METHODOLOGY.

All data is public: raw eth_getLogs re-indexes, public explorer APIs, the public ENS ReverseRecords contract, and public web endpoints. No private endpoints, no privileged access.

Stage 0 · Census.

Every protocol's event logs were pulled raw via JSON-RPC and decoded independently: Umbra 11,714 registrations + 24,992 announcements on mainnet plus 266,339 L2 announcement logs (266,321 unique stealth addresses); Fluidkey 83,020 module logs and 117,664 decoded fund movements; Cloaked 12,951 orchestrator intent logs with 550 of 573 multi-EOA spends fully decoded. Explorer APIs were used for labels only, never for measurements.

Stage 1 · Identity anchoring.

Every Umbra registrant address (~40k ENS reverse-resolutions) was checked for a public ENS name. Result: 37.0% of mainnet registrants registered from ENS-named wallets, and the May–July 2022 crisis cohort was 51.3% named against a 37% baseline.

Stage 2 · Withdrawal graph (Umbra mainnet).

Full transaction histories for every mainnet stealth address via the public explorer API (25k+ addresses). Withdrawals were classified by destination: is it a registered key-holder? an active sender? the original sender itself?

Heuristic	Result	Statistical note
Withdrawal to registered key-holder	8,932 stealth addresses (35.7% of all)	destination is itself in the public StealthKeyRegistry
Withdrawal to active Umbra sender	41.5% of all withdrawals	recipient is a visible public actor
Self-payment loops	391 users paid themselves	174 from ENS-named wallets
Same-block sweeps	1,102 stealth pairs	~20 expected by chance → 55× lift over null
Destination clustering (union)	48.3% of active stealth addresses	largest cluster: 131 addresses
ENS-named registrants	4,286 / 11,578 (37.0%)	reverse-resolved via public ReverseRecords contract
ENS-named payments	21.3% of all payments	naming propagation from registrant set

Stage 3 · Clustering with null calibration.

Shared withdrawal destinations plus same-block sweeps were used to cluster stealth addresses into single-owner groups. Every heuristic was run against shuffled controls: 265 same-block withdrawal pairs observed against ~5 expected by chance (55× lift). Union result: 48.3% of active mainnet stealth addresses fall into clusters; the largest spans 131 addresses.

Stage 4 · Amount fingerprints.

Fingerprint	Finding	Null model
Exact wei repeats	455 distinct values across 1,617 payments (6.9% of ETH payments)	collision probability ≈ 2⁻⁶⁰; every repeat is a proven relationship
Fiat price-point families	108 families	e.g. 70 payments in 0.00445–0.00455 ETH over 13 months = one merchant's revenue
Fluidkey dust suffixes	.552850 family: 12 Safes	balances decrement in exact 9-USDC steps

Stage 5 · Fluidkey Safe welds.

Each edge type was computed independently, with protocol-internal counterparties (earn vaults, share burns) excluded: naively counted, those fake-link 91% of users. CEX hot wallets were identified by extreme nonce and fan-out (a 478,873-nonce wallet funding thousands of strangers proves nothing; a quiet 74-nonce wallet that seeded 836 Safes proves everything).

Weld type	Safes linked	Share	Note
Shared cash-out destination	6,293	23.3% of active	decoded from 117,664 fund movements out of 27,013 active Safes
Shared personal-wallet funder	5,026	—	CEX hot wallets excluded by nonce/fan-out heuristic
Born together (atomic co-init)	3,335	10.7% of census	1,224 transactions initialized multiple Safes atomically
Union of all three	12,697	44.3% of observable	edge types overlap; union < sum

Stage 6 · Cloaked merge decoding.

573 multi-EOA spend transactions; 550 fully decoded from calldata, 100% coherent single-user consolidations. 40.1% of Base spenders and 61.1% of Ethereum spenders welded their accounts in the act of spending; one user merged 51 one-time addresses in a single transaction.

Stage 7 · Username enumeration (Cloaked).

Channel	What leaks	Scale
CCIP gateway oracle	name existence (address vs zero)	dictionary probeable
username.clkd.id web app	name existence (200 vs 404)	faster oracle
Google index	usernames with no probing at all	site:clkd.id
Result	69 registered usernames recovered	from ~9,000 probes; 0 of 105 CT influencer handles registered
Seed harvesting	1,475 live receive addresses	across 59 known usernames; weekly re-check watchlist

Stage 8 · Cross-protocol joins.

Every address set was intersected across protocols and across chains. The address layer is perfectly firewalled (zero overlaps); the wallet layer is not.

Join	Result
Umbra ∩ Fluidkey wallets	8 addresses (3 ENS-named: ckxpress.eth, leo888.eth, zyrav21.eth)
Cloaked EOAs merged on Base AND Ethereum	115
Cross-protocol address reuse	0 across every address set held

Stage 9 · Negative results.

Null results are published as credibility for the positive ones:

Checked	Result
Ephemeral-key reuse (Umbra)	3 cases in 25,000 payments
Stealth-to-stealth payment chains	≈ 0
Cryptographic-layer failures	none in any system

Stage 10 · L2 withdrawal graphs (complete).

The mainnet withdrawal analysis was extended to all four L2s. Method: (1) nonce prefilter via batched eth_getTransactionCount on dRPC, dropping addresses that never sent a transaction; (2) full transaction history pull for the survivors via Etherscan V2. A pure-RPC token-sweep scan was evaluated and rejected: dRPC kills selective log scans beyond ~10k-block ranges, making it slower than the pull it would replace.

Chain	Censused payments	Unique stealth addrs	Nonce = 0 (dropped)	Nonce > 0 (pull list)	Withdrawal txs pulled
Base	5,604	5,604	1,637 (29.2%)	3,967	done
Optimism	58,048	58,048	4,199 (7.2%)	53,849	done
Arbitrum	100,138	100,138	11,587 (11.6%)	88,551	done
Polygon	102,549	102,531	48,074 (46.9%)	54,457	done
Total	266,339	266,321	65,497 dropped (24.6%)	200,824	all 4 chains

The welding heuristics (shared destinations, same-block sweeps, amount families) ran over all four L2 graphs: 73.3% of active L2 receivers are linkable (Arbitrum 76.0%, Base 77.4%, Optimism 69.8%, Polygon 72.2%), replacing the 2023 paper's 25.8–65.7% estimate with measured values.