Privacy Model

Mechanisms: session isolation, randomized settlement windows, probabilistic batching, and slicing.

Mechanisms

Ephemeral session ids to decouple UI context from wallet addresses.
Randomized settlement windows (uniform or truncated normal) to blur timing.
Probabilistic batching with bin width to increase crowding.
Intent slicing via Dirichlet weights within minSize constraints.

Profiles

Fast: minimal jitter; near‑instant settlement.
Balanced: moderate jitter and batching.
Stealth: wide windows, slicing, optional mixouts.

Note: On‑chain settlement is public — goal is to reduce trivial correlation, not guarantee anonymity.

Mathematical Model

We minimize expected linkage score \(L\) for an intent given privacy parameters \(p=(w, b, s)\) where \(w\) is jitter window, \(b\) is batch bin width, and \(s\) is slicing factor:

L(p) = α·P(timing|w) + β·P(size|s) + γ·P(counterparty|b)

Coefficients \(α,β,γ\) are calibrated from historical inference attacks. The coordinator samples \(w\) from priors per profile and enforces \(b\) during inclusion; slicing uses a Dirichlet distribution with concentration parameter \(s\).

Parameter Ranges

Profile	Jitter Window	Batch Bin	Slices
Fast	0–2s	—	1
Balanced	2–8s	250–500ms	2–3
Stealth	8–30s	500–900ms	3–6

Implementation Pseudocode

function samplePrivacy(profile):
  if profile == 'fast':  w = U(0,2);  b = 0;       k = 1
  if profile == 'bal':   w = U(2,8);  b = U(250,500)ms; k = 2..3
  if profile == 'stealth':
      w = U(8,30); b = U(500,900)ms; k = 3..6
  weights = Dirichlet(k, α = 1.0)
  return { window:w, bin:b, slices:k, weights }