In the week since the call with Lydia, Marcus has been sending the contact two, sometimes three messages a day — a pace he hadn't planned, that came from having both datasets and understanding more clearly what the two together suggested. He'd started asking directly: Who are you. What does FMI stand for. What do you do with the data you collect. The replies had been specific about methodology and careful about identity, answering each question as asked without answering the question behind it. He has enough experience with that mode of communication to recognize it and enough patience to wait it out.
Tonight the laptop is open on the kitchen table at 9:20, the notebook on the left side where it's lived since February. He'd eaten earlier — the Thai place on Lincoln, ten minutes on foot, the kind of meal that's satisfying without being memorable. He'd been re-reading the last three exchanges from the contact, tracking the shape of the approach: lead with data, distinguish carefully between what the evidence shows and what it implies, never push toward conclusions Marcus hasn't reached himself. Intellectual integrity or very good calibration. The two aren't mutually exclusive.
The thread shows unread at the top, the timestamp 8:47 PM — he'd had the phone on silent through dinner. He checks the length before reading, a habit from case files, and this one runs past the first screen, longer than anything the contact has sent before. He reads it through once without stopping, taking in the whole shape of it before assessing any piece.
The message runs three paragraphs. The first acknowledges the work, his and Lydia's. What you've found independently is part of a larger pattern we've been documenting for three years. It's the first time the contact has used a plural. The second paragraph answers, directly and for the first time, the questions Marcus has been asking about identity. The third paragraph is a single sentence.
I think it's time you saw the full picture.
The name arrives first: Dr. Rena Vasquez. Then the title and the organization — Director, Futures Mapping Initiative — with a link below that Marcus opens in a second tab before he's finished reading the paragraph.
The FMI site is minimal. That tells him something too. A single-page design: the name at the top, a brief description, an About section, a methodology page, a contact form. No press citations, no advisory board, no publication highlights. Founded 2025. Staff of twelve. Oakland. The site describes its work using the same vocabulary the contact has been using in messages for weeks — decision-to-outcome chains, causal inference, longitudinal tracking — without specifying what FMI does with the patterns it finds or who the findings go to.
Three years. He sits with that number. He'd been doing this for three months and thought of himself as someone paying sustained attention. Vasquez's organization has been doing it since before the 15% Threshold Act was law, before the Henderson County override incident, before Marcus had opened his notebook to the first blank page. Three years and twelve people and a funded research foundation, watching the same gap he'd been watching from a desk in Corktown.
He goes back to the methodology page and reads it through again. Causal inference applied to AI decision outcomes across domains — parole, lending, hiring, insurance, educational admissions. Not a single sector but all of them, running in parallel, since 2025. The word domains appears three times and each time it covers both what Lydia tracks in mortgage processing and what he tracks in corrections and whatever a twelve-person team in Oakland tracks that spans both. He scrolls back to the message and reads the section he'd gone past quickly on the first time through.
You and your colleague each found the gap from within your respective systems. That's unusual. Most people who encounter the gap treat it as noise. You treated it as signal. We've been watching the forum for three years for people who treat it as signal.
He reads it again. He searches Vasquez the way he'd search anyone: name, institution, publicly available record, anything that complicates the initial picture. Credentials first. Full professor at Stanford before 2024, computational social science, fourteen years. The department's cached faculty page still has her photo — a woman in her early fifties, direct camera contact, the settled quality of someone who has been in professional photographs long enough to stop managing them. She'd left in the fall of 2024. The institutional record shows emeritus status and a final semester's sabbatical. No coverage, no controversy around the departure. She'd stopped being a Stanford professor and become something else.
Publication record: sixteen papers between 2015 and 2024. He reads the titles in order. The early ones are foundational — causal inference methodology, propensity score matching, applications to policy analysis. The later ones narrow. Detecting Structural Leverage in Decision Networks (2022). Outcome-Sensitive Decision Weighting: Evidence from Administrative Data (2023). Leverage Points in Networked Decision Systems: Implications for Algorithmic Governance (2024). He opens it.
The abstract takes two readings. The methodology section takes longer. He reads with the notebook open, pen in hand — not to take notes, but because the physical presence of the pen slows him to a useful pace. The paper is about what Vasquez calls leverage points: nodes in decision networks where small decisions produce disproportionate downstream effects. The examples in the paper are synthetic, constructed for the argument, not drawn from real cases. But the structure is the same structure he's been looking at in twelve COMPASS-NG decisions for the past eight months, and the same structure Lydia has been looking at in twenty-four UnderwRite denials.
She'd been asking the same question. In formal mathematical language, with a Stanford affiliation, peer-reviewed. Since at least 2022. The difference between his work and hers was the difference between a notebook and a published framework. But the question was the same question, and she'd been asking it longer, with better tools, and hadn't answered it either. That last part mattered. At 10:43 he closes the browser tabs.
He calls Lydia at 10:50. She answers on the second ring. "I was waiting," she says — he'd texted her at 9:30, one line, Got something, I'll call tonight, and she'd been up for it. They've developed a loose rhythm over the past week, checking Signal at similar times in the evenings after David and Amara are settled. He reads her the contact's message in full. Then the relevant sections of the FMI website: the description, the methodology summary, the founding year, the staff size. She's quiet through all of it, the listening quiet he'd learned to distinguish from processing silence. When he stops, she takes a moment before answering.
"Three years," she says.
"Three years before she reached out." The A/C runs a cycle in her house; he can hear it through the phone.
"She found us because we found the pattern," Lydia says.
"She found us because she was already watching for people who find the pattern." He'd worked out this distinction during the research — it's the difference between a response and a selection. "FMI_obs and outcome_base_r were on the forum before either of us posted. She wasn't responding to our work. She was positioned to receive work like ours." A pause.
"Which means she's found others."
"Her methodology page covers six domains."
Lydia's processing quiet again. Then: "Pull up the foundation's 990."
He'd been looking at academic publications. He opens a second tab, searches the IRS database for FMI under its registered name — Futures Mapping Initiative, 501(c)(3) — and finds the filing from 2026.
"Funding from the Baird Philanthropy Fund," he says. "2026 shows $2.3 million in grants, $1.1 million in operating expenditures." He reads the board section. "Four directors. Two I don't recognize, one is a former DARPA program officer, one is Vasquez."
"Baird is legitimate," Lydia says, and he can hear the mental-math quality of it — the quick internal database she carries that holds things like this. "Endowment-funded, not a single-donor structure."
"So she's not someone's instrument."
"She's her own instrument." A pause. "I want to see what she's tracked. The school district pattern — if she has three years and six domains, she has the same pattern I found in thirteen months. At scale."
"Yes." She asks the question he'd been holding for the end: "What would she want from us?"
He thinks about it. He'd been working the credential side and the funding side and hadn't answered this one yet. "We're case studies. Two people who found the gap independently, from different domains, and documented it with enough rigor to be worth talking to."
"That's not enough. She has researchers. She has twelve people in Oakland doing this as their job."
"We're also the ones who are inside the systems," he says. "She can track outcomes from the outside. We can tell her what the decisions feel like from the decision point." He can hear her consider this. "She's been watching for three years and she's choosing now to make contact. That's a decision."
"Yes," Lydia says. A beat. "Reach out to her."
After the call, the apartment settles to its late-night register. Michigan Avenue below, traffic thinning toward eleven. The refrigerator cycles on. The laptop screen has dimmed to half-brightness, the forum thread still open.
He opens the notebook, not to write. He turns past the first two blank pages — he'd left them empty at the start, the same way he leaves the front of any new notebook, in case there's something worth putting there later — to page three, where the entries begin.
Williams, D. AI approved. My call: deny. Outcome: watch. Feb 14.
The date sits above it in his handwriting: February 14, 2028. Eight months ago, sitting at this same table, the same window with the same view of Michigan Avenue. He'd written that entry because something in the Williams interview hadn't settled, and he'd wanted to track it, and writing it down was the only form of tracking available to him at the time. He hadn't known it was the beginning of a series. He'd thought he was being careful and a little obsessive, the way he'd been careful and obsessive in other phases of the work.
Terrence had told him to stay in his lane. Terrence wasn't wrong exactly — the lane was real, the lane had a logic to it, the lane protected the institution and, at the margins, protected the people the institution processed. But the lane didn't run where the question ran, and the question had kept running.
Janet had asked what he was really looking for. He'd said he didn't know. That was true in February and was less true now, though what he was looking for had kept changing its shape on him. He'd been looking for whether the AI was right. Then he'd been looking for whether the AI was right in a pattern. Then for whether the pattern was his domain only or broader. Now he's looking at three years and twelve researchers and a paper on leverage points in networked decision systems, which is a formal way of asking the same question he'd written in a notebook at a kitchen table in a month when the question still felt manageable.
He closes the notebook. The laptop screen wakes when he moves the cursor. He opens the message thread with FMI_obs and types a reply. The message is four words. Before sending, he adds Lydia's forum handle — delta_42 — to the recipient field, the first time he's included her in any communication that isn't between the two of them directly.
We'd like to see your data.
He sends it at 11:08. The timestamp locks. Vasquez is three hours behind in Oakland. By the time his message arrives, she'll still be at her desk.