Webscan Documentation

Lesson 17 · Prioritisation & the verdict model

Phase 4 · Auditing at scale — Lesson 17 of 21 · All lessons

Objective: understand how webscan decides which findings matter most, how your verdicts shape that order, and how the optional verdict model learns from your decisions.

Before you start: some findings you've given verdicts to (Lesson 6). This lesson explains machinery you've already been benefiting from.

The problem prioritisation solves

A big scan can return hundreds of findings. You can't fix everything at once, and raw severity alone isn't enough — some "critical" items are false positives, some were already accepted. You need the list ordered so the findings that actually matter show up first.

Prioritisation — the ordering webscan gives your findings so the most important, still-actionable ones rise to the top. That's the priority indicator you saw on the scan results page (Lesson 4).

The important promise: it's deterministic and transparent. The score comes from rules you can understand, not a black box — so you can always see why something is ranked where it is.

How the base score works

Every finding starts with a score built from two clear signals:

• Impact (severity): Critical = 100, High = 70, Medium = 40, Low = 20.
• A bonus for WCAG Level A: the most fundamental barriers (Level A criteria) get a small boost over AA-only ones, because they block the most basic access.

That base score sorts findings by how serious they inherently are.

How your verdicts change the order

This is the key idea: your human verdict (Lesson 6) is authoritative and reshapes the priority. Once you've judged a finding, webscan respects that judgement everywhere it appears:

This is why recording verdicts pays off: it doesn't just label one finding, it cleans up the whole list — confirmed work rises, noise sinks. Findings end up with a priority label (critical / high / medium / low / accepted / suppressed).

The verdict model: learning from your decisions

Setting verdicts one by one is powerful, but webscan can also learn from them.

Verdict model — an optional, per-organisation classifier (a small machine-learning model) trained on the verdicts your team has recorded, so it can predict a likely verdict for a brand-new, not-yet-judged finding.

Two rules keep it honest:

• It's a ranking hint, not a decision. When a finding has no human verdict, a confident model prediction gently nudges its priority (a likely false positive sinks a bit and is labelled likely-fp; a likely confirmed rises a bit), and the prediction is shown as a hint. It never fully suppresses a finding — only your human false-positive verdict does that.
• You stay in control. The model informs the order; it never hides anything or overrides you.

This is the same engine behind the Suggest button from Lesson 6.

The Prioritisation model dashboard

Open your organisation's Prioritisation model page to manage it:

• Your labelled data — how many verdicts you've recorded and their breakdown. This is the training data; more and better-balanced labels make a better model.
• Current model — its status: Training, Ready, or Failed (with the last error if training didn't succeed).
• Train now — kicks off training in the background from your labelled data. It needs enough examples across more than one verdict class to succeed. (Training is an admin/superuser action.)
• Export dataset (JSON) — download your labelled findings as a dataset, for your own analysis or record-keeping.

Tips & pitfalls

• Prioritisation improves the moment you set verdicts — even with no trained model at all. The deterministic rules plus your verdicts already do the heavy lifting.
• Don't chase a model too early. With few labels it has little to learn from. Build up honest verdicts first; train when you have a solid, varied set.
• Set impact on custom rules thoughtfully (Lesson 16) — it feeds this same score.
• Trust the order, but skim the tail. Suppressed ≠ deleted; occasionally a "false positive" verdict was a mistake, and the finding is still there to revisit.

Recap

• Prioritisation is a transparent, deterministic ordering from impact + a WCAG Level A bonus.
• Your verdicts reshape it: false positives suppressed, won't-fix downweighted, confirmed boosted — cleaning the whole list.
• The optional verdict model learns from your verdicts to predict verdicts as a ranking hint only — it never suppresses or overrides you.
• Manage it on the Prioritisation model page: labelled data, status, Train now, Export dataset.

Try it yourself

Open the Prioritisation model page and look at Your labelled data. Mark a few more findings as False positive and Confirmed barrier, then reopen the scan results and watch the order change — confirmed items up, false positives gone from the top. If you have plenty of varied labels and admin access, try Train now.

Previous: ← Lesson 16 · Custom rules Next: Lesson 18 · Monitoring →