Two weeks of messages from the anonymous contact, and Marcus has started keeping a second notebook.
He'd bought it at the Rite Aid on Michigan Avenue three blocks from his apartment — unlined, cloth-covered, the same brand as the one in his desk drawer at the office. That notebook stays between shifts. This one lives at the kitchen table beside the laptop, its pages filling with what the contact sends. The messages follow a consistent format. Domain. Decision pattern. The gap between stated criteria and observable outcomes. Time horizon. Insurance. Health risk assessment overrides clustering in counties with above-median educational attainment, ages 28 to 45, downstream event horizon 18 to 24 months. Hiring algorithm rejections, qualified candidates near founding-stage companies in specific sectors, elevated entrepreneurship rates in denied cohort tracked over 36 months. Educational admissions scoring exceptions, geographic distribution, relationship to residential patterns unclear. No names. No employers. No cities he could verify. He'd taken this as the terms of the exchange and written nothing back asking for more.
Three domains in two weeks, and in each one the same question underneath the surface detail: what is the system optimizing for that its stated criteria don't describe? He'd first asked it in February, watching the confidence percentage settle on 73% approved for Dewayne Williams. He's been asking it in different forms ever since. The contact asks it at scale, and has been asking it for longer than Marcus has been writing case names in notebooks. What the contact's messages have built, across the three domains, is a picture of something wider than a single system glitch or a single period of bad training data. The time horizons — 18 months in insurance, 36 months in hiring — are longer than any standard review cycle. The contact doesn't say this directly. The contact sends data and lets Marcus read what's there.
He'd sent the direct question two days ago: Who are you? The response had come at 11:42 PM. He knew the exact time because he'd been watching the interface with the laptop open on the kitchen table, a mug of coffee he'd forgotten beside it. Someone who has been watching longer than you have. He'd typed two follow-up questions and deleted them both. You don't push at a non-answer when the person delivering it is also deciding what comes next. He sat with the second notebook open to a blank page and listened to the refrigerator, the traffic on Michigan Avenue, the quiet that settles into a one-bedroom apartment after midnight.
The next morning the link arrives, the message reading: Read this. Forum thread. He follows the URL to the Outcome Tracker before his coffee has finished brewing. The thread is eight days old, posted by a handle he doesn't recognize: delta_42. Title: D-42 cluster, Phoenix area, school district correlation — mortgage/underwriting experience?
He reads the thread start to finish without opening another tab. The poster has done serious work — mortgage denial dates mapped against school enrollment deadlines, a specific DTI band isolated above baseline, twenty-three cases controlled for income and comparable property values. Her methodology is careful in the way that distinguishes someone worried about being wrong from someone who's already decided they're right. She's held back where the evidence is thin. She hasn't held back where it's solid. Her final post: The D-42 code indicates property risk. I cannot locate the property risk in these cases. Standard underwriting criteria don't explain the pattern. I'm not claiming intentionality — I'm documenting what the data shows. Has anyone in other domains seen a similar structure? Decisions diverging from stated criteria, clustering near educational or developmental access gaps? It's the same question he's been asking from a corrections database for fourteen months, translated into the language of DTI ratios and denial codes. He'd been asking it in case files and risk tiers. She's asking it in mortgage denial clusters near school district boundaries. The domain is different. The gap is the same. He goes back to the thread and reads the analytical posts a second time, tracking the math.
He opens a private message window, his username in the field: casefiles_313. He'd put that in when he registered four weeks ago without thinking too hard about it — the 313 is Detroit's area code, the casefiles is what fills his desk and his notebook. He types the recipient: delta_42. He writes: Your question matches a pattern I've been tracking in a different domain — not mortgage. Twelve cases over fourteen months. I'm saying: same structure of the question. He reads it back, deletes the second sentence, which over-explains, then adds: If you're willing to compare notes — no names, no employers — I'd like to understand whether this is one question or two.
He doesn't usually reach out. The investigation has been private by design — a notebook practice, conducted alongside regular duties, not shared with Janet or Terrence or anyone in his professional orbit. Posting on the forum four weeks ago had already crossed one line. This is another one. He reads the message a third time and decides it says what he means, sends it, and closes the laptop. Outside, Michigan Avenue is quiet. He finishes the coffee, which is still hot.
Lydia has been on the forum for eight days. She'd created a reading account the night after the meeting with Elena, spent three evenings absorbing threads — mortgage processors from six states describing variations of the same structure, insurance clusters, a parole case involving a drowning rescue that she'd returned to twice because the shape of it matched her own work. Then she'd made a second account, delta_42 — the denial code that started all of this, made into a name she could commit to — and posted her analysis. The DTI band, the school district proximity, twenty-four cases stripped of any identifier that could point back to her company or her city.
The forum had answered. A mortgage processor in Oregon with an anomalous cluster near two STEM magnet districts, no explanation that held. A processor in Ohio with a different denial code but the same unanswerable gap. An insurance adjuster in Atlanta who'd been sitting on a zip code anomaly for seven months. She'd cross-referenced the Atlanta data against her spreadsheet. The patterns came from the same structure of question.
Then: casefiles_313 to delta_42 — the message had arrived three days ago while David was putting Amara to bed. She'd looked at the handle before opening it — the 313 is Detroit's area code, which she registered the same way she registers zip codes and DTI cutoffs, as a number that came with location attached. She read the message — parole, not mortgage, twelve cases over fourteen months, same structure of the question — and replied: Received. AI decision doesn't match stated criteria, outcomes exceed the stated risk parameters. My cases involve families with school-age children denied near above-average educational access. Is your pattern population-specific?
His answer came within the hour. Population-specific, yes, but different structure from your cluster. Single-individual cases. Clearest example: thirteen days between decision and observable outcome. The outcome was tied to the individual being present at a specific location at a specific time. Human assessment would have prevented that presence.
Thirteen days. She'd calculated before she finished reading the sentence — thirteen days in his case, eight days in her clearest case, a drowning against an enrollment window. She typed: My pattern is temporal as well. There's a window between the denial date and an access deadline. Clearest case: eight days.
He wrote back: Temporal precision is the variable the stated criteria don't include. She considered this. It wasn't wrong, but it wasn't complete either. She sent: Or the system is wrong in a consistent direction. Both explanations produce the same gap between what it claims to optimize and what the data shows.
They'd exchanged five messages over two evenings, each one adding a piece, neither one claiming more than the data warranted. By the third exchange she understood he was in corrections. By the fourth he understood she was in mortgage processing. Neither had said so plainly, because neither needed to — each was giving the other enough to verify the pattern without giving enough to close the loop of identification.
What they kept returning to: the question of mechanism. Whether the gap between stated rationale and observed outcome was a property of the AI system itself, or of the training data, or of something harder to categorize — whether the system had weighted some feature of consequence that neither its stated criteria nor its documented methodology described. They came at this from different angles across three exchanges and set it down each time without resolving it. By the end of the fifth message, the agreement was that they were asking the same question.
I have a dataset, she typed on the second evening, after Amara was down and David's reading light was off and the house had gone quiet enough that she could hear the pool pump cycling three houses down. Twenty-four cases, structured by denial date. If we're comparing notes I'd rather not do it through this interface.
His Signal username arrived twelve minutes later. She added him, watched the confirmation settle at the top of the chat window: Messages to this person are end-to-end encrypted. She took a screenshot — not because she was certain she'd need it, but because Elena Hernandez had kept copies of everything and that had seemed, in retrospect, like the correct instinct. She set the phone face-up on the desk beside the binder. David was asleep. Amara's nightlight threw its bar of orange under the bedroom door. The A/C ran a cycle and stopped.
She typed: My spreadsheet is twenty-four rows, denial date through school district proximity. I've controlled for income but may have missed something. What's your structure? The typing indicator appeared on his end. She watched the three dots and kept her hand flat on the binder's cover, the brochure still inside with its box of blue text, the six prerequisites, the enrollment window. The pool pump ran and stopped. His message came through, and she read it.
Then she opened the spreadsheet.