Quark Core

Training models without surrendering data.

Centralization is not an engineering default — it is a political choice.

Overview

Quark Core is an end-to-end reference implementation for training machine learning models across a network of participants without any of them surrendering raw data to a central authority.

The build is the argument: every design decision is also a position on where epistemic power should sit, who is entitled to refuse, and what the protocol guarantees when participants do not trust the aggregator.

What it is

A minimal, auditable framework — server, client, and aggregation primitives — designed so that the security properties are legible without reading the entire codebase.

The goal is not feature parity with industrial federated learning toolkits. The goal is to write down, in code, what the simplest defensible system looks like.

What it is not

A product. The framework is research infrastructure. It will not ship with hosted services, paid tiers, or customer success motions. It exists so that the position papers it grounds can be falsified by anyone willing to read the source.

Goals

01
Publish a working secure-aggregation protocol resistant to partial collusion among participants.
02
Document the threat models the framework does and does not defend against — explicitly, not implicitly.
03
Produce one position paper per major design decision, each one falsifiable against the running code.

Field notes