DRAPE stands for Detection Reliability And Precision Efficiency which is basically a unified index (score) that quantifies:
- Reliability: Are the detections actually finding true attacks? (TP)
- Precision Efficiency: Are they doing so with minimal noise? (FP)
The algorithm outputs a simple heuristic metric to score detection performance using TP/FP counts. The index score can be either negative or positive with the extremes implying the following characteristics:
- Negative = noisy, unreliable detection.
- Positive = precise, efficient, valuable detection.
The index is a composite measure of the individual TP/FP data points which can be used to measure detection/rule performance over time.
It reveals whether a rule is signal-dominant (+) or noise-dominant (–) by inspecting the number of its TP and FPs alerts over a period of time.
The formula rewards rules that spots a decent amount of threats (TP) with good precision, while penalizing noisy rules, especially those that haven't yet "proven" themselves with enough TPs.
The formula captures the real-world tradeoff between TP and FP figures better than precision or simple raw counts. Most single-number metrics are either:
- too forgiving of noise (TP-heavy rules look great even when they're a FP factory), or
- too punishing (precision destroys high-value, high-volume rules).
DRAPE models that tradeoff how detection engineers think in practice:
- High TP → good
- High precision → better
- Extremely noisy rules → bad
- Slight noise on high-value rules → acceptable
This mirrors actual triage experience so I encourage you to try it out.
| Input | Meaning |
|---|---|
TP |
True positives: alerts confirmed as real threats. |
FP |
False positives: alerts confirmed as not a threat. |
IC |
Inconclusive alerts (optional, new in v1.1): alerts that got no clear verdict, such as a benign or neutral outcome, or alerts closed with no disposition. |
| Weight | Default | Meaning |
|---|---|---|
k |
0.20 | FP penalty: how much false positives pull the score down. |
w |
0.38 | Precision reward: how much a rule gains for being right most of the time. |
a |
0.5 | Inconclusive alert cost (new in v1.1): each inconclusive alert counts as this share of one false positive. a=0 ignores them, as v1.0 does. |
The formula, with EFP = FP + a × IC (effective false positives):
precision = TP / (TP + EFP)
score = 10 × [ log10(TP + 1) × (1 + w × precision) − k × log10(EFP + 1) / (log10(TP + 1) + 1) ]
The score is rounded to 2 decimals. When there is nothing to score (TP + EFP = 0), it returns an empty value.
A new rule that raises 100 alerts in its first week, none of them with a clear verdict, still costs analysts time without proving anything. In v1.0 it had no score at all, so it sat with the rules that never fired. In v1.1 it scores -3.42, among the rules that need attention: worse than 10 false positives (-2.08), but not as bad as 100 confirmed false positives (-4.01). Rules that catch threats barely move.
Without inconclusive alerts (IC=0, the default), v1.1 returns exactly the v1.0 score.
- Python: drapeindex.py
- Splunk (SPL): drapeindex.spl
See CHANGELOG.md.
Intro blog post on Detect.FYI website: https://detect.fyi/introducing-the-drape-index-how-to-measure-in-success-in-a-threat-detection-practice-154fd977f731

