Understand the topic
What this means in daily life
AI for women’s safety should be examined as a support tool for accountable staff. The proposal keeps uncertainty, privacy and the ability to challenge an output visible instead of treating a computer prediction as a judgment about a person.
The cost of an error depends on what staff do with the output. Sorting duplicate reports and identifying a person as suspicious carry very different consequences; the proposed evaluation must examine the actual decision and the route for correcting harm.
Data reference table
Evidence Snapshot
Read each value with its reporting period, population and limitations. A historical national reference is not a local baseline or proof that an IFT proposal works.
5 measurement records still need data and a review of their relevance.
| Signal | Reference Value | Period / Population | How to Read It | Source Status |
|---|---|---|---|---|
| Report-to-acknowledgement time | Not established Data required — not zero | Reporting period: Information Required Population and geography: Information Required | Proposed measure; no observed value is established here. Limitations: Define the measure, unit, reporting period and covered population; obtain a traceable dataset and review its suitability for this topic before using a number. | Source organisation: Information Required Direct source URL: not yet verified Source publication date: not established Verification status: Data required |
| Evidence preservation success | Not established Data required — not zero | Reporting period: Information Required Population and geography: Information Required | Proposed measure; no observed value is established here. Limitations: Define the measure, unit, reporting period and covered population; obtain a traceable dataset and review its suitability for this topic before using a number. | Source organisation: Information Required Direct source URL: not yet verified Source publication date: not established Verification status: Data required |
| Offline-risk escalation | Not established Data required — not zero | Reporting period: Information Required Population and geography: Information Required | Proposed measure; no observed value is established here. Limitations: Define the measure, unit, reporting period and covered population; obtain a traceable dataset and review its suitability for this topic before using a number. | Source organisation: Information Required Direct source URL: not yet verified Source publication date: not established Verification status: Data required |
| False-positive/negative rates | Not established Data required — not zero | Reporting period: Information Required Population and geography: Information Required | Proposed measure; no observed value is established here. Limitations: Define the measure, unit, reporting period and covered population; obtain a traceable dataset and review its suitability for this topic before using a number. | Source organisation: Information Required Direct source URL: not yet verified Source publication date: not established Verification status: Data required |
| Appeal and correction time | Not established Data required — not zero | Reporting period: Information Required Population and geography: Information Required | Proposed measure; no observed value is established here. Limitations: Define the measure, unit, reporting period and covered population; obtain a traceable dataset and review its suitability for this topic before using a number. | Source organisation: Information Required Direct source URL: not yet verified Source publication date: not established Verification status: Data required |
Different surveys, administrative records, estimates and reporting years are not interchangeable. No combined IFT score is calculated from these rows.
Evidence coverage and review log (CSV) · Structured evidence register (JSON)
10 ideas for this topic
What could be done — and what needs testing
These are proposals for research and local testing. They are not claims that a project has been funded, implemented or proved successful.
One defined AI task
Choose a limited task such as sorting duplicate service reports before evaluating a tool.
What to check: Reject undefined multipurpose prediction claims
Human decision record
Record who accepted or rejected an AI suggestion and the information used in that decision.
What to check: Check human reasons for consequential decisions
False-alert review
Examine incorrect alerts and the consequences they could create for women or wrongly identified people.
What to check: Assess harms caused by false alerts
Missed-case comparison
Compare the tool with existing work to find important cases it overlooks.
What to check: Compare missed cases with existing practice
District transfer test
Check performance before using a model in a different language, place or population.
What to check: Evaluate local language and population differences
Training-data permission
Review whether source records can lawfully and ethically be used for the proposed analysis.
What to check: Verify lawful permitted training-data use
Restricted identity access
Separate identifying records from the information needed for routine planning analysis.
What to check: Test role-based access to identity information
Plain-language challenge route
Let affected users question an output and obtain human review through an understandable process.
What to check: Check that challenges reach an accountable human
Vendor exit rehearsal
Test whether the authority can retrieve its records and continue essential work if the supplier leaves.
What to check: Rehearse data retrieval and essential service continuity
Harm-based stop rule
Agree which errors, exposure or unfair effects require suspension before any wider deployment.
What to check: Suspend use when agreed harm limits are exceeded