Accuracy & confidence

WebLens pairs AI agents with deterministic checks, and we continuously validate each criterion against expert manual review. Every criterion carries a confidence tag so you always know how much to rely on it. This page explains how those tags are set.

How we score confidence

We track three measures for each criterion, all on a 0–100% scale:

• Precision — of the issues WebLens flags, the share that experts confirm as real.
• Recall — of the real issues experts find, the share WebLens catches.
• F1 — the balance of the two, which we use as the single headline confidence score behind each tag.

The confidence tags

Each criterion's F1 maps to one of three tags:

• High accuracyF1 ≥ 80%. Agrees strongly with expert review, so you can rely on its findings and spot-check as you see fit.
• Medium accuracyF1 60–79%. Performs well; treat findings as strong leads and confirm before remediating.
• Use with careA criterion that is either still building up testing data, or currently scoring below its expected range. Its findings are worth confirming before you act on them.

How we measure it

For each validated criterion, experts review a real assessment run and confirm every finding, and note any issues WebLens did not surface. From that we compute precision, recall, and F1. Coverage grows as we validate against more applications, and a criterion can move to a stronger tag between releases. You can see the tag on every criterion in the criteria reference.

What a tag is not

A confidence tag is not a measure of whether your application is good. It reflects how much to lean on the agent's call. Use it to set review priority, not to dismiss findings.

Overall performance

Across everything validated against expert review so far, the breakdown of findings is:

• 48% confirmed — real issues WebLens flagged that experts agreed with.
• 38% to confirm — additional candidates WebLens surfaced for a reviewer to accept or clear.
• 13% missed — real issues that experts caught and WebLens did not.

WebLens is tuned to surface issues comprehensively, so it rarely misses a real problem. The remaining candidates are quick to confirm or clear in a short triage pass, keeping you in control with high confidence in what ships.