In September 2023 an Albany court denied a records request for the formulas and decile cutoffs behind COMPAS, the risk assessment tool New York’s prison agency uses for parole decisions. The decisive argument was not safety or privacy: it was that with the code public, another vendor could rebuild the system far more cheaply. Meanwhile, in actual courtrooms, New York law already requires pretrial tools to be disclosed.
On 14 September 2023, the New York State Supreme Court in Albany County ruled that residents of the state have no right to see how the software that helps decide who leaves prison actually works. The case is Matter of Rayner v. DOCCS & Equivant, Index No. 908549-22. The petitioner, Martha Rayner, is a clinical law professor at Fordham Law School, represented by Yale Law School’s Media Freedom & Information Access Clinic. Under New York’s Freedom of Information Law (FOIL) — the state’s public-records statute — she had asked for the formulas, the coefficients, the decile cutoffs and the “Norm Core data group” behind COMPAS. Not source code as a trophy: the numbers that turn a questionnaire into a score.
COMPAS is a statistical risk and needs assessment instrument, built in 1998 by a company then called Northpointe and renamed equivant in January 2017. It belongs to the family of systems the industry markets as artificial intelligence applied to criminal justice: a model trained on historical data about incarcerated populations, which takes structured questionnaire answers as input — criminal history, housing, employment, relationships, stated attitudes — and outputs a risk score expressed in deciles, that is, a position relative to a reference group. It does not predict anything about the individual: it estimates how closely that combination of answers resembles profiles that, in the training sample, had certain outcomes. The New York version is called COMPAS Re-entry. The Department of Corrections and Community Supervision (DOCCS) has used it since 1998, after a competitive procurement awarded to equivant, and it feeds into the state’s Parole Board.
Where the software actually sits
One misconception needs clearing first, because it travels widely outside the United States. COMPAS, in New York, is not in the courtroom. State law — Executive Law §259-c(4) and regulation 9 NYCRR 8002.2(a) — requires the Parole Board to use a periodically validated risk and needs instrument, and the Board adopted COMPAS Re-entry. This is an administrative stage after conviction, not a stage of trial. A judge imposing a sentence does not receive a COMPAS decile; the people deciding whether a prisoner returns to supervised release do.
The distinction matters, because at the earlier stage — deciding whether someone awaits trial in jail or at home — New York went the opposite way, and did so years ago. The 2019 bail reform introduced CPL §510.45, effective 10 January 2020. It requires that any questionnaire or instrument used to decide release be made available to the defendant and counsel on written request, and that a blank copy of the form be handed to anyone who asks. Subsection 3(b)(ii) requires the instrument to be empirically validated and revalidated, with the studies and the underlying data publicly available on request. Subsection 2 bars the state, counties and cities from contracting pretrial monitoring out to for-profit private entities. In New York City, the instrument in use is the CJA Release Assessment, run by the Criminal Justice Agency and updated in 2020 with the University of Chicago Crime Lab, the Mayor’s Office of Criminal Justice and Luminosity; its methodology and reports are published.
The paradox is sharp. At the pretrial stage, where the stakes are the liberty of someone presumed innocent, the state legislated transparency and shut out the for-profit vendor. At the parole stage, where the stakes are the early release of a convicted person, a proprietary product bought twenty-seven years ago remains closed.
The reasoning nobody expected
Rayner’s FOIL request was filed on 4 January 2022, denied on 13 June 2022, with the administrative appeal rejected on 21 July. In the ensuing court proceeding, equivant intervened — by order of 27 April 2023 — to defend its trade secrets. The judge found that the “COMPAS-NY” technology is a bona fide trade secret, therefore exempt from disclosure under Public Officers Law §87(2)(d), and added that any change to that rule “must come from the state legislature.”
So far, routine. The interesting part is the economic argument used to justify the secrecy. The judge observed that if the formulas were made public, another of the state’s IT vendors could replicate the system’s functionality at far lower cost. He cited State Finance Law §163(1)(h), which requires agencies to document the reasonableness of contract costs, noting that in that scenario it would be doubtful whether DOCCS could keep renewing non-competitive awards to equivant.
The competitive harm that justifies the secret is, read backwards, evidence that the state is overpaying. If publishing the formulas would let a rival do the same job for much less, then the gap between those two prices is precisely what the secret protects.
The logic is the same one that underpins pharmaceutical patents and much commercial software, and in the abstract it is defensible: whoever funds the research deserves a return. But here it applies to a product in service with the same public buyer since 1998, where the “research” being rewarded is no longer a risky investment but an entrenched position. And it applies to a model which, according to the petition (paragraph 29), has no published predictive validation studies for the Re-entry version used in New York, while such studies exist for other instruments in the same family.
The real ledger
What New York pays for COMPAS is not known: the ruling mentions the contract without stating its value, and no public figure surfaced. That itself is a finding. But the relevant accounting is not the licence price. It is this:
- The cost of non-contestability. A statistical model of this kind is not a neural network with billions of parameters: it is a set of coefficients applied to questionnaire answers. The barrier to entry for a competitor is not technical, it is informational — and trade secrecy manufactures it.
- The cost of non-validation. Without published studies on the population the model is applied to, there is no way to know whether the score performs better or worse than a simpler rule. Nobody can tell whether the state is buying information or noise.
- The cost of the reversed burden. New York’s pretrial statute puts validation on whoever supplies the instrument. On parole, whoever objects must establish a right to see — and in a FOIL proceeding, loses to a trade-secret exemption.
One precedent, not a wave
It is often said that American states are dismantling trade secrecy over risk models. The documented evidence is much thinner. Idaho’s House Bill 118 of 2019, sponsored by Representative Greg Chaney, requires transparency for pretrial risk assessment tools and bars both builders and users from invoking trade secrecy to withhold them. One academic source dates it to 28 March 2019 and identifies it as the first state law to strip that protection entirely in this field. During its passage through the Idaho House Judiciary committee, however, the clause requiring a demonstration that the tool was free of bias was removed.
One state, a law confined to the pretrial stage, and the most uncomfortable provision dropped along the way. That is not a wave: it is an isolated precedent which, moreover, addresses the same segment New York had already covered on its own. The uncovered segment — decisions about people already convicted — remains where trade secrecy holds up best, probably because that is where legislators find the least political upside in intervening.
One final detail is worth more than most arguments about algorithmic fairness. The Albany judge did not write that COMPAS is accurate, nor that it is fair: he was not asked to decide that and he did not. He decided only that the formulas fall within the FOIL exemption for trade secrets. Which means, literally, that New York State feeds part of its judgment about a person’s liberty through a model whose quality no court has ever examined — and that such an examination cannot begin, as things stand, while the numbers remain a corporate asset.