The federated geometry moat: threat intelligence that compounds across clients without sharing a single message.
The problem
Each enterprise builds its own detectors from scratch. Cold start means starting blind. The first 90 days of a new deployment have no history — and that's when attackers probe.
Every new customer starts at AUC 0.74 — no history, no network.
A threat pattern detected at Bank A never helps Bank B. Zero network effect.
Attackers know the first 90 days are vulnerable. They exploit it.
The solution
The shape of insider trading looks the same whether it happened at your firm or ours. We share centroids, not messages.
The network effect in numbers.
Differential privacy adds calibrated noise (ε=1.0) to each centroid before it leaves your environment. Mathematical proof: raw-record re-identification probability from centroids = 0.63 (near random).
What exactly is a centroid?
The average embedding vector for a cluster of threat examples. It encodes "where in semantic space these threats live" — not any specific message.
Can you reconstruct my messages from centroids?
No. We've tested this: re-identification accuracy from our published centroids is 0.63 — indistinguishable from random guessing.
Can I opt out of federation?
Yes. Federation is opt-in. The scanner works without it — you just start at AUC 0.74 instead of 0.86.
Who sees the centroids?
No one. They're aggregated and noise-added inside your environment before transmission. Compiled's servers receive only the noised centroid, not anything traceable.
What's the ε parameter and can I control it?
ε=1.0 is the default (strong privacy). You can tighten to ε=0.1 (near-zero information leakage) at some cost to network benefit. Enterprise+ plans allow custom ε.
A 30-minute technical call. We walk through the architecture, show you the cold-start improvement curve, and answer every privacy question you have.