Top ↑

Police investigations are information problems before they are artificial-intelligence problems. A single case may require officers to reconcile interview notes, incident reports, CCTV references, receipts, phone records, property lists, forensic returns, vehicle descriptions and unanswered evidence requests. When those records are fragmented, investigative time is consumed reconstructing the file rather than testing the facts.

Saint Lucia is not starting from zero. A customised police records-management system is already reported as active through the regional CariSECURE work; officers have received recent training in intelligence analysis, crime-scene work, cybercrime and cryptocurrency investigations, and the justice system is adding digital case-management capacity. The right next move is therefore not a free-standing chatbot. It is a controlled support layer connected to approved records and built around the existing legal case file.

The first useful functions are deliberately modest: transcribe approved material, build a source-linked chronology, extract people, places, vehicles and objects, flag factual conflicts for human review, show missing documents, search authorised prior reports and draft routine summaries with every proposition linked back to its source. The system may suggest questions. It must never decide guilt, manufacture evidence, hide uncertainty or turn an unverified association into a fact.

SLPA proposes a staged Evidence Intelligence Pilot: test the system first on closed, quality-reviewed property-crime files; measure accuracy and time saved against a human baseline; move to live shadow mode only after legal, security and disclosure checks, and permit operational use only when an accountable investigator verifies every output. Predictive person-scoring, automated arrest recommendations, emotion recognition and uncontrolled facial recognition should sit outside the pilot.

Developed from an SLPA field observation about the information burden inside a routine Saint Lucia investigation. The design was checked against Saint Lucia’s digital-policing foundations, justice reforms, data-protection law and the INTERPOL–UNICRI responsible-AI framework in July 2026.

01

Better case records are the foundation for useful AI.

Hiring and training matter, but investigators also lose time when facts must be repeatedly copied, searched and reconciled across paper, messages and separate digital systems. Standard records, consistent identifiers and a complete audit trail are prerequisites for useful AI.

Evidence-led finding
02

Start with timelines, extraction and document comparison.

Chronologies, entity extraction, document comparison, request tracking and source-linked drafting are easier to verify than forecasts about who will offend. They can reduce administrative burden without converting probabilistic output into coercive state action.

Evidence-led finding
03

Every AI output needs a link to its source.

Every generated sentence, conflict flag and suggested lead should open the exact report passage, image, transcript segment or record that produced it. Unsupported output must be visibly labelled and excluded from the evidential record until independently established.

Evidence-led finding
04

Build disclosure and chain-of-custody rules into the system.

The system should preserve originals, access history, model and prompt versions, officer corrections, export history and the boundary between evidence and analysis. Prosecutors, courts and defence counsel need an intelligible route to test how a material output was produced.

Evidence-led finding
05

Small-state scale makes privacy controls more important.

Saint Lucia can run a bounded pilot, train a small user group and learn quickly. But a small population makes re-identification and informal access especially consequential. Role-based access, purpose limits, retention rules, breach response and independent audit are core operating controls.

Evidence-led finding
IIR baseline diagnosisRedesign the workflow, preserve accountable judgmentThe inherited human evidence workflow is slow and fragmented, but it performs rights-critical functions that an automated assistant must support rather than displace.3 functions · 5 dependencies

Institutional Intervention Readiness

What does the current arrangement do, and what depends on it?

This is a provisional SLPA function-and-dependency screen, not a legal or operational assurance.
Bounded inherited arrangement

The current processes for receiving, organizing, evaluating, disclosing, contesting and retaining police and prosecution evidence.

Current legitimate functions
  • Human evaluation of context, credibility and investigative significance
  • Source provenance, disclosure, contestability and accountable coercive decisions
  • Institutional memory with legal responsibility attached to named officials
Current harms or dysfunctions
  • Fragmented records, omission risk and repeated manual reconstruction
  • High administrative burden and weak retrieval across complex files
  • Inconsistent institutional memory
Critical dependencies
  • RSLPF and DPP authority
  • Defence disclosure and court process
  • Secure hosting and access control
  • Data protection
  • Independent benchmark and review capability
Choice-relevant unknown

Task-specific omission, hallucination, bias, access, security, disclosure and contestability performance has not been established for live Saint Lucian use.

RIPPLE-4 recursive reviewClosed-file pilot only; no operational scale without rights and accuracy gatesThe proposal now models investigator overreliance, adversarial behaviour and downstream court obligations. Citation accuracy alone is insufficient: omission, access, bias, disclosure and contestability each receive a separate gate.4 orders · 2 triggers

RIPPLE-4 recursive review

Trace direct effects, public responses, system effects and long-term consequences.

Decision after recursive reviewClosed-file pilot only; no operational scale without rights and accuracy gates

The proposal now models investigator overreliance, adversarial behaviour and downstream court obligations. Citation accuracy alone is insufficient: omission, access, bias, disclosure and contestability each receive a separate gate.

Instrument
A source-linked assistant benchmarked on reviewed closed files, with access control, immutable logs, disclosure rules, prohibited uses and human authority over every coercive decision.
No-policy counterfactual
Investigators continue manual chronology and contradiction work across fragmented files, with variable completeness and slow review.
Binding constraint Binding constraintThe scarce capacity, dependency or rule most likely to determine whether the policy can work.
Productivity gains cannot be purchased by weakening provenance, disclosure, privacy, equality before law or accountable human judgement.
O1Direct incidence+

Case material becomes faster to organise and cite

The system extracts chronology, entities and contradictions with links back to source material.

O2Adaptation+/-

Investigators and adversaries change behaviour

Users may rely on machine salience; offenders may exploit predictable collection or classification gaps.

O3Propagation+/-

Errors or gains move into prosecution, defence and courts

Disclosure, credibility and evidential decisions inherit the quality and auditability of earlier processing.

O4Inherited state+/-

Accountable capability or automation dependency forms

Good logging and contestability can raise institutional memory; opaque reliance can weaken judgement and trust.

Critical tail Critical tailA low-frequency or hard-to-reverse failure that deserves explicit protection even when the central case looks favourable.

Automation bias or a hidden data error shapes suspicion and propagates through charge, bail, disclosure and public trust before a human notices the original mistake.

Shared nodes Shared nodesPeople, systems or assets used by several policies at once. Overloading one can make otherwise sound projects fail together.
  • Digital Trust Layer
  • RSLPF case records
  • DPP and courts
  • Cybersecurity capacity
  • Public legitimacy
Adaptive policy triggers, automatic responses and accountable owners
Leading indicatorTriggerAutomatic responseOwner
Source, omission and contradiction benchmarkAny safety-critical error class exceeds its pre-registered toleranceStop the affected use case, preserve logs and retrain or redesign before retestIndependent evaluation lead
Unauthorised access or undisclosed model useAny verified material incidentSuspend access, notify the lawful oversight route and commence incident reviewRSLPF data owner + oversight

These are conditional causal pathways, not forecasts disguised as facts. A live appraisal must add evidence vintage, probability ranges, distribution and an authorised review date. The Sovereign Option review below converts this map into a bounded decision posture.

IIR final readiness gatesNot ready for live coercive usePermit closed-file benchmarking only. No live coercive use should proceed until every separate rights, accuracy, disclosure and security gate passes.1 met · 3 conditional · 1 blocked

IIR final readiness test

Is Saint Lucia ready to make this change?

RIPPLE-4 may reveal functions, groups or failure paths that require the baseline to be revised before authorization.
Inspect the five readiness gates
01Task-specific

Bounded causal account

Each permitted task must be defined and tested separately; a general claim of AI capability is insufficient.

02Live use fails

Rights and legitimacy

Live coercive use cannot pass before legality, equality, privacy, disclosure and defence contestability are demonstrated.

03Mapped

Dependencies and alternatives

Police, prosecution, defence, courts, data protection, cyber security and independent review are included.

04Human route required

Coverage and transition

Accountable judgment, source provenance, correction, audit and manual fallback must remain intact.

05Closed-file only

Learning and recovery

Benchmarking is controllable only with isolated datasets, logged outputs, stop rules and independent replication.

Function coverage

Any new tool must preserve human decision authority, immutable source links, defence access, disclosure, audit, correction and a manual route for every consequential task.

Transition controllability

Limit work to quality-reviewed closed files; isolate use cases; log every output and access; stop the affected task on a safety-critical error or material breach.

Strongest alternative

Replace human evaluation with an automated system. It may appear efficient but fails the rights, accountability and recovery floors.

What the IIR review changesPermit closed-file benchmarking only. No live coercive use should proceed until every separate rights, accuracy, disclosure and security gate passes.
Sovereign Option reviewPause operational use; test closed filesPermit benchmarked closed-file learning, but no live coercive use until rights, omission, disclosure, access and contestability gates pass.Commitment posture

Sovereign Option review

How far should Saint Lucia commit now?

RIPPLE-4 maps what the system does next. Sovereign Option Theory converts that map into a bounded decision.
Open the decision architecture
SLPA recommended decision posturePause operational use; test closed files

Permit benchmarked closed-file learning, but no live coercive use until rights, omission, disclosure, access and contestability gates pass.

Fix before action

Public purpose

Improve evidence organisation and institutional memory while leaving suspicion, credibility and every coercive decision with accountable humans.

Do not trade away

Viability floor Viability floorA condition Saint Lucia should not trade away while pursuing the policy, such as legality, fiscal resilience, safety or essential access.

Legality, equality, privacy, source provenance, disclosure, defence contestability, human judgement and an immutable audit trail.

Keep adaptive

Instrument

Model, vendor, interface, permitted tasks, hosting arrangement, benchmark set and workflow placement.

Minimum viable commitment Minimum viable commitmentThe smallest serious first step that can produce useful evidence without locking the country into the full exposure.

Stage 1 benchmark probe on quality-reviewed closed files; Stage 2 only after independent replication and security review.

Option and capability effect Option and capability effectWhether today’s choice expands or narrows the country’s future choices, skills, institutions and bargaining power.

Source-linked records, benchmark discipline and secure review capability can improve future investigation even if a particular model is retired.

Structural-transformation test Structural-transformation testWhether the policy builds sustained productivity, knowledge and domestic capability rather than only financing consumption or importing a finished system.

Imported technology adds national value only when local institutions can understand, govern, test, replace and improve it.

Evidence required for the next commitmentStop the affected use case on any verified safety-critical error or material access breach; scale only a task that passes every separate gate.

Current public-safety signal

RSLPF reported declines in five major offence categories.

% reduction reported

Percentage change for 1 January–10 July 2026 against the corresponding 2025 period. These administrative figures describe reported offence counts; they do not by themselves measure investigative quality, detection, case attrition or the effect of any technology.

Use this data

Copy the visible figures or download them with the unit, claim label and method note attached.

SLPA policy proposal

RSLPF Evidence Intelligence Pilot

SLPA–10 / DRAFT

Build a secure, human-accountable evidence assistant that makes case material easier to organise, search, test and disclose, without delegating suspicion, guilt or coercive decisions to a model.

01

One secure evidence system

Operate inside an approved environment connected to the police records system, with role-based access, encryption, retention rules, immutable originals and no use of public consumer AI services for identifiable case material.

02

Source-linked case summaries

Generate timelines, entity tables, contradiction matrices, missing-evidence lists and draft summaries only when every output can be traced to an authorised source and its confidence is visible.

03

Human approval for every decision

Keep an accountable officer responsible for verification and every investigative decision. Ban automated guilt scores, arrest recommendations, witness-credibility ratings and unsupported person-risk rankings.

04

Full audit access across the justice system

Preserve model versions, queries, outputs, corrections and exports so prosecutors, courts, defence counsel, auditors and authorised oversight bodies can examine material uses.

05

Test results before wider use

Benchmark closed files, then run a live shadow pilot in a bounded case class. Expand only if accuracy, security, timeliness, workload, fairness and user discipline meet published thresholds. Separate source accuracy, omission, contradiction, access, bias, disclosure and contestability gates. Preserve immutable logs and prohibit any automated suspicion, credibility or coercive decision.

01First 100 days

Prepare the records and benchmark the tool

  • Map the current police-record, evidence, forensic, disclosure and court handoffs; identify the authoritative system and every uncontrolled copy.
  • Complete legal, privacy, cyber-security, records, procurement and human-rights impact assessments using the INTERPOL–UNICRI toolkit.
  • Create a de-identified benchmark of quality-reviewed closed property-crime files and score timeline, extraction, contradiction and citation accuracy against trained investigators.
  • Pre-register safety-critical error tolerances and the automatic stop, evidence-preservation and independent-review response for each affected use case.
02Months 4–12

Run a live shadow test with independent review

  • Select a small trained unit and a bounded lower-risk case class; keep model output outside the official file until an investigator verifies and adopts it.
  • Test source linking, request tracking, disclosure export, chain-of-custody logs, access alerts, failure reporting and supervisor review under real operating pressure.
  • Publish aggregate pilot results, material error types, security incidents and the decision to stop, redesign or advance, without exposing victims, witnesses or investigative methods.
03Year 2

Expand the functions that pass every test

  • Allow only the functions that passed accuracy, legality, security and workflow tests; retrain users before adding a new evidence type or case class.
  • Create standing judicial, prosecutorial, defence, privacy, technical and community review routes for material system changes.
  • Join verified caseflow measures to the wider justice dashboard so speed, disclosure quality, fairness and case outcomes are governed together.
Public accountabilityMeasures for public accountabilityQuarterly operating signals and one independently reviewed annual outcome report.7 measures
01Median time to a verified first chronology

Tests whether the system returns investigator time without rewarding rushed conclusions.

02Material outputs accepted, corrected or rejected

Makes accuracy and human correction visible by function and evidence type.

03Source-link and disclosure completeness

Tests whether every consequential output remains inspectable across the justice chain.

04Unauthorised access, leakage and security incidents

Treats privacy and evidence integrity as operating outcomes, not policy language.

05Case-stage time, attrition and quality measures

Checks whether administrative savings translate into stronger, fairer case movement.

06Safety-critical omission, contradiction and source-link error by class

Prevents a strong average score from hiding the error that could propagate through a case.

07Unauthorised access, disclosure completeness and appeal or correction outcomes

Tests the institutional and rights effects beyond investigator productivity.

4 limits · 6 sources

Limits of this analysis

  • The reported 2026 offence reductions are RSLPF administrative comparisons and do not establish causation, detection rates or investigative performance.
  • AI can amplify incomplete reports, historic bias and data-entry errors. Human review does not make a weak system safe unless reviewers have time, training, authority and recorded accountability.
  • This brief proposes a pilot architecture, not a claim that a specific model or vendor is ready for evidential use. Procurement should follow testing requirements rather than determine them.
  • Some investigative techniques and security controls cannot be published in operational detail. Public accountability can still cover purpose, prohibited uses, governance, aggregate performance and rights protections.