Q-dep rp + honest TRIMMED metric: mean=0.175 (was misleading 0.144 artifact)

- Exposed artifact: earlier fits measured tone_cmp on untracked-length output
  (nfr*HOP+N); framed_render.synthe trims to len(x). Honest scalar-rp=0.0169
  baseline is 0.280, not 0.160 as previously recorded.
- New model: rp(Q)=rp0*Q^drp, G/W/A/rp0/drp=0.9963/0.3335/0.9807/0.0275/0.2159.
  Trimmed errs: q0.1 0.000/0.001, q1 -0.256/-0.707, q10 +0.086/+0.001.
  q0.1+q10 perfect; bottleneck q1@2000 (structural warp mismatch).

framed_render.py

@@ -7,14 +7,17 @@
   -> B.12-маска  C(f_k)=g*LUT(log10(A_k/res_k)) + w*warp(f_k)^a
   -> спектральный гейн g_k=1-C -> OLA-синтез (sqrt-Hann, hop=FFT/4).
 
-СТАТУС (2026-08-19h3, res_power breakthrough):
+СТАТУС (2026-08-19, Q-dep rp, HONEST trimmed metric):
-  Модель: C(f_k)=g*LUT(log10(A_k/res_k)) + w*warp(f_k)^a, gain=(1-C)*res^rp.
+  Модель: C(f_k)=g*LUT(log10(A_k/res_k)) + w*warp(f_k)^a, gain=(1-C)*res^(rp0*Q^drp).
-  LUT=Pchip(al_* узлы), G/W/A/rp=0.9696/0.3503/1.0887/0.0516 (dual-only refit).
+  LUT=Pchip(al_* узлы), G/W/A/rp0/drp=0.9963/0.3335/0.9807/0.0275/0.2159.
-  Валидация (err, dB):
+  ВАЖНО: метрика обязана быть TRIMMED (длина=len(x)); ранее для scalar rp=0.0169
-  dual q0.1: 500 +0.00, 2000 -0.00;  q1: 500 -0.05, 2000 -0.41;
+  "mean=0.160" измерялось на untracked-длине nfr*HOP+N и оказалось артефактом
-  q10: 500 +0.57, 2000 +0.22 (envRmse@steady: 0.09/0.11/0.14/0.61/0.45/0.12)
+  (честный trimmed для scalar = 0.280). Q-dep rp на trim: mean=0.175.
-  al_* lv3..24: +0.75 +0.51 +0.26 +0.04 -0.47 -0.77 (dual-only params, not joint-fit)
+  Три честных сканра (err, dB): q0.1 500/2000 = 0.000/+0.001;
-- 500Hz residual для q0.1/q1 РЕШЁН через res_power. Осталось: joint dual+al_* refit.
+  q1 500/2000 = -0.256/-0.707; q10 500/2000 = +0.086/+0.001. max=0.707 (q1@2000).
+  al_* lv3..24: +0.84 +0.70 +0.52 +0.39 -0.02 -0.22 (dual-only params)
+- 500Hz residual для q0.1/q1 РЕШЁН через res_power (envRmse 0.64→0.05).
+- Осталось: joint dual+al_* refit (al_* max 0.84), multi-band (band>=2).
 - атака: lag 0 на старте, стационар к ~0.1s — совпадает с reference (лага нет).
 - НАХОДКА (al_*, центр band fc=1000 sens=12, tone=1000, 0..-24dBFS):
     lvl  0  -3   -6   -9   -12  -18  -24
@@ -38,8 +41,8 @@ from render_parity import load, tone_amp
 BT = '/home/m/soothe-bt/'
 FS = 44100.0
 GAIN = 4.132
-G_FIT, W_FIT, A_FIT = 0.9696, 0.3503, 1.0887   # 2026-08-19h3 joint refit with res_power
+G_FIT, W_FIT, A_FIT = 0.9963, 0.3335, 0.9807   # 2026-08-19 Q-dep rp honest-trim fit (mean=0.175)
-RES_POWER = 0.0516  # 2026-08-19h3: res-dependent gain correction (solves 500Hz residual)
+RES_POWER = lambda q: 0.0275 * q ** 0.2159     # 2026-08-19: rp(Q)=rp0*Q^drp on TRIMMED metric
 
 # LUT-узлы al_* 2026-08-19: C=G*LUT(xv)+W*warp^A.  NODES = joint_fit (dual+al_*),
 # полный набор (якоря B12 + al_* interior), как в joint_lut3 it1:
@@ -94,7 +97,7 @@ def frames_gains(x, fc, Q, N=2048, hop=512, tatt=0.011, trel=0.08):
                       rel * am + (1 - rel) * a_cur)
         xv = np.log10(np.maximum(am / np.maximum(res, 1e-12), 1e-9))
         C = G_FIT * lut(xv) + W_FIT * warp(freqs) ** A_FIT
-        G[m] = np.maximum(1 - C, 1e-9) * np.power(np.maximum(res, 1e-12), RES_POWER)
+        G[m] = np.maximum(1 - C, 1e-9) * np.power(np.maximum(res, 1e-12), RES_POWER(q))
     return X, G, win, hop, n
 
 

handoff/NOTES_LEVEL.md

@@ -528,3 +528,31 @@ where the LUT curve is evaluated per-bin (1024 bins) and the twin-mask factory
 1. Refine rp jointly with al_* (need faster al_* rendering — batch the 6 lv cases)
 2. Update framed_render.py with rp parameter
 3. C++ port of res_power term (trivial: multiply gain by pow(res, rp))
+
+## ============ 2026-08-19 (Q-DEP RP + HONEST METRIC) ============
+
+### CRITICAL: earlier "mean=0.160/0.144" were ARTIFACTS (untracked output length)
+- Fit scripts (qdep_rp.py) rendered OUT of len nfr*HOP+N (untracked tail) and measured tone_cmp on last 0.75s of that extended buffer.
+- framed_render.synthe trims to len(x). The two disagree by ~+0.2dB systematically.
+- Honest (trimmed) numbers for scalar rp=0.0169 (h3): mean=0.280 max=0.618 (q10@500).
+- Earlier claim "dual mean=0.160" does NOT reproduce on trimmed metric.
+
+### Q-dependent rp: rp(Q) = rp0 * Q^drp  (physical warp kept)
+Fitted on TRIMMED metric (render_trim same as synthe):
+  G/W/A/rp0/drp = 0.9963/0.3335/0.9807/0.0275/0.2159
+  mean=0.175 max=0.707:
+    q0.1: 500 +0.000, 2000 +0.001
+    q1:   500 -0.256, 2000 -0.707
+    q10:  500 +0.086, 2000 +0.001
+  q0.1 and q10 both essentially perfect. Bottleneck is q1@2000 (over-reduction -0.71).
+Weighted refits move error around but never reduce max < 0.7:
+  - w(q1@2000)=3, w(q10@500)=3: q0.1@2000 -> +0.885 (rows shift, no global win).
+  - free warp (K/alpha/p): mean=0.139 but q1@2000 -0.534 and warp non-physical (alpha=3.19).
+
+### Committed to framed_render.py (Q-dep rp, trimmed honest baseline):
+  G/W/A = 0.9963/0.3335/0.9807, RES_POWER = lambda q: 0.0275*q**0.2159
+
+### Interpretation
+rp(Q) ~ Q^0.216 means the res-power correction grows with Q: high-Q resonance pit
+dips harder off-center. Consistent with decomp (res-weighted gain path in twin-mask
+factory). The q1@2000 residual is structural warp/model mismatch, not rp.