Data → inference → attention → consequence · New in v1.13
Compare how public systems forecast places, classify people, resolve identities, allocate screening, maintain watchlists, or structure behavioral threat assessment—and inspect the human review, notice, retention, evaluation, equality, and unknowns around each program.
Answer-first summary
Predictive enforcement is a decision architecture that connects historical or administrative data to identity resolution, inference or prioritization, human interpretation, state attention, retention, and review or redress.
Read the supporting sectionNo. They ask different questions, use different evidence, affect different decisions, and create different error, rights, notice, and evaluation problems.
Read the supporting sectionWhen a forecast or designation changes where officials look or whom they scrutinize, the resulting contacts and detections can become future data that appears to confirm the original priority.
Read the supporting sectionThe unit of analysis
The consequential chain starts before a score and continues after it. Collection, identity resolution, inference, interpretation, intervention, retention, and redress can be owned by different institutions and can fail in different ways.
Place-based
A map can redirect patrol without naming an offender. It can still reproduce reporting, deployment, and data-quality patterns.
Person-based
A score or tier can persist across encounters and affect surveillance, visits, diversion, screening, or partner treatment.
Identity and status
Identity resolution is not itself future-conduct prediction, but a match can trigger consequential downstream rules.
Professional assessment
Structured behavioral threat assessment is not automatically an actuarial forecast, but it can drift toward pre-crime when vague indicators become durable suspicion.
Comparative program explorer
Use the filters to compare function, status, authority, input, output, human review, notice, retention, evaluation, equality findings, public unknowns, and what the record does not establish.
Netherlands
CAS
A national gridded place-risk model. The public register describes the variables and place-based purpose but does not make the complete implementation or performance record public.
Police teams and information specialists interpreted the output; the register does not describe an automated coercive action against a person.
The registered output concerned neighborhoods, not a named-person decision. Public material reviewed here does not establish a resident-specific notice or correction process.
The algorithm register records phase-out and an end date; detailed forecast-retention arrangements remain outside the bounded public record used here.
The register states that CAS was regularly evaluated and was phased out after an evaluation session. It does not establish a distinct crime-reduction effect.
The system was described as depersonalized and place-based, but incident data and deployment choices can still encode reporting and enforcement patterns.
The Dutch government algorithm register records CAS as out of use, with an end date in December 2025.
United States · federal support to multidisciplinary teams
FBI BTAC · FBI TATM Initiative · U.S. Secret Service NTAC
Structured, case-specific professional judgment and management rather than a population-wide actuarial probability of attack.
Multidisciplinary teams of law enforcement, analysts, mental-health practitioners, educators, or community partners gather and weigh evidence.
Processes vary by local institution and intervention. The federal public guidance does not establish one national notice or appeal mechanism for every referred case.
Good-practice guidance emphasizes reassessment and management; actual retention and closure rules vary across participating agencies and institutions.
Public research can identify recurring behaviors and support structured inquiry, but extremely rare attacks make individual probability prediction and causal prevention claims difficult to validate.
The main risk is drift from behavior-based, contextual inquiry toward profiling based on ideology, disability, grievance, protected speech, associations, or vague indicators.
Current FBI and Secret Service materials describe ongoing operational support, research, training, consultation, and multidisciplinary prevention guidance.
United Kingdom · London
GVM · VHA
The GVM used scores and red/amber/green bands associated with gang violence. The VHA uses explicit inclusion criteria and harm scoring across designated violence categories.
The VHA has central governance, periodic refresh, an SOP, published impact assessments, and human intelligence processes.
The Met publishes information about access and removal requests for the VHA. The former Matrix was criticized for weak notice, accuracy, retention, and sharing controls.
The VHA uses published inclusion and refresh rules; the page reports people entering and leaving assessment. The historical Matrix no longer exists.
The ICO identified serious data-protection problems in the former Matrix. MOPAC oversight and periodic publication apply to the VHA, but causal effectiveness remains a separate question.
Disproportionate representation was central to criticism of the GVM. The Met publishes quarterly VHA demographic data; as of July 3, 2026, 1,632 people met the criteria and 44.91% were self-recorded as Black, Black British, Caribbean, or African.
The GVM was decommissioned and the Met now uses the VHA. The current public VHA page, SOP, and quarterly figures remain available.
United Kingdom · County Durham and Darlington
HART
A supervised random-forest model described in public research as using hundreds of decision trees to classify two-year reoffending risk as low, moderate, or high.
The tool was advisory. A custody officer retained discretion and was expected to consider other available information.
A person could be informed about Checkpoint eligibility, but public material did not identify a dedicated mechanism to challenge the model category itself.
The reviewed public sources describe periodic model refresh and police-held data but do not establish a current public retention and deletion schedule for HART outputs.
Published validation examined classification accuracy and intentionally weighted “cautious” versus “dangerous” errors. That is not the same as proving that use of HART improved long-term justice outcomes.
Published debate focused on age, gender, postcode-derived variables, unequal error costs, and whether advisory use was meaningfully contestable.
Public evidence reviewed for this release establishes implementation and use but does not reliably establish whether HART remained operational in August 2026.
Germany · Bavaria
Pre Crime Observation System
Near-repeat pattern analysis over recent crime records. The reviewed public material describes rule- and pattern-based forecasting rather than person scoring.
Analysts review candidate warnings before operational use.
The forecast concerns an area rather than a named person. Public material reviewed here does not establish individual notice or a route-specific challenge process.
Not sufficiently established in the reviewed official records.
Official and promotional sources describe use and claimed utility; independent evidence does not establish a distinct causal crime-reduction effect attributable to PRECOBS alone.
The released method is place-based and does not directly score named people, but deployment and recorded-crime data can still distribute police attention unevenly.
Public official evidence reviewed for this release confirms historical operational use but does not establish the current 2026 deployment footprint.
United States · aviation security
TSA Secure Flight Program
Identity matching and screening allocation. Secure Flight is not accurately described as a single future-crime prediction model.
Potential matches and exceptions can involve TSA and Threat Screening Center personnel; downstream action follows agency authority.
Travelers can use DHS TRIP for screening and watchlist-related complaints. The public process does not reveal every underlying rule or record.
DHS privacy documents describe system records and protections; list modification and removal are owned by relevant list authorities rather than Secure Flight alone.
Public materials establish workflow and privacy controls. They do not publish a complete precision, recall, subgroup-error, or counterfactual security-effect evaluation.
Public demographic error and disparate-impact metrics are limited. Identity matching can affect people whose names or identifiers resemble listed records.
Current TSA and DHS privacy materials continue to describe Secure Flight as an active passenger prescreening program.
United States · Chicago, Illinois
SSL · CVRM · Party to Violence models · Heat List
Six successive statistical models estimated future “party to violence” involvement over an eighteen-month horizon. Earlier SSL versions produced scores; CVRM produced risk tiers.
Operational use occurred through police personnel and programs; the score did not itself constitute a warrant, charge, or adjudication.
Some people received Custom Notifications, but the public record did not establish a general score-notice, explanation, or direct appeal process.
The program was decommissioned; OIG identified governance and data-quality concerns and recommended stronger retention, correction, and removal controls for any future model.
OIG reviewed RAND evaluations and found major weaknesses in purpose, documentation, data quality, operational guidance, and evaluation. The public record did not demonstrate reduced shooting or homicide risk from the model.
Race was not a direct model variable in the reviewed versions, but arrest, victimization, gang, and policing records can carry structural disparities and proxies.
Chicago’s Office of Inspector General records that CPD decommissioned the Party to Violence risk-model program on November 1, 2019.
United States · federal interagency system
TSDS · terrorism watchlist · former TSDB
A federated nomination, sufficiency, identity-resolution, acceptance, designation, dissemination, encounter, modification, review, and removal lifecycle—not one self-executing list or one prediction model.
Nominating agencies, NCTC or FBI components, TSC analysts, and downstream agencies perform distinct review and action functions.
Direct notice is limited. DHS TRIP handles many travel complaints; review and disclosure vary by consequence, citizenship, and legal process.
The PCLOB report describes review, modification, redress, and annual removals while recommending more systematic re-review and accuracy controls.
PCLOB’s 2025 report describes the enterprise, approximately 1.1 million persons as of August 2024, and recommendations concerning accuracy, review, and redress. These counts do not provide a single false-positive rate.
The overwhelming majority of listed people were foreign nationals; under 6,000 were U.S. persons as of August 2024. Rights and remedies vary substantially by status and consequence.
The FBI renamed the Terrorist Screening Center to the Threat Screening Center in March 2025. The terrorism component remains a distinct watchlisting function within the broader mission.
No program matches the current filters. Reset the explorer or broaden the search.
Descriptive comparison
No overall score, vendor ranking, legality rating, effectiveness grade, or “most advanced” label is calculated.
Select two to four program records to compare their functions, consequences, human review, redress, evaluation, equality findings, and unknowns.
Synthetic feedback-loop laboratory
Change the type of designation and level of added attention. The exercise shows how deployment can influence the records available to the next analysis without claiming that every contact, report, or forecast is invalid.
Recorded incidents, contacts, intelligence, or case events form the available baseline.
A map, score, tier, match, or professional formulation redirects attention.
Patrol, screening, visits, review, investigation, support, or another response follows.
Contacts, detections, arrests, referrals, non-events, or corrections become records.
The next analysis may learn from records partly produced by the previous intervention.
A fictional forecast directs more observation to one grid cell. More patrol can reveal more events or officer-initiated records even if underlying event prevalence is unchanged.
The demonstration does not claim that all patrol creates bias or that reported incidents are invalid. It shows why deployment exposure must be measured separately from underlying event prevalence.
Predictive Enforcement Program Explorer ready.
Publication boundary
A system may be sophisticated but low-consequence, simple but coercive, discontinued but functionally replaced, or operational without public evidence of effectiveness. A dashboard, list, match, risk tier, and multidisciplinary case formulation must not be treated as interchangeable.
Operational means an output was available for live use. It does not mean a computer made the final legally consequential decision.
Review must be timely, informed, empowered to reject, and supported by access to underlying evidence and uncertainty.
Official and vendor records establish bounded facts about purpose, procurement, status, or claims—not independent effectiveness.
Names, vendors, models, and administrative owners can change while prioritization, data fusion, patrol allocation, or watchlisting functions continue elsewhere.
Review the evidence methodologyOpen the source librarySubmit a correction