REI pipeline

Internal property and owner state is already in the sandbox at (FIPS, PropertyId, Period) granularity, streamed from the First American silver layer at Hudi period/FIPS layout, deduped at read by QUALIFY ROW_NUMBER() … ORDER BY _hoodie_commit_time DESC = 1. The feature builder reads only the as-of-T0 columns (base_globs = _globs([t0], fips)), so a row's state cannot draw on snapshots dated after its own T0.

This phase predicts ANY sale. The arms-length document-type filter is intentionally OFF — every sale event counts as y_sold = 1 (foreclosures, quit-claims, probate, divorce included). This is scope, not an oversight (CLAUDE.md hard rule #3); re-running with the filter is a planned second pass once Eduardo signs off. Do not read this as a bug.

Three Tier-D features are under active leakage audit — listing_duration_months, months_since_prev_sale, mortgage_age_months (CLAUDE.md hard rule #7). Do not cite their contribution as validated until the ablation closes. The exact feature count differs across pages — see finding 45 and the model card for the curated set rather than asserting one number here.

Horizon H = 6 months; the eval window is embargoed by horizon_months so train-label and eval windows do not overlap. Finding 32 caught the original layout leaking a 5-month label-window overlap (eval started one month after train-end with H=6); the default now pushes eval start to train_t0s[-1] + horizon + 1 month. Six walk-forward folds run on the 5-county set; the seventh fold is the locked March-2025 test.

The modeling ladder (MASTER_PLAN §6) climbs three rungs — logistic regression (interpretable floor) → LightGBM (production candidate) → state-stratified / hierarchical — each with a go/no-go gate over Alpha. HistGB is the current default per the ablation; treat the matrix as the authority for which architecture wins each county rather than a single verdict here. Class weighting is none — the table arrives downsampled 1:K, and prior-correction back to the true base rate is the calibration step's job (stage 6).

A trained ranker is not a forecaster: raw boosted scores rank well but do not read as calibrated probabilities. Calibration maps scores → probabilities against a held-out, non-downsampled, true-prior slice so the prior-correction the downsampling deferred is applied here. The ±15% calibration tolerance is the macro-model win-condition target (MASTER_PLAN §12), not a measured result — Apollo is still in build.

Cross-county scores are not directly comparable — each county is ranked against itself, so a sidecar (per-county base rate / calibration context) is required before any cross-county comparison. The Alpha head-to-head is read at the top of each county's ranked list. The win condition — top-decile recall ≥ Alpha and ≥ Camilo, with 30/60/90-day calibration within ±15% — is a target on the locked March-2025 test, which remains untouched until Eduardo + Camilo sign off in writing (CLAUDE.md hard rule #4).