Domain Atlas / Child welfare & family services
Insight Bristol / Think Family Database
Bristol's Think Family Database drew on roughly 30 to 35 fused council, police and other datasets covering about 55,000 families (some 170,000 residents in 2021 reporting), and its child sexual and criminal exploitation risk models were quietly withdrawn in 2023 as 'not fit for operational use' after an independent evaluation judged the risk-scoring models the weakest element and staff reported victims of exploitation scoring below people involved in burglary; FOI responses indicate no record was kept of why the models were switched off, and auditors could not locate their source code or variable lists.[5]
What happened
Insight Bristol, a joint data-analytics team of Bristol City Council and Avon and Somerset Police, was formed in 2015 under the national Troubled Families (now Supporting Families) Programme, and built the Think Family Database — a shared store fusing roughly 30 to 35 datasets from council, police and other public agencies on about 55,000 families (some 170,000 residents by 2021 reporting, with around 450 staff able to access it). On top of that store the team built at least three predictive models — a Child Sexual Exploitation model, a Child Criminal Exploitation model, and a NEET model — scoring on a 1 to 100 scale; the CSE model, introduced in 2016, drew partly on anonymised data covering roughly 1,000 children known to have experienced sexual abuse (the CCE model's start date is given variously as 2016 or 2019). Researchers and a 2021 Centre for Data Ethics and Innovation review warned that several inputs acted in practice as proxies for poverty and could produce indirect discrimination, and that data gathered through "legal gateways" meant "legality is not the same as legitimacy." An independent 2023 evaluation judged the risk-scoring models the "weakest element" of the database, staff concluded the CSE and CCE models were "not fit for operational use" — with reports that victims of exploitation scored below people involved in burglary — and the two exploitation models were quietly withdrawn in 2023; FOI responses indicate that neither authority retained records explaining why, and that the models' source code and variable lists could not be located when auditors sought them. The NEET model and the underlying database remain operational as of 2026, the NEET model documented in a 2025 GOV.UK Algorithmic Transparency Record as advisory indicators only that should "not replace professional judgement." In June 2026 the Children's Commissioner for England said it was "deeply concerning that tools used by staff were not fully understood, trusted, and may have compromised their professional judgement."
The sociotechnical reading
Most cases in this Atlas fail through over-reliance — a human waves the machine through. Bristol failed the other way. Front-line staff distrusted the scores enough to stop citing them, yet the system kept running: kept collecting across roughly three in five families with children in the city, kept writing to a record read across agencies, and kept scoring from a substrate no one could later reconstruct. In map terms the store is the center of gravity — a fused cross-agency database feeding several models, not any one clever model — and the two pathways that would have made it governable were dormant. The independent second read had gone quiet (an ethics committee that reportedly stopped revisiting the analytics after 2017; outside scrutiny that arrived years late), and no reconciliation ever checked the distrusted models against ground truth. When auditors finally looked, the institution could not produce its own system: no source code, no variable list, no record of why two models were switched off. The lesson is distinct from Oregon's clean, documented discontinuation and from Arkansas's un-explainable determination: a distrusted tool is not a harmless one — it still surveils, still shapes the record, still consumes governance attention — and a system its own owners cannot audit can never be bound by anyone else. The Practice Library's provenance labeling, connection authorization (legality is not legitimacy), oversight cadence, and "understand the system" patterns all trace back to a failure of exactly this shape.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.