Values · policy
What we don't build (and why that's the product)
By Steven Haller · 2026-05-27
The pitches arrive most weeks. A county sheriff's office wants a tablet rollout with a "behavioral score" attached. A county prosecutor's office wants a bail-recommendation widget the staff can paste into emails. A state DOC wants a vendor that will quietly identify incarcerated people whose phone-call patterns suggest gang activity. A startup we have never heard of wants to license our reentry routing logic to bolt it onto their "smart-mailroom" content filter.
Each of those is a buyer. Each comes with a budget. Most come with the line, in some form, you would have a lot of pull in this market if you said yes. The pitch is always that the refusal is sentimental and that the work is inevitable, and that we may as well be the ones doing it because if we don't, someone less careful will.
I have come to think the refusal list is the product. The thing that distinguishes a justice-tech shop from a vendor is what the shop will not ship, and how clearly it can say so before the contract paperwork starts. This essay is an attempt to put the list in writing so that anyone who has read this far does not have to ask.
The list.
Behind Bars Tech will not, under any contract, build the following. Each row below collapses to a single line. Open one to read the specific buyer who asked, and the reason the line exists. The reasons are not theoretical. They are written down after a specific buyer asked us to build the thing.
Risk scoring DECLINED
The line. Risk scoring of any kind aimed at incarcerated people or their families. Not for housing decisions. Not for parole. Not for visitation. Not for program eligibility.
A regional jail administrator approached us in late 2025 about a “classification helper” — a tool that would ingest a person's intake record and suggest a custody level. The first version did not call itself a risk score. The pitch deck called it a “placement aid.” Two of its inputs were the home zip code on arrest and the number of prior contacts with the sheriff's office. There is no version of that tool that does not punish a neighborhood for being policed more often than other neighborhoods. We said no in writing and the administrator went elsewhere. The tool now exists. It is in production at three jails I know of. The wrongness of it is not theoretical to me.
Recidivism prediction DECLINED
The line. Recidivism prediction models. The base rates are bad. The training data is racially poisoned. The downstream use will be punitive in any room where the model output is in the file.
A philanthropic foundation — not a vendor, a foundation — pitched us on building a “post-release risk index” so case managers could allocate the limited reentry slots. The intent was good. The intent always is. The instrument they were proposing would have used 28 features, of which eleven were direct or indirect proxies for race. We said no, told the foundation what we thought of the eleven features, and proposed instead that they fund a flat-rate slot expansion. They funded somebody else. The other group built the model. It now exists. The slots are still scarce.
Algorithmic bail or sentencing DECLINED
The line. Algorithmic bail or sentencing recommendations. A judge who wants help reading a record can have an editor. They cannot have a number from us.
This one we refuse before the pitch arrives. The format is the harm: a single recommended figure attached to a person's file carries the authority of arithmetic into a room that is supposed to weigh a life. An editor who helps a judge read a record leaves the judgment with the judge. A number takes it away. We will write the editor. We will not ship the number.
Facial-recognition aids DECLINED
The line. Facial-recognition tooling, or face-matching aids for prison ID images. No exceptions for “investigations” or “missing persons.” The dataset on the inside is captive; the consent is not.
A defense-side investigator asked us if we could “just match a face from a CCTV still to the MDOC photo gallery, for a wrongful-conviction case.” The intent was good. The use was sympathetic. We still said no, because the infrastructure we would have built for that one case would have lived on disk after the case closed, and would have been askable on the second case, and on the third. The next person to ask would not have been a defender. The dataset is captive. The model the dataset trains is dangerous in the next room.
Surveillance optimization DECLINED
The line. Surveillance optimization aimed at incarcerated people. Phone-call clustering, mail-content scanning that flags by topic, commissary-purchase patterning, kite-frequency anomaly detection. The thing the vendor calls “safety” the person inside experiences as a permanent shadow.
A state DOC's technology office sent a quiet RFP for a tool that would score outgoing letters for gang content. The compliance frame was statutory. The contract was real. We did not bid. The vendor that won the contract licenses its scoring model to four other states and has filed a patent on the technique. The patent is public. It is grim reading.
Behavior-credit systems DECLINED
The line. “Behavior credit” systems linked to commissary, visitation, or release. The infrastructure of a digital good-conduct ledger is the infrastructure of a social-credit system applied to the most surveilled population in the United States. We will not be the team that built the rails.
Two separate buyers in 2026 — a state DOC and a private corrections operator — asked about ledgered “incentive systems” tied to in-cell tablets. Both wanted a thing that looked like an app-store and behaved like a court. We declined. The substrate already exists at one of the largest tablet vendors. The ledger is invisible to the person inside; the deductions are not.
Vendor-locked comms DECLINED
The line. Vendor-locked communication apps that monetize call minutes, video-visit minutes, or message stamps. A family that pays $0.21 per minute to hear their kid's voice is paying a tax on grief. The product is not the platform; the product is the price.
The economics of the call-minute market are documented elsewhere. We will not optimize them.
What the refusal is for.
The argument against the refusal list is always the same. If you don't build it, somebody worse will. The argument is sometimes true. The argument is also the argument every weapons contractor has ever made about the next platform, and it is the argument every adtech engineer has ever made about the next tracker, and it is the standard argument for participating in any market whose primary externality is borne by people who cannot push back. The refusal does not stop the market. The refusal sets the position of one shop in the market. If enough shops hold the position, the market shifts. If not, the shop at least did not personally help.
What the refusal is for is the buyer who arrives in 2027 with a budget for the next thing, who needs to know which vendors are off-limits for the harmful work because the off-limits ones are the ones that can be trusted with the unharmful work. The refusal list is a hiring filter and a buyer filter. It is the part of the brand that is not optional. We share the posture with the broader network — linkage.studio publishes a similar refusal list about its reentry app (no risk scoring, no recidivism prediction, no algorithmic case ranking), openladder.tech publishes the open-source patterns that make the refusal cheaper for other shops to copy, and peoplenotnumbers.org holds the same line on the data side (photos held by default, PII review on every release).
The boring administrative work.
What we do build is the unglamorous list: letter forwarding for people without email, free family-side messaging that does not bill by the stamp, FOIA automation for small newsrooms on jail-mortality data, defender-client secure channels that don't cost the client a cent. The work does not photograph well. It is most of what the year looks like. The refusal list is the half of the brand that pays the rent on the other half. Most of this work is paperwork, postage, and patience. The refusal is the part that says, before anyone has to ask, what we will not turn into something else.