On November 19, the UK government published an Algorithmic Transparency Recording Standard record for the Risk of Reoffending Tools used by His Majesty’s Prison and Probation Service in England and Wales. It describes five tools, the data behind them, validation results and the points at which staff use their scores. The record identifies the tools as already in production. Some of the models described were released in February 2026 and the two sexual reoffending models date to April 2024. Publication makes an existing system easier to examine. It provides a basis for asking how a score influenced a decision to allocate supervision or recommend an intervention. HMPPS transparency record
The disclosure matters because the score can travel. It may begin as a calculation in a probation assessment, inform which services someone is offered, and appear in a report to a court or a parole decision-maker. The agency says practitioners use the scores as one input and retain responsibility for the broader assessment. The public record allows a reader to distinguish what the tool estimates from what officials must decide. It still leaves a practical question. When a decision is contested, can someone reconstruct the contribution the score made alongside the rest of the evidence?
What the scores measure#
The record names predictors for all proven reoffending, violent reoffending, serious violent reoffending, combined serious reoffending and specified sexual reoffending. The combined score brings together estimates for serious violent and specified sexual offences. The sexual tool covers two categories with different definitions. These distinctions matter because a score for any recordable offence addresses a different outcome from one for a serious offence. A report that says someone is in a high band should identify the corresponding predictor and the outcome it covers. Without those details, readers can easily attach a more serious meaning to the label than the underlying measure supports. HMPPS summary and definitions
The target is proven reoffending within two years from the calculation date, such as the start of a community order or release on licence. In the record, proof generally means a conviction or a qualifying police caution. It is different from all offending that actually occurs. Some offences go undetected, some cases do not lead to prosecution and others take longer to conclude than the available follow-up. HMPPS says the model development allows at least another year for offences to reach a proven outcome. That still leaves some offences outside the resulting count. The two-year horizon describes what the models are designed to estimate at a defined assessment point.
There is a reason to use a documented outcome. Training on arrests or allegations can make the model learn from claims that did not lead to a proven result. Using proven cases makes the label easier to define. It is still produced by a justice system. Reporting, detection, investigation and the time taken to resolve a case help determine which conduct appears in the records. Two groups with the same underlying conduct may be observed differently. That is a measurement problem to investigate using evidence about the dataset. It cannot be answered by treating the label as either perfectly neutral or entirely meaningless.
The other distinction is between the chance of a specified outcome and an assessment of serious harm. A forecast covers a category and time window. A practitioner also has to consider how harm might arise, whom it may affect and what practical action is available. Repeated conduct may cause serious cumulative harm even when no single event fits a model’s most severe outcome category. A recently recorded threat can be relevant to an immediate safety decision without changing every statistical input. HMPPS describes the combined serious score as a starting point for a wider assessment. The two questions call for different evidence. HMPPS deployment context
A score reaches a decision through people and reports#
HMPPS says the scores are calculated as part of the digital risk assessment process for people receiving pre-sentence reports after conviction, at the start of community or custodial sentences and at later reviews. They can inform the grade of practitioner assigned to a case, the level of community supervision and eligibility for certain interventions. Scores or bands can also appear in written pre-sentence, parole and recall material. That is a substantial administrative role. A number can affect how scarce time and services are allocated even when the final decision rests with a judicial office holder or a probation officer.
A new supervision case shows how the influence can accumulate. An officer might read a risk band before seeing the person’s housing situation and history of engagement with services. The band can help flag a case for attention, but it can also become the first impression against which the rest is interpreted. A later officer may only see the written summary. An auditable report should identify the predictor, when it ran and the information a practitioner independently considered. That lets a reviewer see where different evidence reinforced the numerical score or led the professional to a different judgement.
There can be a benefit for consistency. HMPPS argues that a standard measure provides a common starting point for staff who would otherwise make assessments with different amounts of experience and local information. Consistency has value when two people in comparable situations receive different treatment simply because of where their cases are handled. It also has a cost when the same data limitation appears across many cases. A shared instrument can spread a bad assumption more widely than a single poor decision. The outcome to examine is improved consistency between comparable cases while preserving room to respond to differences that matter.
The independent HM Inspectorate of Probation describes actuarial estimates as one part of assessment and points to their limits for questions such as the nature, severity and imminence of harm. Its national inspection in 2025 found examples of analytical court reports and also weaknesses in adapting practice to people’s circumstances, including disability, ethnicity, employment and language barriers. Those are broader findings about probation. They do not establish that this set of predictors caused a particular failure. They help explain why staff capacity and the quality of written reasoning should be examined along with the model. Inspectorate assessment evidence · 2025 national inspection
What happens when information is incomplete#
The record explains that several predictors have a static and a dynamic version. Some inputs come from probation and prison systems. Practitioners provide others through the assessment, including information about circumstances and needs where relevant. A dynamic version can incorporate changing factors, such as accommodation and employment needs, if the required fields are present. Where the relevant dynamic data are unavailable, the system uses the static score if it can calculate one. If a required static input is missing, the system returns no score. A reader should be able to tell which of these situations produced the number in front of them. HMPPS operational data specification
That difference can affect how a score changes over time. If a later assessment supplies more information, a change may reflect a different available model as well as changed circumstances. If information is stale, a score can continue to reflect circumstances that no longer apply until someone updates the underlying record. A person’s arrangement for accommodation, for example, may have changed since an earlier report was written. A reviewer should be able to see the assessment date, which version generated the score and what information was current. Otherwise a change in the number can look like evidence of progress or deterioration when part of the change was administrative.
The record also describes a population limit for the individual sexual reoffending estimates. They are calculated for men with a qualifying current or past sexual-offence history. HMPPS says there were too few proven sexual reoffending outcomes among women in its development data to validate a reliable model specific to women. For women with relevant offending history it uses a base rate within the combined serious calculation. A missing individual estimate and a baseline incorporated into a separate aggregate mean different things. That distinction matters when someone compares reports across cases with different information or tries to understand what a combined score represents. HMPPS scope and model limitations
A public description of these rules can help people ask for facts to be checked without exposing another person’s record. The published document describes a role-restricted system and says inputs and outputs follow the retention rules of the underlying record. A score can be carried into several places while access to underlying material stays restricted. That makes provenance and correction important. When a source record is corrected, the agency should be able to identify which assessment used the old information and which reports may need attention. This is a practical consequence of using the same output across multiple decisions.
Consider two points in that process. At the first, an assessor learns that a piece of information already used in a calculation was wrong. The disclosure says practitioners update relevant inputs and may recalculate the score. At the second, someone has already read a report containing the earlier band and used it to inform a recommendation. A corrected entry in the assessment system tells that second person very little unless the correction reaches them. Part of the accountability test is whether the organisation can find the affected report, explain what changed and indicate whether the earlier recommendation is being reconsidered. The public record describes recalculation. How corrections travel between systems and reports is an operational question worth auditing. HMPPS human review and recalculation
Validation helps interpret the numbers#
The annex reports a concordance index and, for several models, a mean calibration measure. Concordance tests how often the score ranks someone who had the specified outcome higher than a comparable person who did not. In the full evaluation data the static all-reoffending predictor had reported concordance of 0.803 and the static serious violent predictor had 0.777. Those values describe ranking on the specified historical samples. They do not mean the system correctly predicts an individual’s future in 80.3 or 77.7 percent of cases. A reader should keep the measure, population and outcome attached to each number. HMPPS validation annex
Calibration addresses a different question. If a set of people receive similar probability estimates, does the fraction with the recorded outcome correspond to the estimates? The record reports a mean comparison at population level where available. An overall average can look close even if a model overestimates one group and underestimates another. Decision thresholds add another issue. At any operational boundary, some people put above it will have no proven reoffence and some below it will. The balance matters when a threshold changes access to a service, the intensity of supervision or which evidence is highlighted for a decision-maker.
Consider a hypothetical population of 1,000 cases in which 50 people have the measured outcome. Suppose a rule flags 150 cases, including 30 of the 50. It catches 60 percent of the observed outcomes, while 120 of the 150 people flagged do not have one. That is a simple illustration of the role of the base rate. It is not an estimate of HMPPS performance. In practice, the effects depend on the outcome frequency, the rule for acting on the score and the cost of extra supervision or missed support. Concordance alone cannot supply those operational counts.
The development and evaluation dates also matter. For several predictors the disclosure describes fitting on a community caseload snapshot from June 2018 and evaluating separately on people who began the community part of a sentence in the second half of 2018. Some analyses were limited to people with full assessments. The record also explains exclusions for incomplete inputs and events that prevented observation over the full follow-up, such as a return to custody. Those exclusions help define the test, and they mean the reported statistics are about the population observed in that design. A case being reassessed today can differ from one starting a community sentence in that period. HMPPS development-data specification
The Ministry of Justice had already published revalidation work on earlier actuarial tools in 2024. It examined the ability to distinguish different levels of risk and to estimate outcome rates for groups, and it reported a deterioration in predictive validity for an earlier sexual predictor that prompted further analysis. A separate 2025 study examined possible new methods and data for serious violence. These publications show a history of revision, rather than a November 2026 launch of algorithmic scoring. They also illustrate the need to check how a tool behaves when offence patterns, recorded outcomes and assessment practice change. 2024 revalidation · 2025 serious violence study
The November record says the Ministry considers retraining when revalidation or monitoring shows a need, or when an updated model could improve on the operational version. It gives no fixed schedule. That creates a different accountability question from the one posed by a single table of results. Are recipients of an assessment able to identify the version behind it and compare results on consistent terms across an upgrade? A release that improves ranking can still change the number of people in a band. If supervision assignments depend partly on that band, staff need to understand whether a change came from a new model, a corrected record or new facts about the person. Periodic reporting on version changes and their practical effects would make the evaluation more useful to those reading the reports. HMPPS maintenance specification
Equality and oversight have to be observable#
The agency says it carried out impact assessments on data protection and equality. The record names age and gender among the protected characteristics used as model features and says the equality analysis found performance appropriate for the relevant groups it examined. It identifies overreliance and unequal performance as risks. Those are disclosures of what the responsible organisation assessed and how it describes its safeguards. A public assessment can go further by showing which subgroups were tested, how much data were available, which error measures were used and the uncertainty on those figures. That would help distinguish a well-supported finding from an average that conceals important variation. HMPPS risk and impact assessments
Data quality is part of that assessment. Some fields arrive through administrative systems, while other entries depend on interviews and staff recording. A full assessment may be more common in some settings than others. The effect on a model could come from differences in what is recorded as well as differences in the people assessed. Monitoring outcomes by assessment type and region could reveal whether a static fallback is routine in places where it was expected to be occasional. It could also show which processes need better information rather than implying the people assessed are responsible for the absence of it.
The downstream effect matters even if a model performs similarly across groups. A recommendation for a rehabilitation programme depends on whether that programme is available locally. More intensive supervision may bring useful support, additional requirements or both. A score used to assign a case to staff could have different consequences when experienced practitioners are scarce. Parliament’s Public Accounts Committee has documented substantial pressure on probation services in England and Wales. Its inquiry is separate from the validation of these scores. Staffing and service availability are nevertheless relevant when testing whether practitioners have time to consider context and whether people can receive the intervention suggested. Public Accounts Committee report on probation
A human decision-maker can examine context. The existence of that person in a workflow does not establish how this works across actual decisions. A practical audit can sample reports, identify where a score was cited and examine the reasoning used when other evidence pointed in another direction. It can look at changes after a correction and at cases where the practitioner decided a band was a poor fit. An audit of this kind can also identify good uses of the score, such as focusing attention on a need that would otherwise have gone unaddressed. The point is to observe judgement in action, with the information and constraints staff actually had.
The record says the tools are introduced during mandatory training on risk, with guidance on entering inputs and additional support for court staff. A useful training case would give two practitioners the same band and different evidence about a person's circumstances. One might have reliable information about a recent change that the system has yet to record. The other might have cause to put more weight on the estimate because the surrounding evidence is consistent with it. Looking at what each writes would test whether they can distinguish the observed facts, the model output and their own reasoning. It would also show what a later reviewer can recover from an assessment without having been present for the conversation. HMPPS training requirements
What the disclosure makes possible#
The Algorithmic Transparency Recording Standard is intended to describe tools used by public bodies, including tools that significantly assist decisions with public effects. Its scope encompasses systems that provide inputs to human decisions as well as systems that automate them. The HMPPS record tells readers that ordinary review and appeal processes apply to the relevant decisions and names the organisation responsible for the tool. That information makes it easier to direct questions to the right place. The record is a tool-level description. Its presence alone does not show whether an affected person can readily obtain the specific reasons or data correction they need in an individual case. UK disclosure policy and mandatory scope
There are several useful forms of follow-up. The agency could publish a dated summary of the versions in use, the amount of missing data, and rates of correctable errors seen in assessments. A recent validation could distinguish first assessments from later reviews and report the performance of thresholds used in practice. It could also look at whether people received different supervision or services after a change to a score. These measures would explain more about the tool’s effect while keeping personal case data private. Where a measure is absent, its limits can be stated explicitly so that a historical evaluation is not read as a current audit.
For an individual decision, good documentation would show which predictor ran, which version was used, which recorded facts mattered and how the decision-maker weighed the result. A record of corrections can show whether the relevant result was recalculated and whether anyone relying on an earlier report was informed. The balance is between protecting case data and allowing meaningful scrutiny of a decision. Aggregate reporting and explanations tied to the individual serve different purposes. One can describe how a system works across a population, while the other can make the reasons in a particular case understandable.
The November publication provides a clearer picture of a set of algorithms that already contribute to decisions across England and Wales. It explains the two-year outcome, lists model versions and shows how numerical estimates fit into professional assessment. The question to follow is how much that transparency changes what can be examined in practice. A person should be able to understand the role of a score when it affects a decision. A practitioner should be able to explain the judgement that followed it. An independent reviewer should be able to see how often the system behaved as the public record describes.