Held-out fit (n=2366 prod deals, 2026-06-27) of
log(actual_sold/expected_sold) ~ year + ln(area) gives a multiplicative
factor that corrects systematic mis-estimation by building era + unit size
(underestimates newer buildings, mis-handles large units).
factor = exp(b0 + b_year*(year-2000)/20 + b_larea*ln(area)), clamped
[0.75, 1.30]; applied to the expected_sold point BEFORE the calibrated PI
range block (so the range follows the corrected point). Headline/asking is
untouched. Gated behind estimate_hedonic_correction_enabled (OFF => exact
old behavior, proven by the regression gate).
Frozen baseline regenerated: overall expected_sold MAPE 18.63->14.24;
per-segment median_bias_pct moves toward 0 (бизнес -21.5->-9.2,
комфорт -8.0->-3.6, эконом 17.2->4.3); элит mildly better, no harm.
Tests that assert the exact ratio relation (expected_sold == headline x
ratio) hold the orthogonal hedonic layer OFF.
CLIENT-VISIBLE pricing (killer-factor, demo 2026-06-01). The same-building
anchor replaced headline median/range/confidence with an over-confident
premium on as few as 2 same-building asking comps. Human-chosen (Anton)
3-part safe-guard, all implemented:
1. min_comps 2 -> 4: anchor fires only with >=4 same-building comps.
2. confidence cap: at <5 anchor comps, confidence cannot be 'high' -> medium.
3. MAD-clip: drop price_per_m2 outliers among anchor comps before aggregation
(k=estimate_sb_mad_k=3.5); if <min_comps survive, anchor does NOT fire
(fallback to radius-median). Guards against price typos inflating anchor.
Anchor still REPLACES the headline (uplift-no-replace variant NOT chosen).
quarter-index/radius-tier/asking-sold untouched. 8 new unit tests; 61 anchor
tests pass, ruff clean.
MERGE-GATE (client-visible, like #764/#858): do NOT merge without a green
backtest #763 (held-out MAPE <= 18.7% + canonical A/B Малышева 125 /
Заводская 44а). Backtest needs live DB — pre-merge human/qa step.
Refs #755, #694
