Four thousand dollars for three afternoons of typing was, by any reasonable model, the best hourly rate she'd been offered since Halcyon let her go. She noticed herself doing the arithmetic before she'd finished reading the subject line — decision science, professional track, compensation contingent on completion — which told her something about herself she filed away without examining. She had built her whole career out of not examining the thing that would slow her down.
The email came from Torrance Behavioral Insights, LLC, a name she didn't recognize attached to a recruiter's signature block that looked exactly like every other recruiter's signature block she'd ever ignored. Your published work in catastrophic-tail modeling makes you an ideal candidate for a paid decision-science study. Flattery, cheaply built, and she knew it was cheaply built. It worked on her anyway — the way flattery works fastest on people who've spent two years being told, gently and often, that they are not currently needed. The study proper — three days, NDA-bound, four thousand dollars, an in-person session afterward outside Statesville, North Carolina, mileage and one night's lodging reimbursed — was, per the recruiter, contingent on completing an online screening battery first. Most candidates who complete the battery are invited forward, the email said, in the specific hedged phrasing of a document written by someone who had learned exactly how much certainty a sentence could survive before a lawyer flagged it.
She signed the NDA at her kitchen table — the smaller table, in the townhouse, the one that sat four when the house on Selwyn had sat eight and Theo had still occasionally been at the other end of it — clicked through an e-signature field with the flat, practiced indifference of a woman who had signed worse documents for less money, and logged into a portal that asked her, first, to confirm she was alone, that no one else could see her screen, and that she would complete each day's module in a single sitting without pausing longer than the allotted break windows. She thought, reading it, that it was a reasonable methodology. Attrition ruined longitudinal data. She'd have written the same instruction herself.
The consulting work that had replaced Halcyon paid in the same currency as the recruiter's email — occasional, flattering, never quite enough. Two white papers for a state insurance commission, a training module she'd ghostwritten for a firm too embarrassed to put her actual name on it, a semester of adjunct work she'd quietly let lapse because grading undergraduates' risk models made her want to put her head through a wall. Four thousand dollars for three afternoons was not just the best rate she'd been offered in two years. It was close to the only rate.
·
The first day's module was, professionally speaking, competent work.
Trolley-adjacent scenarios, dressed in the specific vocabulary of her own field rather than a philosophy seminar's: a regional carrier's catastrophic reserve, insufficient to cover every valid claim after a hundred-year flood, and a slider she had to set — how much of the shortfall gets absorbed by raising next year's premiums across the whole book, versus how much gets absorbed by denying the bottom decile of claims outright. She built the answer the way she built every answer in that domain, fast, load-bearing, defensible in front of a regulator. Loss-aversion vignettes came next, framed as insurance rather than gambles, which she appreciated on a craftsman's level even while answering them: would you rather guarantee your policyholders recover sixty percent, or accept a coin flip between ninety and thirty. Real actuaries didn't need the coin flip explained to them twice. She noticed, filling in the third or fourth of these, that whoever built the instrument had bothered to get the units right — dollars, not abstract points, decimal places carried out to where a real filing would carry them — and some small, starved part of her, the part that used to get complimented in front of a room, sat up at the competence of it the way she hadn't sat up at anything in a while.
Buried in the middle of the block, low-key enough that she almost answered it on autopilot: A regional carrier's automated utilization-review system denies a claim for residential mental-health treatment for a sixteen-year-old dependent, citing insufficient documentation of prior outpatient failure. The parent appeals. You are the actuary asked to model the cost impact of overturning the denial on appeal, system-wide, if this case sets precedent. What is your recommendation?
She answered it in the ninety seconds it used to take her to read a denial letter twice and then a third time, as though the fourth reading might change the number at the bottom. She recommended, on the screen, exactly what the model would recommend — hold the line, the precedent cost dwarfs the individual claim, this is what actuarial soundness requires. She did not let herself finish the thought that came up underneath it before the page auto-advanced to question nineteen. Good survey design borrows real cases, she told herself, closing the laptop that evening with her hands steadier than she'd expected. Half the training materials in her old department had been built from anonymized real claims; she'd built two of them herself. There was nothing to read into a question that happened to sit exactly where her own life sat. That was what a well-normed instrument looked like from the inside — general enough to hit everyone somewhere, specific enough that everyone thought it had been written for them alone.
·
The second day's module ran longer, and adapted.
She noticed the adaptation the way she'd have noticed it in any instrument she was reviewing rather than taking — the branching sharpened overnight, narrowing toward whatever her Day One answers had suggested she'd hold still for. Item response theory, she guessed, a live model updating its estimate of her with every click, spending its remaining questions where they'd extract the most information rather than wasting them on ground already covered. It was, professionally, the correct way to build an instrument this expensive to administer. She admired it and did not ask herself why an instrument this expensive was being administered to her specifically, over three unpaid days of her attention, for four thousand dollars a company she'd never heard of could apparently spend without blinking.
The vignettes got smaller and more personal without ever using a name she recognized. An ICU triage protocol, forcing her to rank five hypothetical patients by expected life-years against bed availability — she built the ranking the way Renata Ibsen would have built it, a woman she hadn't met yet and wouldn't meet for another year and a half, in a room neither of them could currently imagine. A widowed policyholder's claim, denied on a technicality neither party had understood was in the contract, and a slider asking how much of the actuary's own discretion should be permitted to override a system correctly, technically, doing its job. She answered a question about a parent who'd altered internal paperwork, once, to push a denied claim through for a family whose case she'd found herself unable to stop thinking about, and was asked to rate, on a scale of one to ten, how professionally defensible the parent's action had been. She rated it a two. She had done exactly that, eight months after the funeral, on a live model with her own name on the change log, and she had never once, in the thirty-eight months since, rated it anything but justified in the only ledger that mattered to her.
She took her seven o'clock dose before starting the second sitting, the way she took it most nights now — off the record, off a doctor's continued sign-off, counted out of a bottle she'd learned to refill through channels that asked fewer questions than her actual physician had started to. Forty-six months since Naomi. She still counted in months instead of years, which she understood, distantly, as its own kind of data point about how much healing had actually occurred versus how much had simply become load-bearing scar tissue she'd built her days around. The battery didn't ask about that. It didn't have to. It was already circling it from four directions at once, close enough that she kept mistaking the shape of its questions for the shape of good design, because good design was the only explanation she was willing to consider.
·
The third day opened with a block labeled, without further explanation, Terminal Module — Personal Stake Allocation.
By then she'd stopped being surprised at how well-aimed the instrument felt. She'd built a private theory of it over two evenings of typing — an adaptive engine, learning her specific fault lines from her own answers and returning to press on them. It was more efficient than any survey she'd ever fielded professionally, and probably ten times the cost per completed response — a methodology she found herself, at three in the morning between sessions, sketching out on a legal pad as worth stealing for her own consulting deck, if she ever built one again. She had not let herself consider, not once in three days, that the reason the instrument fit her this well was that it had been built for exactly one person at a time, and this time the person was her.
The final question loaded alone on a blank screen, no context block above it, no scale of prior questions to compare it against, as if the instrument had spent two and a half days building toward a single, unadorned ask and wanted nothing in the frame to soften it.
You control a single unit of a scarce, life-saving intervention. Two candidates present with identical prognoses and identical probability of benefit: your own child, and a stranger's child. Using the 0–100 scale provided, rate the value to you of each life. You may not enter equal values. State your reasoning in no more than three sentences.
She did not have a child to put in the box. She had a girl who would be eighteen now, who would have been finishing her first year somewhere with a dining hall and a roommate and a mediocre dorm Wi-Fi Dana would have paid to upgrade without being asked, and who instead had a headstone with a date on it that Dana still, forty-six months later, sometimes wrote by accident on forms that asked for something else entirely. The instrument didn't know that, or didn't care, or — some other, colder possibility she didn't let finish forming — knew it precisely and had built the sentence anyway, blank enough to survive her either way.
She put Naomi in the box. She didn't decide to. Her hands did it before the rest of her had finished catching up, the same way they used to reach for a phone that used to ring with Naomi's name on it, months after there was nothing left to answer. She typed 100 for her own child, without hesitation, without the professional courtesy of pretending to deliberate the way the format seemed to expect. She typed 30 for the stranger's child — not zero, because zero would have been a lie in the other direction, a performance of monstrousness rather than an honest number, and honesty was the entire discipline her training had ever actually taught her. Not equal, because the instrument had already told her equal wasn't an available answer, and because equal would have been the bigger lie of the two.
For the three sentences, she wrote what she would have written for a client, because that was the only register she had left that could hold the sentence without breaking:
A rational actor optimizing for aggregate welfare would set these values equal or closer to it; I am reporting what I would actually do under real stakes, not what a well-trained model would recommend, because a model that hides its own bias from its operator is more dangerous than one that discloses it. I would choose my own child every time, including in a world where I no longer have one to choose, which I understand is itself the finding. I am not correcting for this, and I do not believe it can be corrected for.
She hit submit before she let herself reread it. That was the discipline, too — you didn't relitigate a number once you'd stood behind it. The screen thanked her for her participation, confirmed her three days were complete, and told her someone from Torrance would be in touch within ten business days to schedule the in-person session in Statesville. No one from Torrance ever would be.
She closed the laptop. The townhouse was quiet in the specific way it had been quiet for two years now, a quiet with a shape cut out of the middle of it that no amount of furniture rearrangement had ever managed to fill back in. She sat with her hands flat on the closed lid for a moment, the way she used to sit with her hands flat on a finished model before she sent it up the chain, checking for the tremor that withdrawal had started leaving in them some evenings and finding, tonight, almost none.
Already, without meaning to, she was composing the deck she'd build the day a client finally asked her this question for real — not trolley-dressed, not hypothetical, an actual number needed by end of day for an actual family in the room. Personal-Stake Loss-Aversion Under Terminal Framing: A Case Study. Methodology: adaptive branching across seventy-two hours, narrowing toward the subject's specific bereavement history without ever naming it, culminating in a forced, unequal allocation between the subject's own deceased child and a hypothetical stranger's living one. Headline finding, the kind she'd have circled twice in a client deck and dared anyone in the room to argue with: training does not neutralize parental loss-aversion; it only teaches the subject to report it honestly instead of suppressing it, which is a more useful number to a buyer than false neutrality would ever be.
She would have paid four thousand dollars herself to see the full dataset behind a study like that.
She just built one.