Industries / Government, education and community / Government / Corrections
Intake and records
Intake is the work of receiving a person into custody: confirming who they are, entering the court's orders and working out when they must be released; records keep those facts, and the counts built from them, correct for the whole stay.
6
Milestones achieved, from 1903 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1903
Fingerprints replace body measurements
Prisons once identified people by Bertillon measurements, a set of body measurements such as head length and arm span recorded on a card with a photograph. In 1903 Captain James H. Parke began fingerprinting everyone convicted of crime in New York State, which the US National Institute of Justice's Fingerprint Sourcebook calls the first systematic use of fingerprints for criminal records in the United States. In the often-told story from the same year, the federal prison at Leavenworth, Kansas, received a man named Will West whose measurements nearly matched those of William West, already held there, and only their fingerprints told them apart. From 19 October 1904 Leavenworth fingerprinted everyone it held, and in 1924 its files and those of the National Police Bureau of Criminal Identification, 810,188 records in all, became the Identification Division of the Bureau of Investigation, later the FBI.
LessonTwo people can share the same body measurements, so an identity record needs a mark that does not repeat.
Credit
James H. Parke in New York; John Kenneth Ferrier of Scotland Yard and Major M. W. McClaughry at Leavenworth
1926
A national prisoner count
The US National Prisoner Statistics program has counted the people held in state and federal prisons every year since 1926. The Bureau of Justice Statistics (BJS) now sends its forms to the 50 state corrections departments, the Federal Bureau of Prisons and five territories, about 56 reporting bodies, and asks for totals by sex and race, the number held in private prisons and local jails, prison capacity, noncitizens and people aged 17 or younger. These are totals, not records of individuals.
Credit
US Bureau of Justice Statistics
2000
A world prison count
The World Prison Brief, an online database, was launched in 2000 from figures compiled by Roy Walmsley and now holds prison statistics for 224 jurisdictions; it is run by the Institute for Crime & Justice Policy Research at Birkbeck, University of London. In Europe, the Council of Europe's annual penal statistics, known as SPACE and compiled at the University of Lausanne, have collected prison figures from member states since 1983 and probation figures since 1992.
Credit
Roy Walmsley and Birkbeck's Institute for Crime & Justice Policy Research; Marcelo F. Aebi and the SPACE team at the University of Lausanne
2009
A single database, cut back
In 2004 the National Offender Management Service of England and Wales began C-NOMIS, an IT system meant to give prisons and probation one shared record for each person, due by January 2008 at a lifetime cost of £234 million. By July 2007 it had spent £155 million, was two years late and its estimated cost had risen to £690 million, and a minister halted it; a cut-down program followed at an estimated £513 million. The National Audit Office reported in March 2009 that the single shared database would not be delivered, though the number of databases in use would fall from 220 to three, and it blamed weak oversight, underestimated technical complexity and no monitoring of the budget.
LessonA national records system needs a realistic budget and basic project controls from the start; the auditors found both missing.
Credit
reported by the UK National Audit Office
2015
Release dates miscalculated for years
In December 2015 Washington State's governor announced that an error in the Department of Corrections' records system had led to as many as 3,200 people being released early since 2002. A 2002 state Supreme Court ruling required good-behavior credit earned in county jail to count toward prison sentences, and the code written for it gave too much credit to people whose sentences included enhancements, extra time that by law cannot be reduced for good behavior. Most errors were 100 days or less. The department was first told in December 2012, by a victim's family who did the arithmetic themselves, but the fix was put off again and again; until it was made, staff checked affected release dates by hand. Outside investigators appointed by the governor, from the Seattle law firm Yarmuth Wilsdon, found "systemic failures of process and management".
LessonA change in sentencing law is also a software change, and it needs tests and someone who owns it.
Credit
the family who checked the figures; the governor's outside investigators
2022
Deaths in custody undercounted
The US Death in Custody Reporting Act of 2013 requires states that receive certain federal justice grants to report deaths in custody to the Justice Department. In September 2022 the Government Accountability Office (GAO) compared states' records for fiscal year 2021 with public sources and found nearly 1,000 deaths that potentially should have been reported but were not; 70% of the records states did send lacked at least one item the law requires. In December 2023 the department issued compliance guidelines that tie each state's reporting plan to its annual grant application.
LessonA legal duty to report did not produce a complete count until someone checked the reports against other sources.
Credit
US Government Accountability Office
Watch for
Digital release checks in England
People released from prison by mistake in England and Wales rose from 115 in the year to March 2024 to 262 in the year to March 2025, according to the prison service's official annual digest. After an independent review by Dame Lynne Owens made 33 recommendations, the Ministry of Justice said in April 2026 that its Calculate Release Dates service already covers 98.5% of sentence types, that a Digital Rapid Response Unit is building AI tools that pull data from court papers and link the records of people known by more than one name, and that fingerprint and face checks would be trialled within six months. It reports 179 releases in error in the year to March 2026, a 32% fall.
Achieved whenthe official annual digest shows releases in error below the level of the year to March 2024 for a full year, and an outside body such as the National Audit Office or HM Inspectorate of Prisons credits the new tools.
Claimed by
UK Ministry of Justice, April 2026
India automates early release lists
On 27 May 2026 India's National Legal Services Authority launched an Early Release Processing Module inside e-Prisons, the prison records platform built by the National Informatics Centre (NIC). NIC says the module identifies people eligible for remission or premature release before their eligibility date, handles the proposals online and sends status alerts by SMS and WhatsApp. It was piloted at Central Jail, Agra and District Jail, Lucknow, after a Supreme Court ruling on delays.
Achieved whenstates adopt it and an outside body, such as a court-appointed monitor or an audit, publishes how long cases took before and after.
Claimed by
National Informatics Centre, July 2026
Open
01
A shared release-date test
A public set of at least 1,000 made-up sentence cases, each written in plain words and as structured data, with the correct release and parole eligibility dates worked out by hand from the law and checked by a second person. Start with three jurisdictions whose sentencing laws are public, and include the hard cases: consecutive and concurrent sentences, jail credit, good-behavior credit and enhancements. Use it to score calculators and AI models. No case describes a real person.
The briefCredit · unclaimed02
Made-up intake papers
A benchmark of at least 2,000 synthetic court papers of the kinds a prison's reception desk receives, such as sentencing orders, commitment warrants and remand orders, in many layouts and some scanned badly. Each comes with the correct fields: charges, sentence lengths, whether sentences run together or one after another, and days of credit already served. Score document-reading tools on getting every field right, and report separately the fields that change a release date.
The briefCredit · unclaimed
Open 01 · the brief
A shared release-date test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 1,000 made-up sentence cases, each written in plain words and as structured data, with the correct release and parole eligibility dates worked out by hand from the law and checked by a second person. Start with three jurisdictions whose sentencing laws are public, and include the hard cases: consecutive and concurrent sentences, jail credit, good-behavior credit and enhancements. Use it to score calculators and AI models. No case describes a real person.
Why it is still open
Release dates are worked out by records staff or by agencies' own software; New York's corrections department says the records coordinator at each prison does the calculation. England's Ministry of Justice publishes the code of its calculator under the MIT license, with about 1,100 test cases, but they cover England and Wales only and are written in that program's own input format. Recidiviz says it tests how it encodes each state's rules for its time-computation module, but it publishes no test set. Searches of the web and of Hugging Face found no public test set that spans jurisdictions.
Built with what exists
State statutes and published sentencing guides; the structure of England's open-source calculator as a model; Python for a reference calculator; language models such as Claude, Gemini and GPT, each run on every case 20 times to see whether it gives the same answer each time (see the Note "All twenty times").
First step
Write 50 cases for one state's good-behavior credit rule, including sentences with enhancements like those behind Washington's error, and have a second person check every date.
Done looks like
1,000 cases across three jurisdictions, every date agreed by two independent checkers, and published scores giving the share of cases each tool gets exactly right and the share it gets right on all 20 tries.
Open 02 · the brief
Made-up intake papers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A benchmark of at least 2,000 synthetic court papers of the kinds a prison's reception desk receives, such as sentencing orders, commitment warrants and remand orders, in many layouts and some scanned badly. Each comes with the correct fields: charges, sentence lengths, whether sentences run together or one after another, and days of credit already served. Score document-reading tools on getting every field right, and report separately the fields that change a release date.
Why it is still open
Reception staff still type these papers into prison systems by hand. England's Ministry of Justice is now building AI tools to pull data from them and a "smart inbox" to sort court emails sent to the wrong prison, and it has published no accuracy figures. The nearest public benchmark, Extend AI's LongArray-Extract, holds only eight synthetic federal indictments, the charging papers that start a case, and no sentencing or commitment papers. A search of Hugging Face found no dataset of court orders of this kind.
Built with what exists
Blank judgment and commitment forms that courts publish; a language model to write varied made-up cases; an HTML-to-PDF generator and image noise to imitate scans; small low-cost models for the extraction, such as Muse Ember (see the Note "The jobs agents repeat"), compared with commercial OCR (optical character recognition) services.
First step
Collect 20 blank sentencing and commitment forms from court websites in two US states and in England, and fill 100 of them by script with made-up details.
Done looks like
2,000 documents in at least 20 layouts from at least two countries, with field-level answers, and published scores for at least five tools, including the error rate on sentence length, consecutive or concurrent terms and credit days.