spb/wp2_uqo

N2: scripts/12_specifications_v2.py (11 designs x 5 encoder variants x {raw, centred contrasts, Löwdin}
x {S_main, S_v11, S_nodup} x {rule, langdetect} = 530 cells; unscaled dummies; per-0.10-cosine scale;
HC3 / FSA-clustered t and F(q, G-1); Holm, BH, Romano–Wolf (FSA-cluster bootstrap B=999) on the 55
main-grid cells with a pre-stated 90 % survival criterion; LMG decomposition; domain tests with Holm;
Models A–E on the v2 design; leverage of the raw lot field; --legacy reproduces the 14 v1.2 rows at 1e-14),
scripts/15_spatial_inference.py (Moran's I k=10 with permutation p; Conley 5/10/20 km; decision rule).
N3: scripts/13_multiverse.py (200 phrasing draws x 2 encoders x 3 specs x {raw, centred} x {S_main, S_v11};
the three v1.2 variants reproduced). N4: scripts/14_oos_v2.py (random x20, FSA x20, leave-one-region-out,
GroupKFold by region, k-means distance blocks + 10 km buffer, duplicate-grouped, S_nodup; training-fold FE
with municipality/region/grand-mean fallback; Clark–West, Diebold–Mariano, bootstrap CIs, MCS; price-masked
encoders; --legacy reproduces revision_oos.csv at 1e-16). N5: scripts/16_pca_benchmark.py (Table 9 from a
script; PC1 correlations), scripts/17_inline_numbers.py (results/inline_numbers.json, 2,607 keys),
run_all.sh, tests/ (data, legacy results at rtol 1e-6, embeddings at 1e-5, leakage), environment/
(lock with transitive pins, pip freeze, hardware, HF revisions, SHA-256), data/DATA_AVAILABILITY.md
(users-table exclusion, scrape window 7–11 Jan 2026, shareable extract), LICENSE (MIT), data citation and
model-card entries in references.bib, README runtimes. L2: scripts/18_tables_v2.py (16 tables with labels,
Table 9 regenerated from CSV without self-ratings) and scripts/19_figures_v2.py (fig13–fig16).
results/v2/*.csv|json committed; no data, embeddings or texts.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

v1.1 → v1.2 : analyse de la langue des annonces (40 % anglais, η² = 0,86 sur la
similarité moyenne), effets fixes RTA/municipalité + erreurs groupées, terrain en m²,
sous-échantillon superficie, tests par domaine, validation croisée aléatoire et
spatiale, sensibilité au libellé des références, encodeur multilingue ; sections
réécrites, nouvelle section 7, 8 références vérifiées ajoutées (86 → 94),
docs/REVIEW-REPORT.md + docs/CHANGELOG-REVISION.md, README et data/README mis à jour.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Restructured from immo-wp2-spb-20260519: modular src/ package, numbered
reproduction pipeline (01-06), machine-generated LaTeX tables, regenerated
figures (incl. previously missing fig2/fig3), corrected paper statistics
(see CHANGES.md), 47-page PDF compiling clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>