Fraud investigation · evidence-grade
A customer was scammed.
Where did the money go?
Give The Judge the address the value left from and the transaction it left in. It follows that value forward — hop by hop, never sideways — names the exchange, mixer or bridge it reached, and produces the freeze request a desk can actually send. Every figure carries where it came from, because a finding that cannot be checked cannot be filed.
What a desk does with it
Where did the money go?
The victim's address and the theft transaction in; the trail out — every hop, where it split, and where it stopped. A trail that ends at an exchange is still recoverable; one that ends at a mixer is not.
Trace fundsIs this customer safe to onboard?
A verdict band with the signals that produced it, each carrying the exact transaction that fired it — and a validation block that says which rules were measured against real addresses and which were only evaluated on synthetic data.
Screen a walletThe funds reached an exchange — now file it.
A terminal that resolves to a named institution produces a freeze request with the institution, jurisdiction, amount and transaction hashes attached: the artefact a desk sends, not a paragraph about one.
Build a freeze requestShare the finding with another desk.
Publish a signed, content-addressed batch any desk can verify without trusting this one — Merkle root, signature, per-finding proofs. A second desk screens against it on arrival.
How sharing worksMeasured, not claimed
- 9 of 10
- rules measured against real addresses
- 815
- addresses in the tuning corpus
- 297k
- venue-published addresses
The validation block in every response splits fired rules into those measured against real addresses and those evaluated only on synthetic data — so no one can over-trust a result.
Precision 0.818, recall 0.818, false-positive rate 0.063 — those weights were fitted while viewing that corpus, and the docs say so rather than presenting the figure as production accuracy.
Institutions' own proof-of-reserves publications, attributed with the source URL and retrieval date, scoped to the ledger the claim is about — the difference between a name and an assertion.
How it holds itself to that
Forward only, never sideways
Value moves from one address to another, so the trail follows it that way. An undirected hop search would trace the victim's own history and report a path that predates the theft.
An unknown stays unknown
An unlabelled address is never assumed benign and never guessed at. A missing classification is reported as an absence of data, because a false 'this is fine' is worse than a missing one.
Provenance on every figure
A label carries the publisher's URL and the date it was read. A claim the venue scoped to one ledger is never applied to another. A truncated read says so on the face of the result.
Open a case
The victim's address and the theft transaction are enough to start. Nothing is stored, nothing is shared, and the trail you get is the same trail you would get again tomorrow from the same inputs.