Ordellum AnalyticsLaunching
Evidence-based case review
Before a high-risk alert becomes a report, it is tested against the strongest evidence for and against the suspicion.
Rule-based AML produces many false positives, and each one consumes analyst time. Traditional systems also struggle to explain why a pattern is suspicious in a way an auditor can follow.
ASTRAEA — Ordellum’s forensic evidence engine — reviews high-risk alerts against three sources at once: versioned regulation for each country, the published rationales behind sanctions, and anonymised past analyst decisions. It never decides on its own; it prepares a calibrated, auditable evidence bundle for the people and models who do.
Input
High-risk alert
Prosecution lane
Supports the suspicion
- Regulation in force
- Confirmed precedent
- Sanctions bulletins
Defence lane
Challenges the suspicion
- Clean precedent
- Safe-harbour arguments
- Closest contrary case
Neutral adjudicator
Weighs relevance, authority and recency; tests every legal element
- Support a report
- Suppress (with guarantee)
- Human review
Then multi-model review and a human analyst decide.
Diagram of adversarial review. An alert splits into two lanes: the prosecution lane collects regulation, precedent and sanctions bulletins supporting suspicion; the defence lane collects clean precedent and safe-harbour arguments. Both feed a neutral adjudicator, which produces a balanced recommendation for multi-model review and a human analyst.
- Prosecution lane
- Gathers regulatory references, similar confirmed cases and sanctions bulletins that support the suspicion.
- Defence lane
- Searches for clean precedent and safe-harbour arguments that point the other way.
- Closest contrary case
- Finds the most similar case with the opposite outcome and the smallest difference that would flip the decision.
- Legal elements test
- Breaks the suspicion into the elements the law requires; if any element is missing, the engine abstains.
- Adversarial
- Evidence is gathered for and against the suspicion.
- Sufficiency-tested
- Every required element must be supported before a recommendation is made.
- Temporal
- Only the regulation in force at the time of the alert is used.
- Relational
- Wallet networks and transaction motifs inform similarity.
- Abstaining
- When evidence is insufficient, the engine says so instead of guessing.
- Evidence-weighted
- Official regulation outweighs bulletins, closed cases and analyst notes.
- Auditable
- Every recommendation carries its sources and the policy version used.
Step 1: Alert received
Only alerts on confirmed or finalised blockchain events are reviewed.
Step 2: Elements and jurisdiction
The suspicion is split into legal elements and filtered to the regulation in force where and when it occurred.
Step 3: Evidence gathered
Prosecution, defence and network evidence are collected in parallel.
Step 4: Adjudication
Evidence is ranked by relevance, authority and recency, and calibrated into a false-positive probability.
Step 5: Sufficiency check
Missing or contradictory evidence leads to a limited search or a hand-off to human review.
Step 6: Evidence bundle
A bundle with every source cited is passed to multi-model review and then to an analyst.
Critical cases are assessed independently by three AI models. A result counts only when at least two-thirds agree on the risk level; their scores are then combined with fixed weights. If the models disagree significantly, the outlier is asked again with more context, a further model is added, and persistent disagreement goes to a person. Models are referred to through an internal registry rather than by provider, so the review stays consistent when a model is replaced.
- Primary model
- 40%
- Secondary model
- 35%
- Tertiary model
- 25%
For confirmed suspicions, Analytics drafts a suspicious-activity report in five parts — summary, subject information, activity detail, AML analysis and supporting evidence — with a design target of under 60 seconds. Drafts follow jurisdiction templates, including US FinCEN, EU national FIUs, the UK NCA, Singapore STRO, the UAE and Hong Kong JFIU.
Every draft passes automated quality checks and then human review. The client or VASP files the final report.
- Fewer false positives
- ≥35%
- With zero true positives lost
- Evidence recall
- ≥0.90
- Relevant evidence found in the top ten results
- Source coverage
- 100%
- Every statement traced to its source
- Review time
- <2.5 s
- 95th-percentile design target per alert
Targets are proven through the release gates before general availability.