Insights
Research & Technical Deep Dives
Detailed case studies, implementation guides, and results from governed-rank across domains.
Start Here
Understanding governed-rank: How MOSAIC Steers Rankings Without Breaking Them
Every ranking system eventually needs a second objective. MOSAIC orthogonalizes the policy signal, protects confident decisions, and projects the optimal result — in three steps, one function call.
3 stepsZero-interference steering with full audit trail
Tutorial: Content Moderation — Demoting Toxicity Without Killing Engagement
Toxic content is engaging — outrage drives clicks. This tutorial walks through the content_moderation notebook: why naive penalties over-correct, and how govern() targets only the uncertain zone.
Toxicity drops in top-10 while ranking quality preserved
Tutorial: Fairness — Reducing Racial Bias in COMPAS Risk Rankings
The COMPAS recidivism dataset is the canonical example of algorithmic bias. This tutorial walks through steering toward demographic parity with govern(), computing adverse impact ratios, and inspecting the audit trail.
AIR improves from ~0.77 to ~0.92, passing the 4/5ths rule
Tutorial: Fraud Detection — Steering Review Queues Toward High-Value Fraud
A fraud model ranks by probability, but a $50K wire transfer matters more than a $5 candy purchase. This tutorial walks through the fraud_detection notebook: how govern() steers toward high-impact fraud without flooding the queue with false positives.
10x fraud value captured in BLOCK tier vs base model
Tutorial: Objective Discovery — Finding Policies That Work Before You Deploy
Not all policies are worth pursuing. This tutorial walks through the objective_discovery notebook: run 7 candidate policies through govern() to discover which objectives align with users and which fight them.
quality_depth is the only policy with high preference lift AND diversity gain
Tutorial: RAG Safety — Blocking Prompt Injection Without Breaking Retrieval
Attackers craft high-relevance documents with hidden prompt injections. This tutorial walks through the rag_safety notebook: how govern() removes injected documents from the citation window while preserving retrieval quality.
All injected docs removed from top-10 with 65% quality retained
One Algorithm, Six Datasets, Four Domains
Does govern() actually generalize? We ran the same function call — no retuning — across grocery, movies, fashion, news, and music. Policy lifts from 1.15× to 41.7×, budget controls stability smoothly in every domain.
1.15×–41.7× policy lift across 6 real datasets with zero retuning
Why Forcing Diversity Backfires — And What Works Instead
We tested 34 policy candidates on 54,544 real news sessions. Diversity-forcing policies scored 0.25× preference lift. Quality-based steering scored 2.00× lift AND increased diversity by 47%. The data says: optimize quality, get diversity for free.
Quality steering: 2.00× preference lift AND +47% click diversity