How the tracking works

Transparent chain data: balances, amounts, heights, counterparties. Arithmetically exact and independently checkable against any node.

Shielded activity. An upper bound on a candidate set, never a value and never an owner. A bound that shrinks is a bound that bought its reduction with an assumption, and the assumption is printed with it.

Undefined by construction: not derivable from public data at all. Printed as a refusal rather than left blank, because a blank cell reads as an oversight and a refusal reads as a finding.

1.1 balance and history — arithmetic, not inference

bal(a) = sum over u in UTXO(a) of v(u). History is every transaction with an input or output script paying a; counterparties are the other scripts in those transactions. Fees are sum in - sum out - valueBalance terms, and where a transaction has shielded components the fee is split across pools, so the page shows the ZIP 317 conventional fee and the implied logical-action count L = fee / 5000.

1.2 common-input ownership — union-find over input scripts

All transparent inputs of one transaction are spent by one key-holder, so union-find over the input scripts yields clusters C(a). Two exceptions are flagged rather than assumed away: coinjoin-like shapes, rare on Zcash, and P2SH multisig.

t3ev37Q2uL1sfTsiJQJiWJoFzQpDhmnUwYo - ZIP 271 lockbox disbursement multisig

1.3 change detection — soft membership, weighted

In a two-output transparent transaction where one output pays a never-seen address of the same script type and the other is a round amount or a reused address, the fresh one is change with probability p_change. Change outputs extend the cluster with weight p_change.

1.4 exchange-withdrawal shape — behavioural, and only behavioural

One large input paying out once with the change returning to the same address, and many-to-one deposit sweeps, are behavioural exchange signatures. The corpus records one on 2025-12-24: “Withdrawal from the hot wallet to a freshly created address. Input 120,552.69 ZEC in, 29,999.99 out, 90,552.70 change back to itself: textbook exchange-withdrawal mechanics.”

t1PKBiv7mtzD9bNafYaqyxaENeiNDbpKxxQ - Exchange hot wallet, labelled "Binance" by Lookonchain

the audit record, rendered — illustrative shape, not a real transaction

why the chain is generated, not narrated — the contract behind the picture

1. 01The anchor published with a spend fixes a tree position, and nothing above it can be the spent note. This is the only universally correct filter.
2. 02A time window is a behavioural guess. It is anchored at the height the note was created and it is logged as an assumption, with its parameters; the audit record's time_window variant carries windowBlocks and a half-open height range where the lower bound is exclusive and the upper bound is the anchor's own creation height.
3. 03An amount match intersects two intervals with an explicit fee tolerance. Its audit record carries the matched deposit, both amounts in zatoshi, the tolerance used, and whether the match was exact or fee-tolerant - so a reader can reconstruct the step rather than take it.
4. 04Every estimator is a pure function and emits {filter, params, countIn, countOut} with both counts as bigint, so the chain above is generated from the run rather than narrated after it. rawCount and effectiveSetSize are equal when no filter applied, and diverge by exactly what the filters took.

the posterior — four likelihoods, one distribution, one number

w_j  proportional to  L_amount(Y | X_j) . L_time(h - h_j) . L_fp(T, T_j) . L_struct(T_j)
p_j  =  w_j / sum of w
H    =  - sum over j of  p_j log2 p_j        N_eff = 2^H        claim = level(N_eff)

L_amount

1 for an exact match; exp(-(|X_j - Y| / X_j) / eps) for a relative one eps = 10^-4

L_time

the half-life kernel exp(-ln2 . age / tau) tau = 2 days

L_fp

1.0 when the wallet fingerprints agree, 0.5 when they disagree tunable, and it is a soft signal

L_struct

down-weights candidates already consumed by a HIGH link, one-to-one, greedy by weight structural, not behavioural

Every transaction page prints the same five things beside the number: the candidate count before and after each filter, the top three candidates with their p_j, N_eff, the claim chip, and the assumption sentences - W, tau, eps and the fingerprint-table version. A posterior published without them would be a number with no way to disagree with it.

flow estimate over hops — what the drawing on /flows is a drawing of

From a transparent address, follow value through k <= 3 boundary crossings, multiplying p_j along the paths. Edges are drawn with opacity proportional to weight and cut at p < 0.02, and the mass unresolved in the pool is shown as a first-class residual bar rather than as leftover - because it is the honest answer most of the time.

what this is not — the closing refusal, and the reason for the rest

None of it identifies a person. It produces bounded, reproducible estimates from public data, with every assumption visible and every claim capped by the claim level. Three hops of a multiplied probability is still a probability; it never becomes a fact by being followed further.