Accuracy & methodology

How we measure accuracy.

We want a system that catches real problems while keeping false alarms low enough that your team isn’t dismissing noise. Here’s how we measure that, and the thresholds we hold ourselves to.

≥ 0.85
F2-score target

Weights recall over precision

< 10%
False-positive rate

Keeps false alarms manageable

85–90%
Human agreement

Inter-reviewer consistency baseline

What we measure

Recall

Of all the real problems, how many did the system catch?

Precision

Of all the things flagged, how many were truly problems?

F2-score

A harmonic mean of precision and recall that prioritizes catching problems over avoiding false alarms.

The four outcomes

Every finding lands in one of four buckets when we compare the system against a human reviewer.

Human reviewerSystemOutcome
Red flag presentFlaggedHitTrue Positive
Red flag presentNot flaggedMissFalse Negative
No red flagFlaggedFalse alarmFalse Positive
No red flagNot flaggedCorrect blankTrue Negative

The math

Recall

TP ÷ (TP + FN)

Hits ÷ (Hits + Misses)

Precision

TP ÷ (TP + FP)

Hits ÷ (Hits + False alarms)

F2-score

5 · P · R ÷ (4 · P + R)

Harmonic mean, weighted toward recall

False-positive rate

FP ÷ (FP + TN)

False alarms ÷ (False alarms + Correct blanks)

Why F2-score?

Why not just use raw accuracy?

Because red flags are rare. A system that never flags anything could look “99% accurate” while missing every real problem. F2 focuses on catching problems, which is what QC is for.

Why weight recall more than precision?

Because missing a real problem costs far more than checking a false alarm. A false alarm is 20 to 40 seconds of a reviewer’s time; a missed problem can reach the client. We tune to minimize false alarms so your team isn’t dismissing noise, but never at the expense of catching real issues.

See these numbers on your own surveys.

We calibrate to your protocol and run a batch of your real calls in the demo.