Human oversight
AI-assisted triage supports human judgment. People remain accountable for interpretation, escalation and action.
Technology
Child-safety intelligence expressed through
SAFE.
SAFE’s intelligence layer is built around structured child-safety logic, frontline language, expert review and a long operating history. It supports risk visibility and governed review. It does not replace human responsibility.
The system at a glance.
Six capabilities sit beneath the SAFE platform. They work as one operating layer for child-safety intelligence.
Structured assessment
Frontline concern is captured through consistent fields, decision logic, contextual information and structured pathways, not free-form notes that get lost.
Language & pattern intelligence
Machine-assisted analysis helps surface repeated concerns, emerging patterns and signals that may need closer review.
Human-in-the-loop expert review
Qualified child-safety expertise remains central.
SAFE supports judgement, it does not replace it.
Evidence & audit trail
Concern history, review notes, decision context and action records are preserved to support internal, regulatory and legal review.
Dashboards & system visibility
Service, operator and de-identified system-level views help leaders see where concern and vulnerability may be emerging.
Sovereign & secure
SAFE is designed around role-based access, auditability, data minimisation, privacy-aware workflows and clear governance.
Machine-assisted.
Human-governed.
In child-safety settings, AI must be used carefully, transparently and with human accountability. SAFE supports attention, review and prioritisation. People, services, professionals and statutory authorities remain responsible for action.
Model validation
Performance is tested, reviewed and improved over time using appropriate validation methods and labelled child-safety data.
False-positive & false-negative discipline
SAFE understands that both missed risk and over-escalation carry consequences. Thresholds and triage logic are built around that trade-off.
Explainability & reviewability
Outputs support meaningful human review rather than opaque black-box reliance. Reviewers can see why a matter surfaced.
Bias & fairness monitoring
The system should be monitored for unfair or inappropriate patterns across populations, services and contexts over time.
Data-quality controls
Model quality depends on data quality. Structured capture, review discipline and user-behaviour guardrails are built in.
Version control and audit trail
Changes to models, thresholds, decision logic and prompts are governed, versioned and documented for review.
Clear use boundaries
SAFE strengthens the system around people, regulators and statutory bodies. It does not replace them and it does not publicly profile children.
The data moat
A 12-year compound learning loop.
SAFE’s risk-intelligence capability has been shaped by real-world child-safety data, structured assessment, expert review and operating use over more than a decade. That is much harder to replicate than code.
Operating evidence
68,943
Children assessed and de-risked through SAFE’s predecessor system by 2020.
Continuous horizon
12yrs
Child and family data infrastructure, AI-assisted triage maturity, and dedicated child-safety AI infrastructure development over time.
01
Frontline concern
Real staff record real observations, behaviours, disclosures, uncertainty and context.
02
Structured assessment
Concern is organised into usable, reviewable child-safety information - not free-form notes.
03
Free-text language
SAFE captures how child-safety concern is actually described in practice, in human words.
04
Expert review
Qualified child-safety judgment adds context, pattern and review feedback to the record.
05
Model & product learning
ML, NLP and product design are refined through accumulated operating experience.
06
Compound advantage
Each year of operation deepens the system’s understanding of risk language and workflow reality.
Eight layers,
one safety pathway.
SAFE’s architecture maps directly to how risk moves through a service: capture, analyse, review, evidence, escalate and learn. Each layer is designed to be infrastructure, not a feature.
- 01
Secure access &
Identity · permissions · service scope
service registrationServices and organisations access SAFE through controlled registration and role-based permissions. Different users see different surfaces depending on function and responsibility.
Access - 02
Structured assessment layer
Fields · decision logic · pathwaysFrontline concern is captured through consistent fields, contextual information and structured pathways.
Capture - 03
Free-text &
How concern is actually described
language captureSAFE captures the language staff use to describe concern, uncertainty, behaviour, context and emerging risk.
Capture - 04
Risk-intelligence layer
ML · NLP · prioritisationMachine learning and language analysis support triage, pattern recognition and prioritisation of matters that may need closer review.
Intelligence - 05
Expert review layer
Qualified child-safety judgmentRelevant matters are reviewed by qualified child-safety expertise or authorised review pathways. Humans remain responsible for interpretation, escalation and action.
Intelligence - 06
Evidence &
Time-stamped · defensible · reviewable
audit-trail layerConcern history, review notes, decision context and escalation-ready records support internal, regulatory and legal review.
Evidence - 07
Dashboard &
Service · operator · system view
intelligence layerService, operator and de-identified system-level visibility help show where concern, vulnerability and pressure are concentrating.
Evidence - 08
Integration & API layer
Gateway · microservices · modulesFuture integration pathways can support service systems, operators, regulators and sector partners where appropriate.
Roadmap
From product
to infrastructure.
SAFE’s roadmap strengthens the system around the turnkey product: deeper dashboards, clearer review workflows, stronger integrations and sector-specific intelligence layers.
Stronger dashboards
Service-level, operator-level and de-identified system-level visibility, with heat-mapping of emerging concern.
Improved user experience
Simpler onboarding, clearer workflows, better prompts and stronger usability for frontline staff under pressure.
Enhanced ML & NLP
Improved risk-intelligence support from structured data, free-text language, expert review and ongoing learning.
Expert review workflow
Better triage queues, review notes, decision context and escalation support for qualified reviewers.
API gateway
Integration pathways for service systems, operators, regulators, insurers and sector partners where appropriate.
Microservices direction
Scalable architecture that supports independent deployment, flexibility and future modules.
Sector modules
Adaptations for schools, sport, disability, healthcare, aged care and other vulnerable-population settings.
International licensing
Product and compliance pathways for comparable jurisdictions - New Zealand, the UK, Canada and US state-based systems.
Children shouldn’t depend on luck to feel safe.
SAFE operates in one of the most sensitive data environments there is. Privacy, security, access control and careful language are not optional - they are the product.
The intelligence layer is framed as de-identified, system-level visibility, used for planning, oversight and early support - never as surveillance of children.
Role-based access, auditability and data minimisation are built in. Outputs are reviewable rather than opaque, and AI use is bounded by clear governance.
Technical & AI diligence notes.
The numbers SAFE reports are scientific measurements, not marketing claims. Every published figure carries the cohort, method and as-of metadata required for reproducibility and audit.
Model quality & reproducibility
Reported as measurements, not marketing.
The numbers SAFE reports are scientific measurements, not marketing claims. Every published quality figure should carry cohort, method and as-of metadata required for reproducibility and audit.
Accuracy
95.08%
Share of predictions that match expert verdict across the benchmark cohort.
PRECISION
100.00%
When SAFE raises a flag, how often that flag is warranted.
Recall
84.66%
Of genuinely at-risk children, the proportion the system saw. Most consequential metric.
F1
91.69%
Harmonic mean of precision and recall, a balanced view that penalises lopsided performance.
Training & selection
The risk model is a Random Forest classifier built in R using the caret framework. Training uses repeated 10-fold cross-validation: historical data is partitioned into ten folds, each held out in turn, with the procedure repeated across multiple random seeds.
Models are selected on ROC-AUC - the threshold-independent measure of how well the model separates at-risk from not-at-risk cases. ROC-AUC is preferred to raw accuracy because it is robust to the class imbalance intrinsic to child-safety data.
Evaluation cohort
Quality claims are evaluated only on the qualified validation benchmark cohort - records where a qualified child-safety reviewer has issued a verdict. This is the cohort against which the model's predictions can be benchmarked against ground truth.
Operational workflow metrics (review capacity, queue health, time-to-verdict) are tracked separately and are not blended into the quality numbers - the same governance discipline applied elsewhere in SAFE to separate scientific quality from throughput quality.
Why accuracy alone is not enough
A metric suite, not a single headline.
In child safety, the cost of a missed at-risk child is not equivalent to the cost of escalating a not-at-risk child. SAFE reports quality through multiple measures to avoid single-metric bias in decision support.
Precision
TP / (TP + FP)
When SAFE raises a flag, how often that flag is warranted.
Recall
TP / (TP + FN)
Of children genuinely at risk, what proportion the system sees.
F1 score
2·TP / (2·TP + FP + FN)
The harmonic mean of precision and recall.
ROC-AUC
threshold-independent
A threshold-independent view used for model selection and trend comparison over time.
Evidence integrity controls
Cohort separation, exclusion rules, threshold discipline, as-of timestamping and method notes support governance.
Interpretation for stakeholders
SAFE supports two parallel readings of performance: scientific quality and operational quality.
Evidence integrity controls
- Cohort separation. Quality metrics on the benchmark cohort, operational metrics on the full population - never blended.
- Exclusion rules. Test, demo, duplicate and soft-deleted records excluded from quality computations.
- Threshold discipline. Binary classification reported alongside the probability threshold used (default 0.5).
- As-of timestamping. Every reported metric stamped with model version, evaluation window and as-of date.
- Method note attached. Every published quality figure carries a brief method note describing cohort, exclusions and threshold.
Interpretation for stakeholders
SAFE supports two parallel readings of its performance.
Scientific quality is measured on the qualified validation benchmark cohort - the number used for benchmarking and due diligence.
Operational quality is measured on the full population including the unreviewed backlog - the number used to manage review capacity and queue health.
Reporting both, separately and clearly labelled, supports both investor due diligence and regulator-facing governance without conflating the two questions.
Core message
SAFE's quality framework is benchmark-cohort, reproducible and method-stamped. Publicly reported model-quality figures are published as measurements, not marketing claims.
Technical methods & reproducibility metadata
Formulas, definitions & method-stamps.
The equations, definitions and minimum metadata accompany every reported SAFE quality figure.
Accuracy(TP + TN) / (TP + TN + FP + FN)
PrecisionTP / (TP + FP)
RecallTP / (TP + FN)
F12·TP / (2·TP + FP + FN)
ROC-AUC∫₀¹ TPR(FPR) d(FPR)
Confusion matrix definitions
- TP means predicted at-risk and expert-confirmed at-risk.
- TN means predicted not-at-risk and expert-confirmed not-at-risk.
- FP means predicted at-risk and expert-confirmed not-at-risk.
- FN means predicted not-at-risk and expert-confirmed at-risk.
Cohort & threshold definitions
- Benchmark cohort: qualified validation set with expert verdicts recorded.
- Operational cohort: full population including unrevised backlog.
- Threshold: probability cut-off used to convert a continuous score into a binary classification.
Core message
A reported quality number without method, metadata and as-of date is not a measurement. It is a claim. SAFE reports measurements.
Current, roadmap & assumption separation
Bold on ambition.
Disciplined on claims.
SAFE clearly separates four capability bands. This discipline protects credibility and helps stakeholders understand the technology with confidence.
01 - Current
Current capability
What SAFE can already do or has already operated.
02 - Validated
Validated capability
What has evidence from use, assessment volume, expert review, user feedback or historical operation.
03 - Roadmap
Roadmap capability
Planned improvements such as expanded dashboards, integrations, extended safeguards, sector modules and mandatory reporter continuance.
04 - Assumptions
Working assumptions
What depends on future execution, validation, sector localisation, regulations and legal or technical reviews.
Core message
SAFE can be bold on ambition and disciplined on technical claims at the same time.
Ready to protect children?
Join services across the country using SAFE
to detect what matters most.
.avif)
.avif)