PO-LEU

Perceived-Outcome, LLM-generated, Embedding-based Utility. Specification: docs/redesign.md.

Status

Implemented via a wave-based orchestration. See NOTES.md for the running build log, deviations from spec, and unresolved ambiguities.

Wave plan

Wave	Scope	Modules
Scaffold	tree, default config, conftest	configs/default.yaml, tests/conftest.py
1	data prep + cache + prompts	src/data/person_features.py, src/data/context_string.py, src/outcomes/prompts.py, src/outcomes/cache.py
2	LLM generation + diversity	src/outcomes/generate.py, src/outcomes/diversity_filter.py
3	encoder	src/outcomes/encode.py
4	model modules	src/model/attribute_heads.py, src/model/weight_net.py, src/model/salience_net.py
5	assembly + ablations	src/model/po_leu.py, src/model/ablations.py
6	training + eval	src/train/regularizers.py, src/train/loop.py, src/eval/metrics.py, src/eval/strata.py
7	interpretability + ablation configs	src/eval/interpret.py, configs/ablation_*.yaml
Final	smoke test	scripts/smoke_end_to_end.py

Data

Source data lives in amazon_ecom/ (Amazon purchase logs + survey). Upstream v2.0 pipeline stages (load/clean/survey-join/state-features/split) are referenced at their interfaces only — not reimplemented here. src/train/subsample.py is retained verbatim from v1 (Appendix C).

Layout

See redesign.md §14.

Setup

python -m venv venv && source venv/bin/activate
pip install -e '.[dev]'
pytest

Deliverable

Codebase + passing unit tests + scripts/smoke_end_to_end.py that executes one forward + backward pass on synthetic data and emits an interpretability report. No end-to-end training.