Transparent by design

Accuracy begins with refusing to pretend.

Sentinel does not claim to predict violence, diagnose mental health, infer ideology or identify a “type” of dangerous employee. It measures observable workplace conditions and convergence of independent signals, then supports trained human review.

Construct 01

Protective climate

Speak-up confidence, respectful conflict, safety adherence, support accessibility, pressure management and fair-review confidence.

Construct 02

Concerning-change climate

Time-bounded reports of unusual behavioural change, escalating conflict, threatening language, procedural bypass and multiple independent concerns.

Construct 03

Signal convergence

Confidence rises when different channels independently point toward the same team, time window or concrete situation. Correlated duplicates should not be counted as independent evidence.

Construct 04

Human disposition

Support, management intervention, operational investigation, security review or emergency action are distinct pathways. Sentinel should never collapse them into one generic “risk” label.

Scoring logic

Two outputs. Never one accusatory score.

A

Protective Climate 0–100

Higher indicates a stronger climate for speaking up, respectful conflict, procedure adherence and access to support.

B

Attention Level

Routine → Review → Priority Review. This is a routing signal, not an assessment of guilt, intent or diagnosis.

C

Confidence

Confidence is based on independence, recency, specificity and convergence — not on how dramatic a single allegation sounds.

Validation standard

We earn accuracy empirically.

The prototype scoring is a product hypothesis. Production claims should only follow staged validation with real organisations and independent expert review.

Content validity

Industrial-organisational psychology, behavioural threat assessment, HR, security, legal/privacy and employee representatives review every construct and item.

Reliability

Evaluate internal consistency, test–retest stability where appropriate, item discrimination and response-quality effects.

Factor structure

Exploratory and confirmatory factor analysis should test whether the proposed seven dimensions actually emerge in data.

Cross-cultural invariance

Before comparing sites or countries, test whether constructs behave comparably across language, geography, sector and workforce groups.

Calibration, not rare-event prediction

Validate against review outcomes, confirmed policy/safety events, resolution quality and timeliness — not a simplistic promise to predict rare acts of violence.

False-positive governance

Track how often signals lead to no substantiated concern, whether certain groups are disproportionately surfaced, and whether reviewers override the system.

Prospective validation

Freeze a scoring version, preregister evaluation criteria, test prospectively, document performance and only then revise thresholds.

Human oversight

High-consequence action requires documented human judgment. Model output is evidence for review, never the decision itself.

Evidence alignment

Grounded in established principles.

CISA insider-threat guidance

CISA emphasises detecting and reporting concerning behaviours, multidisciplinary threat-management capability, contextual assessment and respectful interventions that consider dignity, rights and privacy.

View CISA guidance ↗

FBI behavioural analysis

The FBI states that no single behaviour means someone is on a path to targeted violence; multiple concerning behaviours may indicate reason for concern.

View FBI guidance ↗

NIST AI Risk Management Framework

NIST emphasises governance, measurement, human oversight and evaluation of AI risks, including the need to retain context when modelling complex human phenomena.

View NIST AI RMF ↗