Claim boundary
This repository contains research infrastructure and a synthetic contract check. It contains no model result, benchmark score, or scientific finding.
The question
When a language model has a stable representation of demographic identity, does that representation causally transport into human-aligned subgroup response differences?
A reproducible identity direction establishes that the representation can be measured. Further evidence is needed to show that the model knows how a human subgroup would answer a survey question, or that steering the direction will express that knowledge correctly.
STRATA Transport treats the following stages separately:
- identity encoding;
- behavioral transport;
- alignment with survey contrasts;
- calibration of location and magnitude.
What is implemented
- compositional normalization and orthonormal ILR transforms;
- cell, item, and crossed bootstrap primitives that preserve multiplicity;
- weighted survey distributions and Kish effective sample size;
- option-count-invariant total variation and adversarial controls;
- deterministic dev and test partitions with canonical hashes;
- typed model, dataset, run, artifact, and result manifests;
- exact run-manifest and result binding;
- fail-closed, structured command execution;
- an executable synthetic-only vertical slice with byte-hashed artifacts.
The audited release has 147 passing tests and 96% branch coverage.
What comes next
The remaining empirical work requires contamination probes, option-order controls, shuffled-persona controls, survey-aware uncertainty, and a frozen model and data plan. The current synthetic run verifies that the software contract passes through all gates; it provides no evidence on the empirical question.
Empirical work will wait until the controls and frozen plan are in place, settling provenance decisions before any result is visible.