How WorkCheck thinks | SpectralSense

02 / House rules

It works from your rules

Most safety cameras ship one detector for everyone. WorkCheck is personalized to you and uses your own work rules to make its judgement. Your written policy becomes a profile for each door: which roles belong at this door, what each role wears, what counts as worn versus carried, which hours and which direction. Your issued gear becomes the reference the device checks against.

It learns who belongs at each door

A door profile is not a list of objects. It describes who belongs at that door and what they look like when they are ready to work. The gate learns what normal looks like at its own door, by hour and by role, and flags what doesn't fit: a visitor with no escort, a trade in the wrong zone, street clothes at the clean-room door at two in the morning. That is a WAIT for a supervisor, not an accusation, and it never needs a name.

It can tell your vest from a knockoff

A knockoff vest and the issued vest look the same to a person and to a webcam. They do not look the same to WorkCheck. At a fixed camera under fixed light, a copy gets the cut, the trim, the print geometry, and the reflective tape slightly wrong, and the reference set sees it. For venues and events that need more, issued garments can carry a mark only the gate can see. No one wearing a copy knows it is missing.

Every verdict is recorded on the unit

Each verdict is appended to a hash-chained record on the device. Entries cannot be edited or removed without breaking the chain. The record syncs to the cloud when there is a network and stays complete on the unit when there is not. It is a record you can hand to an auditor, an insurer, or a lawyer.

The same unit works at any door

Construction gates, food plants, gowning rooms, hospital units, venues, hotels. Each one gets its own door profile with its own rules, its own reference gear, and its own record. It is the same unit at the same price.

03 / Two speeds of thinking

It thinks in two speeds

Most of what happens at a door is routine. So most of the time WorkCheck runs a fast detector on a low-power vision processor. It looks at every frame, takes a few milliseconds, and needs no network.

When it isn't sure

A second model wakes up, looks at the same frame, and explains in plain words what it sees and why it does not match the rule. A supervisor gets a WAIT with that reason attached.

It only learns from a person

The second model never teaches the first one on its own. A person confirms the call, and only that confirmation changes what the gate looks for. The unit gets better at its own door without making things up. Both models run on the unit, and the record on the unit is the only version of the truth.

Why now

Small models that can explain a picture, and an accelerator with its own memory that fits on a card the size of a credit card, both arrived in the last year. Before that, reasoning about a frame at a door meant a server in a closet or a call to the cloud, and the cloud call is the thing a privacy officer will not sign. The first units ship the fast layer. The unit is built to take the second.

04 / The verdict

It gives one of three answers

Missing gear and gear the camera could not see are two different things, and WorkCheck treats them differently. Nobody gets flagged for something the camera simply could not see.

CLEAR

Geared up

Every required item present and confident. Clears automatically, logged.

WAIT

Can't tell

An item isn't clearly visible. A person takes one look. Not a violation.

STOP

Missing

A required item is clearly absent. Stopped with a reason, on the record.

FIG.03 · VERDICTS IN THE FIELD WorkCheck reading crews across a site gate, clinic door, and event line: clear, wait, stop, cleared

Why that is hard

Seeing a vest is not the same as seeing it worn

Spotting safety gear somewhere in the picture is the easy part, and it is where most of this technology stops. Three things separate a picture that looks right from a person who is actually protected.

It has to be worn, not just there. A vest over one shoulder or a hard hat carried in a hand does not protect anyone. WorkCheck requires the item to sit on the body, on the torso for a vest and on the head for a hat, and it uses depth to confirm there is a person underneath it.

A photo is not a person. Hold a picture of a fully geared worker up to the lens and a colour-based detector will clear it. WorkCheck uses stereo depth, so a flat image is never counted as a person in the first place.

A hi-vis vest is reflective, not just yellow. Colour alone passes any bright yellow, which is how a football jersey gets through a gate. WorkCheck looks for the reflective stripes on the torso, because that is what actually makes a vest a vest.

None of this shows up in an accuracy number. It shows up the first week somebody tries to beat the gate.

05 / The reduction

What it keeps from a photo

Each check turns the picture into 160 numbers that describe its colours, and keeps only those numbers. It is like keeping a list of how much of each paint went into a painting. Nobody can rebuild the painting from that list.

Input · one reference photo

Illustrative image (AI-generated): a worker in hi-vis vest and hard hat on a factory floor

DISCARDED AFTER COMPUTE
NOT STORED
NOT SENT

ONE-WAY

→

Output · 160 numbers · 1.3 KB · no image · not reversible

Computed from the photo at left · colour info 53% · hi-vis at 60–75°, workshirt and factory blues at 195–225°

What a photo system carries

A face is biometric data: GDPR / BIPA / state privacy law surface

Retention policies, subpoenas, discovery: every photo is a future legal artifact

Breach liability: a leaked archive of worker photos is a headline

Storage: megabytes per worker per day, forever

Spoofing surface: a photo system can be fooled by a photo

Offline failure: cloud-dependent systems stop checking when the network does

What 160 numbers buy

No face, no image, no biometrics: there is nothing to leak

One-way transform: you can compute the numbers from a photo, never the photo from the numbers

1.3 KB per reference; about 60 bytes per check event (time, kit, verdict, distance, colour info)

Matching runs on-device; the check works with no network at all

The gate adds stereo depth: a flat photo is never granted personhood

An image of a person being checked is never stored. Anywhere.

06 / What it is not

What it is not

This is why it costs more than a webcam.

Not a webcam. A webcam sends pictures somewhere else to be looked at. WorkCheck looks, decides, and records on the unit at the door. Nothing leaves.

Not a security camera system. It keeps no video. There is nothing to review, nothing to subpoena, nothing to leak.

Not facial recognition. It never learns who you are. It checks what you are wearing against the rule for that door.

Not a generic PPE detector. It is trained on your gear, your uniforms, your policy, at your door, in your light.

Not a cloud subscription that goes dark when the network does. The unit works with no connection and keeps a complete record until one comes back.

Not a dashboard you read the next day. It gives a verdict in under a second, while the person is still standing at the door.

Not a camera and a laptop. It is an industrial unit with its own vision processor, a stereo depth vision module, and a tamper-evident log. That is what $1,499 buys. The $99 a month is the rules, the updates, and the record you can prove.

The policy is the model.