Pharos¶
A labeled fleet testbed for federated personalization with a governed disclosure boundary.
Sensitivity Compartments Capacity
What is Pharos?¶
The problem
Federated personalization splits what a model learns into local and shared knowledge. Deciding which stays local is a disclosure problem.
The gap
Public corpora label on a single sensitivity ladder. Real disclosure policy is not a ladder: two holders at the same level with different compartments are incomparable.
Pharos
Synthetic corpora labelled on that lattice, with privacy leakage, over-escalation and personalization measured against content-defined ground truth.
From seed to fleet¶
Generate
Label
Gate
Route
Personalize
Aggregate
A seeded corpus is labelled on the lattice and checked by the shortcut gate. Each item routes to a personal adapter that never leaves its holder, or to a shared fleet adapter aggregated by a robust rule. The architecture walks through every stage and module.
Explore¶
Install, generate a corpus, and read the gate's verdict.
Visual ExplorerThe corpus, the lattice, the shortcut gate, analyst review and every exit priced, in the browser.
FindingsEvery measured result, and the script that reproduces it.
ArchitectureThe problem, the pipeline, and what each module does.
ReferenceThe label lattice, the gate, the release decision and the governance kit.
Releasing a corpusExport a release with its manifest and Croissant metadata.