Skip to content

Self-Assessment

This guide walks a detection engineering team through assessing itself against the framework, picking the right L for each of the ten dimensions, and turning the result into a prioritized roadmap.

For each of the ten dimensions, pick the lowest level for which all descriptors are true of your org today. If a row is partially true, the honest score is the level below it.

Scores aren’t averaged. Treat them as a profile, a 10-dimensional vector, not a single number. An aggregate “we’re at L2” is almost always misleading, because the lowest-scoring dimension is what bounds the program’s safety.

A scorecard with one row per dimension:

DimensionToday (L?)Target (L?)Notes / evidence
Strategy & Governance
People & Skills
Tooling & Infrastructure
Data & Knowledge
Evaluation & QA
Security, Privacy & Safety
Detection Opportunity Ideation
Detection Authoring
Detection Testing & Validation
Tuning, Coverage & Continuous Improvement

The Notes / evidence column is where the assessment is honest or dishonest. A score with no evidence is aspirational.

A radar chart with the ten dimensions on the spokes makes the gaps and balance immediately legible. Two overlays (current state and 12-month target) are usually enough.

A heatmap (dimensions by levels) is an alternative: shade the current level for each dimension and the target level in a different color.

Either visualization makes the “weakest link” problem visible, which is the point.

Score the profile honestly first, then prioritize using these heuristics:

  1. Foundations before Lifecycle. Strategy, Tooling, Evaluation, and Security/Privacy are gating dimensions. Advancing Detection Authoring or Tuning past L1 is unsafe without at least L2 in those four.
  2. Lowest score that touches production. If Evaluation & QA is at L0 and any Lifecycle dimension is past L1, that gap is your top priority. AI is shipping content with no quality control.
  3. Risk before convenience. Security/Privacy gaps trump throughput gaps. A team at L1 in Security/Privacy with engineers pasting raw telemetry into public chatbots is one mistake from an incident.
  4. Don’t skip levels. Going from L1 to L3 in any dimension without passing through L2 means skipping the governance, observability, and eval that make L3 safe.

Re-score on a fixed cadence (quarterly works for most teams) and after material events: a new AI tool adoption, a model-provider change, an AI-related incident, or a shift in regulatory context.

  • Underlying detection program maturity. This framework assumes you already have one. If you don’t, the Detection Engineering Maturity Matrix or the Threat Detection Maturity Framework are the right places to start.
  • Vendor selection. This framework is vendor-neutral. Which model, which provider, and which agent framework are downstream decisions.
  • Cost or ROI in dollars. Mentioned at L2/L3, but a quantitative model is out of scope.