says something about the end-of-month projection and is already two steps away when he finishes the sentence.
She turns back to her monitor. Thirty-two completed files this week against her personal average of thirty-five — she calculated it while he was talking. 91.4 percent of standard. A blip. Her performance record at HomeFirst goes back four years, no write-ups, a regional commendation in 2026 for highest accuracy rate in the division. One blip will not define the quarter.
The spreadsheet with fifteen rows and a column of Ns is on her home laptop. Not this machine. She checked the employee handbook policy on Monday, during lunch, looked it up without appearing to look: the work system logs file access for active cases and recent closures, but archived denials outside the case-handling workflow log only to the standard audit trail. She's accessed archived files before. It's within the scope of her credentials. She has not opened the spreadsheet on the work network and won't.
She pulls the next application into the queue. UnderwRite Pro processes for nine seconds. The verdict field populates green: APPROVED.
She reviews the file. She confirms it.