VSTTeams Edition · Classification

Media Classification · interactive walkthrough

Tens of thousands of files. An interview at nine. Know what you're holding.

A seized device lands on an analyst's desk with more material than anyone can watch, and decisions riding on it — an interview in the morning, a risk assessment that bears on whether a child is safe. VST answers the two questions that decide a case: how old is the subject, and what is happening. It reads every file, estimates an age for every face as a range, grades the act, and leaves every call to the analyst.

Begin the tour → Canned click-through of the real screens. Nothing here is processed.
What the tour shows
Summary

Know what's on the device before you open a file

Before triaging a single item, the folder summary answers the question the case turns on: what is in here, and how much of it could we assess? A likely-minors count sits next to its denominator — the coverage is stated, never hidden — and an age × severity grid shows where young subjects and severe material actually meet. Every figure is ‘likely’: an automated read, never a determination of age.

It counts items, not faces — one image with three children is one item, filed under its youngest subject. The grid crosses the model's life-stage read against its severity, so the eye lands on the cells that matter before anything is opened.

Triage

Start with the worst, not the first

A single seized device can hold tens of thousands of images and videos — far too many to open one by one. VST classifies every file and sorts by severity, so the analyst starts where it matters and works down. Everything is blurred by default.

A red KNOWN band is a hash hit against a known-material list; an amber CONFLICT means two analysts disagreed. Both jump straight to the top.

Filter

Cut the device down to what needs a person

Most of a seized device is ordinary. A saved preset takes the whole collection down to the part that needs judgement — in the screen above, 420 items to 40 — and the clear-cut remainder is confirmed in bulk from the keyboard. An analyst's time goes where it is actually needed.

The certainty sliders isolate a band — say 40–80% — so you work only the grey zone the model is least sure about.

Age

How old is the subject?

Open an item and VST estimates an age for every face it finds, each reported as a range rather than a single number — 12.6 ± 1.9 years, not ‘a child’. The panel counts how many subjects fall under eighteen and names the youngest. Where a band straddles the line it says so, instead of picking a side.

Notice that the picture beside it is still blurred. An analyst can establish that a collection contains children without having to look at it.

Decide

How young, and what is happening

Severity weighs two judgements together: the subject's apparent age, and the specific act — graded from posing to penetrative — alongside aggravating specifics captured as flags: nudity, a weapon, violence, animals. Where a face is hidden, body evidence produces no age of its own; it bears on how the material is graded, against the thresholds your jurisdiction sets. The model's read is labelled as the model's read, with a Why? to open.

Nothing is auto-actioned. The analyst confirms, re-grades the 0–5, or rejects — from the keyboard, with Space to reveal only when they choose to look.

Known material

Known material is a fact, not a guess

A hash hit against a known-material list is shown as exactly that, with its provenance — which list, which record. For video, the timeline surveys every frame and jumps to the worst one, so hours of footage don't mean hours of watching.

The hash match sits separate from the model's opinion — a definite ‘known material’ is never blurred into a probabilistic ‘the model thinks’.

vst · Review · classification queue
⤢ ExpandKnow what's on the device before you open a fileStart with the worst, not the firstCut the device down to what needs a personHow old is the subject?How young, and what is happeningKnown material is a fact, not a guess

1 / 4
Use ← → to move · click a phase to jump

What it's for

An early read, before interview
Walking into an interview already knowing the type and severity of what was found brings clarity to the room — sometimes the difference between a silent or combative interviewee and a cooperative one. We can't promise cooperation. We can promise an early, rapid survey of the material.
Evidence for risk & safeguarding
Many offenders never face a custodial sentence — which makes understanding the risk they pose all the more urgent. The shape of a collection — say, a focus on young boys — is exactly what a risk assessor or social worker needs to know. VST surfaces that shape, not just a count.
Beyond “CSAM or not”
VST reads the specific act — posing → exploitative → non-penetrative → penetrative — plus aggravating flags like a weapon, violence or animals, not a single yes/no. Legal thresholds differ country to country, so your team sets where the line falls for its own jurisdiction.
Age, as a range
The hardest call — is this a child? — is where we're strongest. VST estimates an age for every face it finds and reports it as a band, not a number, saying so plainly when that band straddles eighteen. Where a face is hidden, body evidence produces no age of its own; it bears on how the material is graded. It is the same age engine you can try live on our age-estimation demo.

What makes classification at scale safe

0 – 5
A severity scale the model proposes and the analyst confirms or overrides — never applied silently.
Blur + reveal
Every item is blurred (and can be greyscaled) by default; you hold to reveal, protecting analyst welfare.
Why?
Every classification is explainable — the model’s opinion is labelled as an opinion, with its reasoning one click away.
Hash match
Known-material hits carry provenance — which list, which record — shown separately from the model’s guess.
Keyboard
J/K to move, 0–5 to set severity, Enter to confirm — a trained analyst never leaves the keys.
Human-decided
Nothing is auto-actioned. An analyst confirms every item before it counts.