The easiest version of crime prevention is also the most tempting.
Spend early enough, intervene intelligently enough, rehabilitate properly enough, and eventually fewer people will be harmed and fewer prison cells will be needed.
That was the starting idea behind this research.
The evidence did not earn the simple version.
It did something more useful: it split the package apart.
Some interventions survived strongly enough to define a bounded policy direction. Some remain promising and should be tested. Some matter for welfare, education, housing or opportunity without a demonstrated crime effect. Some popular ideas weakened when contemporary evidence was examined.
And one major idea failed altogether: there is no defensible way to add the surviving effects together and turn them into one national forecast for crime, victims, prison places or savings.
What remains is narrower, but stronger:
Use the strongest interventions where the evidence, population and safeguards fit. Test promising approaches instead of promising their effects. Keep worthwhile welfare policy on its real merits when crime effects are unproved. Stop or narrow interventions that fail. Make cost, capacity, rights and delivery visible before scaling anything.
This is not an argument that prisons are unnecessary. Some people require custody for punishment, incapacitation or immediate public protection. Nor can a long-term prevention strategy solve an acute capacity shortage next year.
The argument is simpler: prison construction should not become a substitute for understanding why demand exists, what prevention can actually change, and whether rehabilitation is being delivered well enough to work.
The prison total hides different pressures
The first part of the project looked at prison demand itself.
Using Ministry of Justice custody-state data, TWIS built a bounded annual prisoner-day accounting for 2015 and 2025. Under that construction, occupied prisoner-days rose from about 31.25 million to 31.85 million — roughly 0.60 million.
But the components moved in different directions.
Remand increased by about 2.15 million prisoner-days. Recall increased by about 2.56 million. Sentenced custody excluding recall fell by about 3.64 million. Non-criminal detention fell by about 0.47 million.
Those figures reconcile inside the accounting. They do not explain why the changes happened.
They are not a causal decomposition. They do not identify the effect of crime, policing, courts, sentencing, release rules or prevention. They do not tell us how many prison places any policy could avoid.
That is the first discipline this research imposes:
Do not use the prison population as a proxy for crime, and do not treat prevention evidence as if it already identifies an avoidable-prison counterfactual.
Youth diversion: the minimum necessary response
Youth diversion remains the strongest first-contact policy direction in the review, but only inside a clear boundary.
For suitable low-risk or limited-contact children, formal processing can itself create harm. The question is not whether diversion is always better than punishment. It is whether formal processing is genuinely necessary in this case, given offence seriousness, repeat behaviour, victim safety, safeguarding and public protection.
The current evidence sharpens that further by separating two decisions:
- should the child be diverted from formal processing?
- does the child need an additional programme, treatment or service?
Those decisions are often bundled together. They should not be.
The Re-Frame trial is a useful warning. It tested a short drug-education programme delivered to children suitable for diversion. It did not test diversion as a whole. Its result means an add-on programme cannot be assumed beneficial simply because it sits inside a diversion scheme.
A defensible approach is therefore the minimum necessary pre-court response, with extra intervention added only for a separate evidence, need, safeguarding or welfare reason. That is a bounded proportionality and evidence rule, not an experimentally proven ranking of every possible pre-court disposal.
Current administrative data show that diversion is already operating at material scale. Around 13,500 Youth Justice Service-involved diversionary outcomes were recorded in the year ending December 2025.
Current inspection and administrative evidence also show why scale alone is not enough: practice varies between areas, decision rationales are inconsistent, outcome data are limited and equality questions remain live.
So the policy needs conditions:
- keep serious, high-risk or persistent offending outside the automatic boundary;
- choose the least intensive disposal consistent with safety and proportionality;
- record why the disposal was chosen;
- measure net widening;
- monitor ethnicity, geography and repeat contact;
- protect victim safety and voice;
- do not make support compulsory merely because it is available.
That is not “diversion for everyone”. It is a bounded rule against unnecessary justice contact.
Prison education: stop confusing volume with success
Where custody is necessary, prison education remains one of the stronger rehabilitation directions.
The evidence does not say that every course reduces reoffending. The important question is what is delivered, to whom, at what quality, with what continuity and with what useful outcome.
Current Ministry of Justice evidence shows substantial educational and additional-needs complexity among prisoners. HM Inspectorate of Prisons has also documented weak purposeful activity and constraints involving staffing, space, infrastructure and budgets.
Those findings matter because an evidence-backed direction can easily become a target factory.
A system can increase course starts while offering qualifications with little labour-market value. It can interrupt learning through transfers. It can exclude prisoners with low literacy or SEND needs. It can count completion without asking whether a person actually gained useful skills.
The better rule is:
Treat prison education as a serious rehabilitation function, and judge it by access, educational progress, quality, continuity, useful qualifications, additional-needs support, labour-market relevance and final outcomes — not by volume alone.
England has substantial prison-education delivery architecture. That proves the system can commission provision. It does not identify a current cost per successful learner or tell us which programmes produce the best final outcomes.
The next prison-education statistics can improve monitoring of access and delivery. They will not, by themselves, prove a causal effect on reoffending.
Restorative justice: an offer, not an obligation
Restorative justice also survives, but only as a bounded option.
Victims must not be pressured to participate, forgive or meet an offender. Cases need professional screening for safety and suitability. A completed conference is not itself proof of a better final outcome.
So the defensible policy is an offer where appropriate:
- voluntary victim participation;
- professional facilitation;
- specialist risk assessment;
- safe withdrawal at any stage;
- offence and context suitability;
- no assumption that restorative justice replaces custody where punishment, incapacitation or public protection still justify it.
That is a credible policy direction. It is not an unconditional national effect claim.
CBT: an intervention name is not an intervention
Cognitive and behavioural interventions remain promising, but “CBT” is too broad to be a policy by itself.
Clinical need, practitioner competence, supervision, dose, engagement and service capacity determine what people actually receive.
The London Your Choice evaluation demonstrates the danger. Fewer than one per cent of participants received the intended dose.
That does not prove CBT is ineffective. It shows why saying a service is “CBT-based” tells us very little if the delivered intervention is materially different from the model being invoked.
The sensible position is to retain clinically appropriate, sufficient-dose CBT for further testing and scaling where evidence supports the specific model, while rejecting the idea that the label alone guarantees an effect.
Mentoring: the category needed splitting
Generic short-term mentoring does not survive as a demonstrated crime intervention.
The July 2026 REACH randomised trial does, however, make the category more precise. REACH provided six months of mentoring and found a small positive high-security effect on behavioural difficulties.
That matters. But it must not be stretched.
Offending and anti-social-behaviour findings were secondary, and school suspensions moved in an adverse direction. REACH therefore does not establish that “mentoring reduces crime”.
It supports a narrower conclusion: a specified longer-duration, high-fidelity mentoring model can produce a demonstrated intermediate benefit worth further testing.
So the policy split is:
- reject generic short-term mentoring as crime proof;
- retain specified longer-duration models for demonstrated intermediate benefits and separate offending evaluation.
Promising ideas that should remain tests
Several other approaches remain plausible enough to test without being ready to promise.
Targeted home visiting and intensive family models have model-specific evidence and can also stand on child-welfare or developmental benefits. Paid youth employment has promising international evidence, but the UK impact question remains unresolved. Focused deterrence has substantial international evidence, but selection, racial fairness, legitimacy and support capacity are part of whether the model is acceptable and transferable.
Intensive Supervision Courts are another clear example. England and Wales has expanded the model, but expansion is not outcome proof. The final impact and economic evaluation still needs to test reoffending, breach, recall, treatment completion, subgroup effects and whether treatment supply can meet court-imposed demand.
A programme can be active, funded and expanded while its causal effect remains unknown.
Good social policy does not need an invented crime effect
One of the strongest lessons from this project is that a humane policy does not need to be sold as crime prevention to matter.
School attendance, SEND support and alternative-provision quality can matter profoundly for children even though the current evidence does not establish a generic causal chain from broad reform to lower offending.
Employment can improve income, stability and dignity after justice contact even though the current UK evidence does not justify a national job guarantee as a demonstrated crime-reduction intervention.
Accommodation is the clearest example. The current matched Ministry of Justice analysis of CAS3 did not identify a statistically significant reduction in the reoffending rate or number of reoffences.
That removes the demonstrated crime-effect claim. It does not make rough sleeping or unstable release accommodation acceptable.
Keep the housing case where the housing evidence supports it. Do not manufacture prevented crime to make humane policy sound harder-edged.
What the evidence does not support
The research does not support a national adult First-Offence Guarantee inherited from youth diversion evidence. Adult police-diversion evidence remains mixed.
It does not support creating a new National Prevention Account by default while existing Treasury appraisal and Place Based Budget mechanisms are still being tested.
And it does not support the original integrated package model.
The surviving interventions use different populations, jurisdictions, outcomes, follow-up periods and causal designs. Some are alternatives for the same people rather than additive components. Several compete for the same staff, housing, treatment, school and employer capacity. Some survive as welfare policy rather than demonstrated crime interventions.
Adding their standalone effects would manufacture precision.
So this research does not produce:
- a national percentage reduction in crime;
- a national avoided-victim total;
- a number of prison places avoided;
- a national package cost;
- a package benefit-cost ratio.
Those are not missing calculations waiting to be filled in. On current evidence, the aggregate model itself is rejected.
Cost, capacity and equality are part of the intervention
A policy that works in a trial but cannot be delivered properly at scale does not retain the same effect by magic.
The research can identify real cost anchors and real delivery systems. It cannot yet build a defensible national package budget or simultaneous national capacity model.
That distinction matters because many of these interventions draw on the same constrained systems: clinicians, probation staff, housing, employers, schools, SEND support and local services.
Rights and distribution are equally central.
Diversion can reduce harmful formal processing and still widen low-level justice contact. Focused deterrence can reduce violence and still operate through unfair or opaque targeting. Attendance policy can improve engagement and still punish disability or poverty. Restorative justice can help and still harm if victim participation becomes expected rather than voluntary.
The subgroup evidence remains incomplete. That means unknown, not “probably equal”.
The strategy therefore has to measure who gets selected, who completes, who benefits, who is excluded, where services are available, whether coercion or breach increases, and whether outcomes differ materially between groups.
A long-term strategy should be designed to change
A 25-year strategy should not freeze a 25-year list of programmes.
It should protect the outcomes, evidence standards and correction mechanisms.
For each intervention, government should be able to say one of five things:
- use within bounds;
- test and evaluate;
- retain for welfare, not crime proof;
- do not generalise;
- stop or redesign.
That is more informative than calling everything “evidence-based”.
It also forces a distinction that public policy often blurs: promise is not delivery; delivery is not outcome; outcome is not automatically causal effect.
The research is finished for now
This project still records researchState:"more-research", because future evidence can change decisions.
But that phrase no longer means “keep searching”.
Routine current-public-evidence research is closed. A lane should reopen only when a named future evaluation arrives or genuinely new causal, cost, capacity, equality or safety evidence appears that could change the current decision.
That includes forthcoming UK results on paid youth employment, focused deterrence, Intensive Supervision Courts, restorative justice and adequate-dose CBT, as well as better prison-education outcome and cost evidence.
Until then, more browsing is more likely to add detail than change a decision.
What survives
The original idea was a generational prevention package strong enough that a future prison might never need to be built.
The evidence does not earn that claim literally.
It earns something more practical.
For suitable low-risk or limited-contact children, use the minimum necessary pre-court response and do not assume an attached programme helps simply because it is attached to diversion. In prison, judge education by access, quality, continuity and useful outcomes rather than volume. Offer restorative justice where victims freely choose it and the case is safe. Retain clinically appropriate CBT and specified longer-duration mentoring models where the evidence justifies testing, without turning programme names into national promises.
Keep housing, education, disability support and employment policy on their real social merits when crime effects are unproved. Keep cost, capacity, equality and public protection visible. Reject interventions and institutions that the evidence does not currently justify.
Most importantly, do not add incompatible effects together and call the result a forecast.
None of this proves that prevention explains national crime trends. None of it identifies a fixed number of prison places that will disappear.
It does something more defensible: it identifies which decisions the evidence currently supports, which remain tests, and what future evidence would justify changing course.
That is the part of the prevention argument that survived.