Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

London's Strategic Insights Tool

One shared memory and thirty-three readers: consolidating a city's rough-sleeping records

A city merges three separately governed record systems - street-outreach contacts, charity casework, and borough statutory applications - into one linked picture of who is sleeping rough, and lets all 33 local authorities read it. Modeled on London's Strategic Insights Tool for Rough Sleeping. It scores no one and decides nothing about any individual; it only consolidates. But the matcher misses roughly 9 in 100 true cross-system matches, by its own report - and because every borough reads the same shared layer, that undercount is not one team's local error, it is the whole city's blind spot at once. The real question is not whether to buy a sharper matcher. It is whether anyone checks the linked layer back against its sources, and whether the readers know the figures run low.

Stylized model of a documented deploymentHousing & homelessness services

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the London-SIT-class rough-sleeping data-consolidation layer network: 8 components and 15 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 3 assumed · 3 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    This models the memory-consolidation pattern documented in the London Rough Sleeping Strategic Insights Tool case file - not a reconstruction of the actual service. The atlas-relevant object is the topology: three separately governed record stores (CHAIN street-outreach, In-Form charity casework, H-CLIC borough statutory applications) probabilistically linked into one shared journey layer read across all 33 London local authorities. The tool makes no individual-level decisions - it is aggregate-only, users cannot access individual records other than their own organisation's, and the delivery team states it is not a substitute for published data and reports. Any reading that implies individual risk-scoring or case adjudication misreads it.

  • baseline

    The defining structural property is city-wide contamination coupling: the matcher accepts an association only above an 85% probability threshold chosen to minimise false positives, and the project's own Phase 2 DPIA (v2.0, 23 October 2023, published on the LOTI site April 2025) reports 91% recall, conceding that roughly 9 in 100 true cross-system matches are missed so that 'numbers subsequently appear lower in places where they should be higher', with recall varying as new data of varying quality is ingested. Because every borough reads the same consolidated layer, that undercount is inherited simultaneously across the whole city rather than staying local to one organisation. The accuracy figures are self-reported by the delivery team; no false-positive rate is published; Splink is named as the matching library in the techUK write-up alone.

  • baseline

    The two load-bearing absences are drawn as inactive checks. First, the record-against-record reconciliation: there is no independent reconciliation of the consolidated layer's recall against the source stores - the conceded undercount is self-reported and drifts with data quality, and nothing routinely reads the linked copy back against its sources. Second, the independent peer evaluation: no independent evaluation of the tool's decision impact exists - the GLA Chief Digital Officer's first-year claims (that the tool helped commissioners challenge assumptions and revealed previously hidden connections between street homelessness and Housing Options services) are qualitative, with no published metrics or counterfactual. These are the case's distinctive safety shape: a strongly governed process (a published data protection impact assessment (DPIA), small-cell suppression, a minimum data set, retention rulings) whose one unmeasured surface is the accuracy of the linkage everyone consolidates onto.

  • assumed

    The privacy surface is real and is drawn on the read and write pathways. The linked identifiers are fuzzy-matched names, National Insurance numbers, dates of birth, and phone numbers; data subjects (people sleeping rough) are not individually notified, processing rests on legitimate-interests / substantial-public-interest and research bases, and individuals have no direct interaction with the tool. During user testing a reader found that stacking filters could reduce a visualisation to 1-2 individuals, which led to ONS-style small-cell suppression (any output of 5 or fewer shows as '5 or fewer'). The DPIA's phase-1 review found several sensitive fields (substance misuse, mental health, pregnancy status, prison history, care-leaver history, benefit entitlement) went unused in the MVP but were retained, not visible to users, for planned features - so what was deleted was out-of-scope and unmatched service-provider data, not those retained sensitive fields.

  • baseline

    Governance consolidation followed the data: the build vendor (Faculty, sole data processor to 2 February 2024) handed ongoing hosting, management and maintenance to Homeless Link, which already operates the CHAIN source system - concentrating a source store and the consolidated layer under one GLA-contracted operator. A 5-year per-individual retention cliff aligns memory decay with how the DLUHC rough-sleeping indicators are defined (someone not seen for 5 years stops being treated as an existing rough sleeper), so long-cycle returners re-enter as new individuals - itself a structural undercount of the recurrence the tool was built to measure.

  • assumed

    Served people - people sleeping rough, and the services they do or do not receive - are not in the dynamics; this Lab reads institutional propagation only. No casework or commissioning outcome to any person is computed here. The harm surface the model reads is aggregate-statistical: a systematic cross-system undercount, a re-identification risk, and the epistemic question of whether operators treat the figures as authoritative despite the conceded undercount. Adoption figures (151 onboarded users, 40-45 organisations) and first-year impact framing are vendor or first-party claims, not independently audited; predictive demand-forecasting is a stated future ambition, not a deployed feature. A count, a journey, or a trend on this map is an institutional signal, never a person.

What this example does not show

  • Served people - people sleeping rough, and the services they do or do not receive - are not modeled here; the Lab reads institutional propagation only. This tool makes no individual-level decisions at all: it is aggregate-only, users see population trends (plus their own organisation's records), and no casework or commissioning outcome to any person is computed here. The harm surface is aggregate-statistical - a systematic cross-system undercount, a re-identification risk, and whether operators over-trust the figures - never an individual determination.
  • The 85% match-acceptance threshold, the 91% recall, and the conceded roughly-9-in-100 miss rate come from the project's own Phase 2 DPIA (v2.0, 23 October 2023, published on the LOTI site in April 2025) and a co-authored LOTI/vendor blog, not from an independent audit; no false-positive rate is published, and no independent evaluation of the tool's decision impact exists. Adoption figures (151 onboarded users, 40 to 45 organisations) and first-year impact claims are vendor or first-party statements, not independently verified. A safe-looking baseline is a property of this model, not a safety promise for any real deployment.
  • Counts vary across the cited sources - boroughs are reported as 33 (techUK) or 32-plus-City-Hall (the CDO retrospective and the DPIA controller table), and contributing organisations as roughly 12, then 14, then 40 to 45 - because they measure different dates and different things (service providers versus all contributing organisations); '33 London local authorities' is the safe formulation. Predictive demand-forecasting and additional datasets (probation, health and care, evictions) are stated future ambitions in every source, not deployed features.
  • This is a record-linkage and consolidation system, not a risk-scoring, generative, or determination system, and it is deliberately restricted to cohort and population-level insight; any reading that implies individual scoring or case adjudication misreads it. Its atlas relevance is the memory-consolidation topology and the city-wide correlated-undercount dynamic, and the mayoral funding decision that expanded the CHAIN source system (MD3331) does not itself name the tool.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • London's Strategic Insights Tool for Rough Sleeping probabilistically links records from three separately governed systems - CHAIN street-outreach contacts, In-Form charity casework, and H-CLIC borough statutory applications - into a single rough-sleeping journey per person that is read, in aggregate form only, across all 33 London local authorities; the tool makes no individual-level determinations, and after the build vendor's data-processor contract ended on 2 February 2024 the Greater London Authority contracted Homeless Link, which also operates the CHAIN source system, for its ongoing hosting, management, and maintenance.

    empirical
    • Trade press techUK, Rough sleeping insights tool: Using machine learning to support decision-making across London (2024) https://www.techuk.org/resource/rough-sleeping-insights-tool-using-machine-learning-to-support-decision-making-across-london.html
    • Government LOTI, GLA and London Councils, Phase 2 Rough Sleeping Strategic Insights Tool DPIA (public version, v2.0, 23 October 2023) https://loti.london/wp-content/uploads/2025/04/Phase-2-Rough-Sleeping-Strategic-Insights-Tool-DPIA-public.pdf
    • Government London Office of Technology and Innovation (LOTI), Rough Sleeping Insights Project (2023-2025) https://loti.london/projects/rough-sleeping-insights-project/

  • The Strategic Insights Tool's matcher accepts an association only above an 85% probability threshold chosen to minimise false positives, and the project's own Phase 2 Data Protection Impact Assessment reports 91% recall - conceding that roughly 9 in 100 true cross-system matches are missed so that 'numbers subsequently appear lower in places where they should be higher' and that recall varies as new data of varying quality is ingested; no false-positive rate is published, the accuracy figures are self-reported by the delivery team, and no independent evaluation of the tool's decision impact exists.

    empirical
    • Government LOTI, GLA and London Councils, Phase 2 Rough Sleeping Strategic Insights Tool DPIA (public version, v2.0, 23 October 2023) https://loti.london/wp-content/uploads/2025/04/Phase-2-Rough-Sleeping-Strategic-Insights-Tool-DPIA-public.pdf
    • Government LOTI (Anna Humpleby) and Faculty (James MacTavish), Using AI to better understand and support homelessness interventions in London (2025) https://loti.london/blog/ai-to-better-understandhomelessness-interventions-in-london/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Housing & homelessness services domain page.

Levers available here and the patterns behind them

Documented case histories