Many commercial classifiers reduce the problem to a binary result: CSAM or not. Very useful for screening and for trust and safety at, say, a file hosting provider, but a thin description of evidence for anyone conducting a criminal investigation.

Rigr treats apparent age and severity as separate signals. VST shows an age × severity distribution, with uncertainty and contextual findings beside it. That supports review order, collection analysis and mapping to local categories.

VST’s models are fast enough to process extracted media across criminal casework at intake. Investigators have repeatedly told us that they encounter CSAM in cases initiated for entirely different reasons — a surprising number have encountered it during terrorism investigations. Limiting screening to specialist referrals makes discovery depend partly on how the case was first labelled.

What severity grading is for

Severity primarily orders review: it moves material most likely to require urgent safeguarding or evidential attention towards the front of the queue and reduces avoidable exposure. That last point is not a side benefit. Investigators in this field carry a cumulative exposure burden with well-documented consequences, and material a machine can set aside is material nobody has to look at.

At collection level, apparent age, severity, volume, recency, repetition and indications of production can shape interview preparation, safeguarding and risk assessment. A positive-file count cannot do that.

Age × severity, not one opaque label

Rigr’s media classifier returns a frame-level severity score from 0 to 5. Its age-estimation model returns an apparent-age distribution. Keeping the axes separate shows what drove the result and routes uncertainty near an age threshold for review.

Level 1 is explicit material in which the model detects no apparent minor. It is operationally distinct from level 0 and the higher severity levels, but is not a finding that the material is lawful.

Context remains separate. Self-generated material, sadomasochistic content and computer-generated elements may change investigative or safeguarding decisions without changing the severity of the depicted act.

There is no universal severity scale. England and Wales has used the ten-level COPINE research typology, the five-level sentencing framework set out in R v Oliver, Hartrey and Baldwin, and the current A/B/C guideline. These schemes answer different questions. Other jurisdictions answer them differently again, and the frameworks used for hash exchange and shared tooling — Project VIC and CAID among them — carry their own vocabularies, because they were built to move data between agencies rather than to sentence anyone.

A classifier should therefore keep three records separate:

  1. the model’s score and uncertainty;
  2. the agency’s versioned mapping to its own categories;
  3. the reviewer’s decision.

In VST Teams, the jurisdictional mapping is configuration rather than part of the classifier. A change in local categories does not require the model to be rebuilt.

A collection is not an average

Severity is ordinal; an average across a collection has no clear meaning. Nor does the highest-scoring item summarise the case. Repeated and distinct files, newly produced and widely circulated material, and the distribution and recency of the collection remain separate facts.

The machine proposes, reviewer disposes

The system should preserve the raw model output, uncertainty, model version, jurisdictional mapping and reviewer decision as separate, attributable records. An override should not erase the proposal that preceded it.

Exact hash matches, robust or perceptual matches, classifier predictions and human decisions should remain visibly distinct. They make different claims and have different error modes. The classification review tour shows where VST presents a model result and where the reviewer changes it. A basic principle across our tools is that we will never attempt to automate police work, but rather to mechanise it: huge backlogs do not serve the public interest, we need to go faster, but not at the expense of removing human judgement.

What to ask when evaluating a system

  1. Does it return only CSAM or not, or separate apparent age, severity and uncertainty?
  2. Is it fast enough to run across criminal casework, not only specialist referrals?
  3. Can jurisdictional mappings be configured and versioned without retraining?
  4. Are contextual findings kept separate from severity?
  5. Can the audit trail distinguish matches, predictions and human decisions?

A useful system does more than flag candidate CSAM. It should create a reviewable age × severity picture early enough to find material outside specialist cases.

If you are working on a digital evidence backlog, we wrote about the wider problem here. Questions or corrections: [email protected].