Anomaly Detection
How SubOps' deterministic rule engine finds settlement variances — no AI guessing, every finding backed by evidence.
Last updated
import { Cards, Card } from "fumadocs-ui/components/card";
Anomaly Detection
SubOps finds settlement variances using a deterministic rule engine — not probabilistic AI. 25+ rules cross-reference your settlement data against contractual rates, known patterns, and your own historical baselines. Every finding links back to specific source document rows.
How the rules engine works
| Component | What it does |
|---|---|
| Rule library | 25+ rules organized in families: stop-level, fuel/SCA, adjustment/clawback, and contractual-rate |
| Evidence graph | Each rule execution is a directed edge connecting a source document location → a calculation → a finding |
| Severity model | Every finding is classified as Critical, High, Medium, or Low based on dollar impact and confidence |
| Review workflow | Findings flow through a review queue — confirm, dismiss, or escalate each one |
Why deterministic matters
Key point: SubOps' rule engine is deterministic, not AI. An LLM never calculates a dollar value that goes into a finding. Standard settlement PDFs parse through structured deterministic validation first; AI is used only for unknown layouts, degraded scans, and other messy edges. The rule layer uses TypeScript math to verify data against contracts, baselines, and patterns. This separation means every dollar figure is reproducible — run the same rule on the same data, get the same result.
Severity model
| Severity | Trigger | Example |
|---|---|---|
| Critical | $500+ variance with high confidence | Misapplied fuel surcharge across 15 routes |
| High | $100–$499 variance, or pattern recurrence | Duplicate stop adjustment on 3 routes |
| Medium | $25–$99 variance | Single misclassified accessorial |
| Low | Less than $25 variance, or insufficient data | Possible rounding discrepancy |
Review workflow
- Findings appear in your queue with severity, dollar impact, and source-linked evidence
- You review each finding — the evidence pack shows the exact PDF row, the rule calculation, and the variance
- You decide — confirm the finding (adds it to your recovery ledger), dismiss it (with a reason), or escalate it (flags for deeper review)
- You export — confirmed findings become dispute-ready evidence packs for your BC
Note: The rules catalog is continuously refined as new FedEx settlement patterns emerge. Each rule has a unique identifier (R001–R025) and a documented scope, so you always know what's being checked.