AITIR 2.0 Reference Risk Scoring Model
Version: 2.0.0 Status: Synthetic teaching illustration; not a production model
Semantic rule
AITIR distinguishes:
- anomaly score;
- heuristic risk score;
- calibrated probability;
- uncertainty;
- impact;
- policy risk;
- response authority.
The example implements only a heuristic risk score. It MUST NOT be described as a probability, confidence, calibrated threat likelihood, or authorization.
Feature table
| Feature | Points | Synthetic rationale |
|---|---|---|
| failed authentication | 10 | repeated failure may merit context review |
| privilege change | 25 | entitlement changes can increase exposure |
| sensitive resource | 15 | potential impact is higher |
| after-hours activity | 10 | temporal deviation may require explanation |
| unmanaged/noncompliant device | 15 | posture can weaken assurance |
| high-risk country flag | 20 | policy-defined context flag, not nationality inference |
| recent threat-intelligence match | 25 | corroborating external evidence may raise priority |
risk_score is the sum of present feature points.
Tier mapping
0-24 -> Low
25-49 -> Medium
50+ -> High
No upper bound is implied by the formula, although the synthetic file uses values below 100.
Version 2 response mapping
Risk tier does not determine response tier automatically.
| Risk level | Default synthetic disposition | Maximum recommendation without separate authority |
|---|---|---|
| Low | observe | T0 enrichment or watch condition |
| Medium | abstain to analyst review | T0 investigation support or T1 verification proposal |
| High | urgent abstention to authorized review | T1 verification proposal; T2/T3 require independent policy and authority |
A real policy may choose a lower-impact control or no action. It must consider evidence freshness, data quality, calibration, counter-evidence, mission impact, target criticality, blast radius, holds, rollback, and authority.
Example results
The checked-in output contains 3 High, 8 Medium, and 1 Low event. The validator recomputes scores and tiers from the event flags and rejects inconsistent counts.
Why this is not calibration
Calibration asks whether stated probabilities correspond to observed frequencies on held-out data. This example has no labels and does not output probabilities. A production probability requires:
- a defined prediction target and timestamp;
- representative training data;
- a held-out calibration period;
- a later untouched test period;
- Brier score, reliability analysis, and uncertainty;
- monitoring for calibration decay and domain shift.
Selective prediction
If a model later provides a calibrated probability, Version 2 still permits abstention. Coverage is chosen from validation data with explicit error and review costs. Required reporting includes:
- accepted-set error or risk;
- abstention/review rate;
- false alerts and reviews per operational unit;
- queue capacity and aging;
- sensitivity to reviewer accuracy and cost;
- behavior under temporal and attack-family shift.
Threats to validity
- feature weights are subjective;
- binary flags discard magnitude and sequence;
- country flags can be crude, discriminatory, or operationally misleading;
- an after-hours event can be mission necessary;
- threat feeds contain false, stale, or low-context indicators;
- additive points ignore interactions and missingness;
- static thresholds drift;
- high impact is not proof of malicious intent.
Use this model only to test documentation, schema, and workflow logic.