force recovery: methods, evaluation, debug log

.gitattributes

@@ -128,3 +128,9 @@ assets/panel_faf.png filter=lfs diff=lfs merge=lfs -text
 assets/panel_feats.png filter=lfs diff=lfs merge=lfs -text
 assets/panel_fota_cnc.png filter=lfs diff=lfs merge=lfs -text
 assets/panel_sparsh.png filter=lfs diff=lfs merge=lfs -text
+assets/panel_cnc_mini_26_zh.png filter=lfs diff=lfs merge=lfs -text
+assets/panel_faf_zh.png filter=lfs diff=lfs merge=lfs -text
+assets/panel_feats_zh.png filter=lfs diff=lfs merge=lfs -text
+assets/panel_fota_cnc_zh.png filter=lfs diff=lfs merge=lfs -text
+assets/panel_sparsh_zh.png filter=lfs diff=lfs merge=lfs -text
+assets/react_normals.png filter=lfs diff=lfs merge=lfs -text

index.html

@@ -59,11 +59,11 @@ newtons in it are estimated from the images alone — reconstructed to a depth
 map, then calibrated on presses of known load.</p>
 
 <div class="cards">
-<div class="card"><b>0.986</b><span>best Spearman ρ vs
+<div class="card"><b>0.968</b><span>best Spearman ρ vs
 ground-truth force</span></div>
 <div class="card"><b>5</b><span>force-labelled datasets, one
 protocol</span></div>
-<div class="card"><b>0.857</b><span>agreement between two
+<div class="card"><b>0.869</b><span>agreement between two
 estimators sharing no calibration</span></div>
 <div class="card"><b>0</b><span>frames of our own rig in any
 calibration</span></div>
@@ -80,8 +80,8 @@ all.</figcaption></figure>
 lookup table learns it from ~700 sphere presses per sensor; a linear
 photometric solve needs only the LED geometry. On markerless GelSight&nbsp;Mini
 presses the calibration-free solve leads by
-+0.27&nbsp;ρ; on a marker gel
-it trails by -0.14, which is
++0.22&nbsp;ρ; on a marker gel
+it trails by -0.12, which is
 what the physics predicts.</p>
 <p class="dim">Full comparison on the <a href="results.html">results</a> page ·
 how it works on <a href="method.html">method</a>.</p>

results.html

@@ -58,7 +58,7 @@ img{max-width:100%;height:auto}
 frames fit a 5-feature least squares with isotonic calibration, half are
 scored; pooled Spearman ρ, five seeds, beside a within-group label shuffle.</p>
 
-<div class="tablewrap"><table><thead><tr><th>dataset</th><th>n</th><th>LUT</th><th>calibration-free</th><th>FEATS U-net</th><th>FeelAnyForce</th><th>shuffle</th></tr></thead><tbody><tr><td>Mini CNC 2026 (markerless, 0-20 N)</td><td>201</td><td>0.986</td><td>0.732</td><td>—</td><td>—</td><td class='dim'>+0.171</td></tr><tr><td>FoTa cnc_Mini (markerless, in view)</td><td>337</td><td>0.946</td><td>0.157</td><td>0.074</td><td>0.828</td><td class='dim'>+0.056</td></tr><tr><td>FEATS (marker gel)</td><td>186</td><td>0.775</td><td>0.615</td><td>0.960</td><td>0.431</td><td class='dim'>-0.003</td></tr><tr><td>Sparsh / Meta (markerless, Sparsh LUT, in view)</td><td>1667</td><td>0.968</td><td>—</td><td>—</td><td>—</td><td class='dim'>+0.264</td></tr><tr><td style='padding-left:26px;opacity:.75'>↳ control: same frames, dots inpainted (depth path; not used for force)</td><td>186</td><td>0.737</td><td>—</td><td>—</td><td>—</td><td class='dim'>-0.010</td></tr></tbody></table></div>
+<div class="tablewrap"><table><thead><tr><th>dataset</th><th>n</th><th>LUT</th><th>calibration-free</th><th>FEATS U-net</th><th>FeelAnyForce</th><th>shuffle</th></tr></thead><tbody><tr><td>cnc_mini_26 (markerless, 0-20 N)</td><td>588</td><td>0.968</td><td>0.838</td><td>0.041</td><td>0.900</td><td class='dim'>+0.195</td></tr><tr><td>FoTa cnc_Mini (markerless, in view)</td><td>337</td><td>0.915</td><td>0.955</td><td>0.074</td><td>0.828</td><td class='dim'>-0.030</td></tr><tr><td>FEATS (marker gel)</td><td>186</td><td>0.775</td><td>0.639</td><td>0.960</td><td>0.431</td><td class='dim'>-0.003</td></tr><tr><td>Sparsh / Meta (markerless, Sparsh LUT, in view)</td><td>1667</td><td>0.968</td><td>—</td><td>—</td><td>—</td><td class='dim'>+0.264</td></tr><tr><td style='padding-left:26px;opacity:.75'>↳ control: same frames, dots inpainted (depth path; not used for force)</td><td>186</td><td>0.632</td><td>—</td><td>—</td><td>—</td><td class='dim'>+0.052</td></tr></tbody></table></div>
 <p class="dim">“—” means not run, not zero. Calibration-free is scored under a
 stripped protocol with no position gain field, so its column is not comparable
 to the LUT's headline value — it is comparable to the LUT under the same strip,
@@ -68,9 +68,9 @@ which is on the <a href="method.html">method</a> page.</p>
 <p>React has no labels, so the test is whether a second estimator sharing no
 calibration does better on React's own calibration domain. It does not:
 0.739 against 0.310 held out by press position. The
-two agree at ρ&nbsp;=&nbsp;0.857 over 1,500 frames, mean
-difference 0.72&nbsp;N. That is consistent with both measuring
+two agree at ρ&nbsp;=&nbsp;0.869 over 1,800 frames, mean
+difference 0.80&nbsp;N. That is consistent with both measuring
 contact and is not a certification of absolute scale — p95 disagreement is
-3.39&nbsp;N against a 7.29&nbsp;N ceiling.</p>
+3.22&nbsp;N against a 7.29&nbsp;N ceiling.</p>
 <p><b>No recompute.</b> The published newtons stand.</p>
 </div></body></html>