For carriers · underwriting & loss-control

Frequency is falling. Severity isn't. See the jobsite behaviors that drive the next claim — before it's a claim.

An audited leading-indicator feed from real construction work — each contact-critical interaction with evidence timestamps, per-field confidence, and a documented truth state. Built to drop into your loss-control survey and underwriting file — not to score or discipline workers.

Per-field provenance: MEASURED / INFERRED / DEGRADEDEvidence span on every flagConsent-gated, PII-blurred captureOSHA Focus-Four–mapped
What this is — and isn't

A safety leading indicator, never a worker score.

The product surfaces risky technique for coaching and loss control — it does not rank, rate, or discipline individual workers. Faces and other direct identifiers are blurred before any release, and the release gate structurally blocks any media with an un-redacted face — a posture built to survive the laws an underwriter's counsel will ask about.

Coaching feed, not a worker score

Findings surface risky technique for loss control and coaching — never an individual ranking or a discipline trigger. (We mirror the explicit stance of safety-AI vendors like Protex AI.)

Consent + PII redaction by construction

Capture is consent-based under signed worker agreements. Faces, plates, badges, and names are blurred before any release, and the release gate structurally blocks media with an un-redacted face — a clip that fails the blur pass is rejected, not shipped.

Audio suppressed where law requires

On two-party-consent jurisdictions (e.g. California) audio is suppressed end-to-end and marked MISSING_SENSOR — not silently assumed present.

Ready for disparate-impact review

Findings map to objective OSHA standards, not subjective worker traits, and human review is a structural pre-condition — built with EEOC/ADA, NLRB, BIPA, and proxy/disparate-impact testing (NY DFS CL-7, Colorado SB21-169) in mind.

Leading, not lagging

EMR, TRIR, and DART tell you what already happened.

OSHA frames leading indicators as proactive, preventive measures and lagging indicators as measures of past effectiveness. Construction frequency has trended down for years while severity and the share of inexperienced workers have not — a per-shift behavioral signal is a severity-and-inexperience early warning your lagging metrics can't give you until the loss is booked.

Commercial auto learned this a decade ago: telematics turned hard-braking events — a leading behavior — into a defensible pricing input for rear-end losses. Un-tied-off edge exposure is to a fall claim what hard braking is to a collision. This feed is that bridge for the jobsite.

Findings map to the OSHA Focus Four
Falls from elevation
29 CFR 1926.501

Detected from egocentric video: elevation height cues, absence of fall-arrest system visible in frame, posture-based SIF potential.

Struck-by
29 CFR 1926.502(j) / 1926.600 (Subpart O)

Overhead object hazards, moving vehicle/equipment proximity, and projectile-risk tool use — detected from video + wrist IMU impact signature.

Caught-in / between
29 CFR 1926.651 (Subpart P) / 1926.300 (Subpart I)

Hand/limb proximity to pinch points, rotating machinery, and unguarded equipment — inferred from video hand-tracking and wrist IMU.

Electrocution
29 CFR 1926.416 (Subpart K)

Unguarded panel work, proximity to energized lines visible in frame, tool-to-conductor approach — detected from video scene context.

Honest coverage — footage cannot assess below-grade excavations, confined-space entry, structural loads, material flammability ratings, equipment weight ratings, or multi-site aggregate frequency. Missing modalities render as MISSING_SENSOR, never as assumed-present.
The bucket taxonomy

Characterized into the buckets carriers actually price on.

The unsafe-condition catalog we characterize findings into was specified by a working carrier actuary — not invented by us. 54 named conditions across 12 insurer categories, each with the visual cues that evidence it and the loss reason a carrier flags it for.

Falls7 conditions
Focus Four

Unprotected edges, missing guardrails, no fall arrest at height, unsafe ladders, uncovered holes, incomplete scaffolds, unstable surfaces.

Struck-by6 conditions
Focus Four

Suspended loads over workers, equipment swing paths, unsecured materials at height, missing hi-vis near traffic.

vehicle-proximity + blind-spot conditions: additional-sensing roadmap

Caught-in / between4 conditions
Focus Four

Unprotected trenches, spoil piles at trench edges, unguarded rotating machinery, worker between equipment and fixed object.

Electrical4 conditions
Focus Four

Exposed or damaged wiring, work near overhead lines, missing GFCI in wet conditions, energized work without lockout/tagout.

PPE7 conditions

Missing hard hat, eye/face protection, hearing protection, respirator, gloves, footwear — and PPE present but worn ineffectively.

Housekeeping / slip-trip5 conditions

Slip hazards on walkways, cords and debris in paths, blocked egress, poor material storage, uncapped rebar.

Fire / hot work4 conditions

Hot work without fire watch or extinguisher, flammables near ignition, unsecured gas cylinders, blocked extinguishers.

Hazardous materials4 conditions

Unlabeled chemicals, uncontained spills, uncontrolled silica dust.

atmospheric/radiation zones: gas-monitor integration roadmap

Confined space2 conditions

Entry without permit or attendant; no atmospheric monitoring or ventilation visible.

Heavy equipment5 conditions

No spotter near people, unsafe forklift handling, outriggers not deployed, unattended running equipment, visible faults.

Environmental / ergonomic3 conditions

Heat-stress conditions without shade or water, inadequate lighting, unsafe manual-lifting posture.

Access / barricade3 conditions

Missing barricades around hazard zones, unauthorized persons in restricted areas.

geofenced exclusion-zone proximity: additional-sensing roadmap

Honest coverage: the feed characterizes what helmet-camera video and motion sensors can evidence. Conditions requiring proximity radar, geofencing, or gas-monitor integration are marked as additional-sensing roadmap above — never silently claimed. What a modality can't see is marked MISSING_SENSOR.

Before you trust it

The pilot path: retrospective first, credit decision later.

We don't ask you to price on our word. The pilot we propose is the one your actuaries already know how to judge.

Step 1

Run on real footage, retrospectively

Consented egocentric capture from a contractor you insure (or our pilot corpus) goes through the full pipeline — no hand-picked clips, no demo reel.

Step 2

Characterize into the catalog

Findings land in the carrier-specified buckets above, each with evidence timestamps, per-field confidence, truth state, and the cited 29 CFR 1926 provision.

Step 3

Human review gates every release

A named reviewer confirms or kills each finding before it's insurer-facing. You see the release gate working — including what it rejected and why.

Step 4

You judge it against your own book

Compare the behavioral frequency we surfaced against the insured's incident and near-miss history. Your actuaries draw the conclusion — we supply the evidence package.

A false positive carries premium and blame consequences — the publish bar is deliberately high, and a pilot that surfaces fifty alerts a day where forty are noise is a failed pilot by our own definition, not just yours.

Risk matrix

Every finding plots to a 5×5 likelihood × severity cell.

The standard carrier risk matrix. Each released finding carries its cell, band, and action tier — so it maps directly onto your corrective-action workflow without translation.

Severity →
1
2
3
4
5
5
5MODERATE
10HIGH
15CRITICAL
20CRITICAL
25CRITICAL
4
4LOW
8MODERATE
12HIGH
16CRITICAL
20CRITICAL
3
3LOW
6MODERATE
9MODERATE
12HIGH
15CRITICAL
2
2LOW
4LOW
6MODERATE
8MODERATE
10HIGH
1
1LOW
2LOW
3LOW
4LOW
5MODERATE
Likelihood → (likelihood proxy, not actuarial frequency)
Likelihood axis note — likelihood proxy, not actuarial frequency. Derived from per-field model confidence; frequency firms up as the footage corpus grows across shifts and sites.
CRITICALImmediate / stop-work reviewscore 15–25
HIGHPriority correction + verify closurescore 10–14
MODERATESchedule corrective actionscore 5–9
LOWMonitor / periodic reviewscore 1–4
What a released finding delivers

Six actuarial fields, every finding, every release.

These are the fields an actuary prices on — the whole pitch in one card. The live report below the gate renders every confirmed finding in this shape; nothing ships until a person has reviewed and approved it.

format example — a released finding renders these six fields; not a specific clip
5×5 likelihood × severity · band · action tier
L4 × S5 = 20CRITICALImmediate / stop-work review
likelihood proxy from per-field model confidence; not actuarial frequency
Cited OSHA standard + compliance gap
29 CFR 1926.501Fall protection — duty to have fall protection (Subpart M)
Gap: worker observed working at elevation without fall-arrest system in use
Evidence spans + reviewer identity (who / when / stage)
161.2s – 162.5segocentric video + wrist IMU · AGREE
reviewed by loss-control inspector · <review date> · our_review
IRPM category + anti-double-count flag
Safety devices — fall protection
⚑ verify it is not already reflected in the experience-mod base rate
Per-hour frequency tag
recurring (3 occurrences) · 1.8/hr point estimate
single-clip point estimate — not a site-wide rate
Per-field truth state + confidence
MEASUREDINFERREDWORKER_CONFIRMEDPROFESSIONAL_CONFIRMEDDEGRADED
confidence: 0.84 (model); truth state promoted to PROFESSIONAL_CONFIRMED after reviewer verdict
Detect → Act → Resolve

A finding can't close without a documented corrective action.

Each released finding carries evidence spans, an action tier, the cited OSHA standard, and the reviewer's identity — so it maps straight onto your loss-control survey and corrective-action workflow.

Detect
Finding surfaced
  • Evidence span (start ms – end ms)
  • Per-field confidence
  • Truth state (INFERRED until reviewed)
  • Cited OSHA standard
Act
Action tier assigned
  • 5×5 cell → band → action tier
  • IRPM category flagged
  • Corrective action documented
  • Reviewer identity recorded
Resolve
Finding closed
  • Documented corrective action REQUIRED
  • Reviewer identity REQUIRED
  • Truth state → PROFESSIONAL_CONFIRMED
  • Finding enters the released report

A finding can't close without a documented corrective action. The Resolve node requires both a corrective action record and a reviewer identity — structural, not advisory.

Built to survive your standards

Every flag carries its source, confidence, and truth state.

Source, confidence, truth state on every field

Each flag carries the modality it came from, a per-field confidence, and a MEASURED / INFERRED / DEGRADED truth state — the explainability ASOP 23 and NY DFS expect of a model-derived input.

Maps to IRPM credit categories — without double-counting

Behavioral findings map to schedule-rating categories (Management, Employees, Equipment, Safety devices, Premises) that static property data can't supply — each flagged not to be re-counted in the experience-mod base rate.

Conservative / fail-closed discipline

Uncertain detections are routed to human review, never asserted. A false positive carries premium and blame consequences, so the bar to publish a finding is deliberately high.

Accuracy stays honest

Formal precision/recall is PENDING_GOLD until a labeled gold set exists; we never present an unsourced ROI number to an actuary.

Schedule-rating · IRPM credit categories

Behavioral evidence for the dimensions static property data can't supply.

Each finding category maps to an IRPM schedule-rating dimension. We supply the evidence package for the underwriter's own credit decision — we do not compute the credit.

Finding categoryIRPM schedule-rating dimensionAnti-double-count caveat
PPE complianceEmployees — selection, training & supervisionverify not in exp-mod base rate
Housekeeping / slip-tripPremises — condition & careverify not in exp-mod base rate
Tools & equipment practiceEquipment — maintenance & safe useverify not in exp-mod base rate
Electrical hazardsSafety devices — equipment & guardingverify not in exp-mod base rate
Falls from elevationSafety devices — fall protectionverify not in exp-mod base rate
Struck-byManagement — safety program & trade coordinationverify not in exp-mod base rate
Caught-in / betweenEquipment — machinery guardingverify not in exp-mod base rate
Ergonomics / manual handlingEmployees — training & job designverify not in exp-mod base rate

Every row above carries the anti-double-count caveat: verify it is not already reflected in the experience-mod base rate. We supply behavioral evidence; the underwriter applies the credit.

Precedent

Video coaching already links leading behaviors to lagging losses.

We cite industry precedent as context, not as our own outcome. Each figure below is the insurer's or vendor's own reported result — attributed, dated, and linked so you can read the source yourself; none is an independently audited number, and none is ours. Our own loss-reduction number stays validation-in-progress until a real peer baseline exists, and formal precision/recall remains PENDING_GOLD.

Reported >50% lower workers' comp claim frequency
Zurich North America + Arrowsight
Zurich's own figure from a 3-year video-analytics + behavior-based-safety pilot with Arrowsight across nine NYC construction sites (cumulatively >$2B in value). Insurer-reported in a joint release — not an independently audited outcome.
Source: Nov 2025 — read it ↗
Reported lower claim frequency and OSHA incidence rate
Nationwide + CompScience
CompScience's own early, self-reported results among program customers (year-over-year), republished by Nationwide's newsroom without disclosed methodology or independent validation — directional, not an audited outcome.
Source: Aug 2024 — read it ↗

External precedent only. These are published outcomes from other video-analytics programs, cited so you can verify them. They are not our measured outcome and are not a projection of this product's effect on your book.

Live loss-control report
Report status

Report under human review.

Safety observations for this worksite clip are being reviewed and face-redacted before any insurer-facing release. Findings stay hidden until a person has reviewed and approved them; the released report appears here automatically the moment review is complete.

Accuracy state: pending gold. Weighted Percent Safe and per-hour at-risk rates stay hidden until release.

Talk to our loss-control team →
Trades Data AI — Real human work, captured and audited frame by frame