​
The claim being made

Clustering asserts that two addresses are the same operator. That is a serious claim about a named party, so the design is deliberately conservative: it would rather return nothing than return a plausible link it cannot evidence.

​
No single behavioural heuristic can cluster alone

Seven heuristics vote against an acceptance threshold of 0.75. Only two are weighted above it: The split is exactly the observe/infer line. An inferred pattern — similar timing, overlapping counterparties, comparable gas behaviour — needs corroboration before it can assert identity. An observed on-chain relationship does not.

​
Every edge carries its provenance

Accepted edges are returned with the evidence that produced them and marked status: 'inferred' — never 'asserted'. The distinction survives into the payload so a downstream consumer cannot promote an inference to a fact by accident.

​
Zero clusters is normal

A low score means “no evidence of a link”, not “no link”.A live depth-2 run over 54 addresses clustered nothing, with a strongest pair of 0.369. That is the system working correctly, not a failure.Reading an empty cluster result as “these wallets are definitely unrelated” inverts the claim being made.

​
Skipped is not no-match

On Hyperliquid, the deployer and gas_fingerprint heuristics skip every pair, because the venue exposes no contract-creation events and no transaction-metadata class to fingerprint. The response distinguishes skipped-with-reason from evaluated-and-no-match. Collapsing them would report “we checked eight ways and found nothing” when the honest statement is “we checked six ways, and two were inapplicable here.”

​
Depth costs runtime

depth accepts 0–2 hops of counterparty expansion. Each hop multiplies runtime; depth 1+ can take tens of seconds by design. That is search breadth, not slowness.

​
Read-only, enforced

Clustering is invoked with persist=False and a test asserts it. Inferred edges are never written back into the entity graph — an edge written on a language model’s say-so would be a write reachable from a prompt.

​
Availability

The backing wallet_clustering module ships on an unmerged branch. On a master checkout, get_wallet_cluster returns module_absent with a reason. Do not repoint it at a different module to make it answer.
Scienter publishes impersonal market signals of general and regular circulation. Nothing here is investment, legal, or tax advice, or a recommendation to buy, sell, or hold any asset. Scienter is not a registered investment adviser or broker-dealer. Trading digital assets can result in the total loss of your funds — see Disclaimers for the full text.