Skip to main content
lab-notes.aiPublic-service AI playbooks
Opportunity 15/17Community safetyJustice and community safety

Violence Risk Pattern Research

Explore the evidence and safeguards violence-risk research would require without creating person-level stand-in data.

Published sources
2
Starter data
None — by design
Domain build partner
build-violence-risk-research
Last reviewed

From the draft strategy

Opportunity

The draft strategy names linking justice and health records to identify the people most at risk of violence against women and girls as a potential public-service application, without establishing that it is necessary, that it is safe, or that there is something worth being identified for at the end of it.

Where the draft says it

Table 2 — potential public-service applications

Read the draft strategy source (opens in a new tab)

Source register

Research already done

Real published source

We reviewed these published sources so you do not have to start from zero. Each source shows what it covers, how it can be accessed, and why it matters.

2 sources

  1. The Executive Office

    Open

    Ending Violence Against Women and Girls

    What it covers
    The Northern Ireland strategic framework on ending violence against women and girls, its delivery plans, the commissioned research behind it, and the funding routed to community and voluntary organisations.
    Why it matters here
    It is the published work this example would have to sit inside, and it describes an approach built on services and prevention rather than on identifying individuals.

    Open the source (opens in a new tab)

  2. Department of Justice

    Open

    Statistics and Research

    What it covers
    The department's collection of justice statistics and research, including reoffending bulletins, the Safe Community Survey, and research undertaken with victims and witnesses of crime.
    Why it matters here
    It is what is actually published in this area, and everything in it is counts and survey findings, which is a long way from the linked person-level records the strategy's example assumes.

    Open the source (opens in a new tab)

Starter data

Starter dataset

No synthetic dataset — by design

Why we stopped

Every stand-in we sketched for this task described individual people, because a risk model's input is a person's history; the aggregate bands that would be safe to publish are exactly the ones that make the task meaningless.

What responsible work needs instead

Anyone taking this further needs formal research access under ethics and data-protection approval, with domain experts and people with lived experience owning the question from the start rather than reviewing an answer at the end.

Clone and build

Domain build partner

Build with a domain-aware coding agent

Clone the repository, then ask your coding agent to load the skill that ships with it. The skill brings the opportunity, sources, starter data, known unknowns and constraints into the conversation before anything is proposed. It follows the open Agent Skills standard, so it works in Claude Code, Codex and any other agent that reads skills.

01 / Clone

git clone https://github.com/Hypership-Software/lab-notes.ai.git

02 / Invoke

Claude Code
/build-violence-risk-research
Codex
$build-violence-risk-research

Conditions, not a checklist

Before you build

These are not footnotes. They are conditions that any responsible prototype would need to address.

  • Linked records create exceptional sensitivity

    Linked justice and health records are exceptionally sensitive, and a risk label seen by the wrong person can leave someone in more danger than they were in before.

  • Both kinds of error cause harm

    False positives bring attention to people who did not need it while false negatives quietly withdraw help from people who did, and both happen at the same time.

  • Reporting patterns are not behaviour

    Institutional reporting patterns are not individual behaviour, so anything trained on who was recorded will mostly learn who gets recorded.

  • Safe services must exist first

    Being identified is only useful if there is something safe to offer, and whether that exists is a question about services rather than about models.