A Ministry of Justice internal review has confirmed that its Core Person Record system automatically links highly sensitive justice records without a measured false-positive rate, a measured false-negative rate, or an independently verified ground-truth dataset.
The system operates across a population of more than one million records relating to defendants, prisoners, and people on probation.
The Ministry’s own Data Protection Impact Assessment acknowledges that false links and missed links cannot be eliminated. Nevertheless, the approved assessment records zero identified risks and a residual-risk score of 0.0.
Those findings do not establish that every link is wrong, or even that the overall error rate is necessarily high. They establish something more fundamental: the Ministry of Justice does not know the population-level error rate of an identity-linkage system being used operationally across the justice system.
In June 2026, I submitted a Freedom of Information request concerning the Ministry of Justice’s use of Splink, an open-source probabilistic record-linkage library.
The request focused on the Core Person Record and Probation in Court services. These systems use Splink to identify and link records across court, prison, and probation systems, including HMCTS Common Platform, Libra, NOMIS, and nDelius.
I requested five categories of information:
The original response largely directed me to public source code and documentation.
That did not answer the central question.
Model configuration is not model performance. Source code can show how a system is constructed, but it cannot tell the public how often that system links two different people together or fails to link records belonging to the same person.
I requested an internal review.
On 10 September 2026, the Ministry partially upheld my complaint.
The internal review states:
It also confirms that the Ministry does not hold recorded information establishing the observed false-positive or false-negative rates.
The reason given is that calculating a population error rate would require an independently verified ground-truth dataset showing which records across the source systems genuinely belong to the same person.
The Ministry confirms that no such reference dataset exists for the prison and probation population.
This is the central finding.
The Ministry is automatically linking identity records across operational justice systems, but it has no verified dataset against which it can calculate how often those links are right or wrong.
The internal review says that assurance is instead provided through conservative matching thresholds, human review during model development, clerical review of sampled record pairs, and the ability to correct an erroneous link after it is discovered.
Those measures may reduce risk. They do not measure the system’s population-level accuracy.
The Ministry itself acknowledges that reviewing sampled pairs cannot identify all missed matches or establish a population error rate.
The internal review also clarifies what happens when the model assigns a sufficiently high score to a candidate match:
There is no separate clerical-review threshold for newly created links.
Human involvement occurs later, when linked information is presented to a member of staff through an operational service, or when somebody notices a possible error and asks for the link to be investigated.
That is not the same as verifying the identity match before it takes effect.
The system also allows incorrect links to be broken and manual overrides to be recorded so that the same match is not created again. This is a reactive safeguard. It depends upon an error first becoming visible to a member of staff or an affected individual.
An existing link may enter a review queue if updated information causes its score to fall below a separate “fracture threshold” of 18. That safeguard still does not provide individual human verification when the initial link is created.
The Ministry’s original response stated that “Splink does not directly make decisions about individuals.”
The internal reviewer conceded that this statement was “too absolute.”
The revised position is that Splink calculates the probability that two records relate to the same person, and the records are linked when the applicable threshold is met. Human beings then make operational decisions using information that may include the linked record.
That distinction matters, but it does not remove the significance of the automated linkage. Before a person makes a decision, the system has already decided which records to assemble and present as belonging to the same individual.
It is also worth noting where the DPIA addresses this directly, and where it does not. Question 2.8 asks how the service supports data-subject rights, and lists among them “rights in relation to automated decision making and profiling.” The recorded response does not mention automatic linking at all. It says that the Subject Access Request team will be a user of the service.
The Ministry also disclosed the methodology used to estimate the model’s matching parameters.
The supplied code contains the following value:
This is not an observed recall rate for the operational system. It does not demonstrate that the system correctly identifies 80 percent of true matches.
The figure is an assumed input used when estimating the probability that two randomly selected records belong to the same person. More precisely, it represents an assumption about the proportion of true matches captured by a set of deterministic matching rules used during that estimation process.
The disclosed methodology provides no empirical justification for selecting 0.8. It includes no ground-truth validation demonstrating that the deterministic rules actually achieve that recall, and no sensitivity analysis showing how a different assumption would affect the resulting model.
The remaining parameters were estimated using random sampling and expectation-maximisation. Those are recognised statistical techniques, but estimation is not the same thing as external validation.
The Ministry therefore used an assumed recall value during training while confirming that it has never measured the operational system’s false-positive or false-negative rates against independently verified ground truth.
The disclosed Data Protection Impact Assessment makes the governance problem more difficult to explain away.
The DPIA states that the HMPPS Person Record processes more than one million records — the volume band recorded is “1,000,001+”. It covers defendants, people in prison, and people on probation, including some children passing between youth justice and probation services. The processing is carried out under Part 3 of the Data Protection Act 2018, the law-enforcement regime.
The categories of information listed in the assessment are not limited to names and dates of birth. They include religion or religious beliefs, sexual orientation, disability or specific condition, citizenship status and nationality, passport numbers, National Insurance numbers, national identity card details, next of kin, home addresses, emergency contact details, and Police National Computer and Criminal Records Office identifiers.
The DPIA explicitly recognises the problem:
of false positives and false negatives. It says that the system is intended to minimise false positives, while accepting that this will increase the number of false negatives. It also acknowledges that neither type of error can be eliminated because the system depends on source data containing inaccuracies, manual-entry mistakes, incomplete updates, and deliberately false information.
Elsewhere, the DPIA states:
Yet the summary of the approved assessment records:
This does not necessarily mean that no controls exist. The DPIA describes security measures, correction procedures, audit logging, access controls, and operational safeguards.
But an assessment that acknowledges unavoidable identity-linkage errors, admits that their frequency is unknown, states that there is no definitive source of truth, and then records no risks at all requires serious scrutiny.
Three further entries in the same document are worth recording. On retention, the assessment says there has been no update from the responsible boards, “therefore we are instructed to continue retaining the data until told otherwise.” On assurance, an internal comment states that the team was “not looking for an approval on this DPIA as we do not have a completed IT Health Check and won’t for a few months.” And an AI Ethics assessment of the service is described as in progress rather than complete. The service was recorded as out of scope for the GovAssure cyber-assurance regime, on the basis that it is not on the list of services subject to it.
The assessment’s own metadata records the dates.
It was created on 11 March 2024. The comment about not seeking approval without a completed IT Health Check sits in the 2024 material. It was then submitted on 22 June 2026 and approved on 25 June 2026, three days later.
My Freedom of Information request was made on 16 June 2026, and the statutory deadline for a response fell on the same date band. The assessment covering a service that has been operational for years was signed off in the days that followed.
That sequence does not establish causation, and I do not claim it does. The dates may be coincidental, and there are ordinary reasons a long-running assessment reaches approval at a particular moment. But the sequence is on the face of the document, and the Ministry is in a position to explain it.
A false positive occurs when records belonging to two different people are treated as belonging to the same person.
A false negative occurs when records belonging to the same person are not linked.
In an ordinary commercial database, these errors may cause inconvenience. In a justice system, the consequences can be much more serious.
An incorrect link could cause information about another person’s history, risk, sentence, or supervision to appear within an operational workflow. A missed link could prevent staff from seeing relevant information that genuinely belongs to the person before them.
The internal review says that linked information is used as supporting information rather than as the sole basis for decisions. That is important, but supporting information can still influence a decision.
The question is not whether Splink independently sentences someone or makes a final probation decision. It does not.
The question is whether the Ministry has adequately validated the automated identity layer that determines which records are presented to the people making those decisions.
The Ministry’s answer is now clear: it has not established a population-level error rate because it has no independently verified ground truth against which to calculate one.
These disclosures do not prove that Splink itself is defective. The evidence points the other way: it is competent, well-documented software, and the DPIA records that it was built in house, with no processor and no third party involved in this service.
“Not held” means unmeasured. It does not mean inaccurate. The Ministry’s defence is that assurance comes from conservative thresholds, clerical review of sampled pairs, and the ability to override a bad link. That defence deserves to be tested on the record rather than dismissed.
These documents do not establish how many people have been affected by incorrect links. They do not establish criminality or determine whether the Ministry has breached any particular law. Those are questions for Parliament, regulators, courts, and other competent authorities.
What the documents establish is that:
That is not a theoretical criticism of algorithmic systems. It is a documented governance gap inside an operational identity infrastructure.
The Ministry of Justice should explain:
For months, the public record showed how the model worked but not whether it worked accurately enough for its operational context.
The internal review has now explained why.
The Ministry of Justice does not hold the evidence required to answer that question.
These are the disclosures themselves, as received. Nothing in the substance has been altered; the only changes are that my name and email address have been removed from the letterhead of the internal review, and the document metadata has been cleared of the same. The Ministry’s own published contact details are left as they were, and the DPIA was redacted by the Ministry before release, not by me.
The earlier response in this request, and the Australian and German responses that preceded it, are published with Nobody Is Counting.