Pillar 5 · Confounder-Gated Bayesian Profiler
Judge Vectors
“Litigation is not objective; it is adjudicated by humans.”
The Judge Vectors module models the doctrinal preferences of federal judges using a confounder-gated hierarchical Bayesian estimator. A deterministic LLM classifier separates procedural rulings from merits rulings before any signal updates the doctrinal vector. Counsel disparity weighting further refines the signal to account for asymmetric representation quality.
How It Works
Ingest every published ruling by the assigned judge.
The ruling-basis classifier (LLM) categorizes each ruling as procedural (standing, jurisdiction, timeliness), merits, or mixed.
The Confounder Gate blocks procedural rulings entirely from updating the doctrinal vector. Mixed rulings propagate at half weight.
For merits rulings, compute a counsel disparity factor: rulings against unrepresented or under-resourced parties carry less signal than rulings against sophisticated counsel.
Extract a doctrinal feature vector from the ruling text — positions on statutory interpretation, evidentiary standards, tolerance for circumstantial evidence.
Apply Bayesian shrinkage: the posterior is the precision-weighted combination of the prior and the new observation. As more merits-classified rulings accumulate, the posterior tightens.
Technical Detail
if ruling_basis == "procedural": BLOCK // Jurisdictional denials tell us nothing about merits doctrine
Hon. Jed S. Rakoff · SDNY · Doctrinal Profile
Statutory Interpretation: shifted from 35 → 20 (more textualist). Daubert Scrutiny: shifted from 60 → 75 (stricter). Based on 47 merits-classified rulings after confounder gate; 23 procedural rulings blocked.
47
Merits rulings in Rakoff posterior (after gate)
23
Procedural rulings blocked by confounder gate
±0.04
Posterior standard deviation (high confidence)