Scheduled Commit

ablation_runs/ablation_section_preview.html

@@ -0,0 +1,263 @@
+
+  <!-- ======================= ABLATION ======================= -->
+  <section id="ablation">
+    <p class="sec-eyebrow">ablation</p>
+    <h2>Ablation: what track, force, and deformation tactile each buy</h2>
+
+    <div class="prose">
+      <p>Four variants of the cascade are trained from the same Wan2.2 foundation
+      weights on the <code>trex_mini_force</code> episode &mdash; identical anchors,
+      LoRA budget (r=4 on q,k,v,o,k_img,v_img,ffn), learning rate, batch size,
+      seed, and 3,000 single-GPU steps &mdash; so every difference below is
+      attributable to the ablated component, not to training budget.
+      <b>B</b> removes the track pathway structurally (no track tokens, encoder,
+      decoder, or loss): the packed block shrinks from 617 to 117 tokens &mdash;
+      exactly the pre-track architecture with the force stage kept. <b>C</b>
+      removes the force stage (no Stage-2 transformer, no tactile input, no
+      force/VQ loss); at inference its coarse model carries the flow all the way
+      to &tau;=0. <b>D</b> keeps the full model and adds ten per-finger
+      deformation-camera tokens (96&times;96 crops &rarr; 4-layer CNN &rarr; d 768)
+      beside the F6 and VQ-history tokens in Stage 2.</p>
+    </div>
+
+  <div class="legend" role="list" aria-label="Ablation variant legend">
+    <span class="k" role="listitem"><span class="vs vs-full"></span>A &middot; full (track + force)</span>
+    <span class="k" role="listitem"><span class="vs vs-notrack"></span>B &middot; no track (pre-track structure)</span>
+    <span class="k" role="listitem"><span class="vs vs-noforce"></span>C &middot; no force stage</span>
+    <span class="k" role="listitem"><span class="vs vs-deform"></span>D &middot; deform tactile + force</span>
+  </div>
+
+
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>Variant</th><th>Stage-1 tokens / block</th>
+        <th>Stage-2 tokens</th><th>Losses optimized</th></tr></thead>
+        <tbody>
+          <tr><td><span class="chip"><span class="vs vs-full"></span>&nbsp;A &middot; full</span></td>
+            <td class="shape">100 obs + 16 act + 1 st + 500 trk = 617</td>
+            <td class="shape">16 act + 10 F6 + 10 VQ + mem</td>
+            <td>dynamics &middot; action &middot; 20&middot;track &middot; force &middot; VQ</td></tr>
+          <tr><td><span class="chip"><span class="vs vs-notrack"></span>&nbsp;B &middot; no track</span></td>
+            <td class="shape">100 obs + 16 act + 1 st = 117</td>
+            <td class="shape">16 act + 10 F6 + 10 VQ + mem</td>
+            <td>dynamics &middot; action &middot; force &middot; VQ</td></tr>
+          <tr><td><span class="chip"><span class="vs vs-noforce"></span>&nbsp;C &middot; no force</span></td>
+            <td class="shape">617</td>
+            <td class="shape">&mdash; (coarse runs &tau;&rarr;0)</td>
+            <td>dynamics &middot; action &middot; 20&middot;track</td></tr>
+          <tr><td><span class="chip"><span class="vs vs-deform"></span>&nbsp;D &middot; deform + force</span></td>
+            <td class="shape">617</td>
+            <td class="shape">16 act + 10 F6 + 10 VQ + <b>10 deform</b> + mem</td>
+            <td>dynamics &middot; action &middot; 20&middot;track &middot; force &middot; VQ</td></tr>
+        </tbody>
+      </table>
+    </div>
+
+  <figure>
+    <div class="quad">
+<div><svg viewBox="0 0 520 230" role="img" aria-label="action flow MSE">
+<text class="th" x="46" y="15">action flow MSE</text>
+<line class="grid-cell" x1="46" y1="157.3" x2="510" y2="157.3"></line>
+<text class="ts" x="40" y="160.3" text-anchor="end">0.5</text>
+<line class="grid-cell" x1="46" y1="108.6" x2="510" y2="108.6"></line>
+<text class="ts" x="40" y="111.6" text-anchor="end">1</text>
+<line class="grid-cell" x1="46" y1="59.9" x2="510" y2="59.9"></line>
+<text class="ts" x="40" y="62.9" text-anchor="end">2</text>
+<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
+<text class="ts" x="46.0" y="214" text-anchor="middle">0k</text>
+<line class="grid-cell" x1="200.7" y1="26" x2="200.7" y2="200"></line>
+<text class="ts" x="200.7" y="214" text-anchor="middle">1k</text>
+<line class="grid-cell" x1="355.3" y1="26" x2="355.3" y2="200"></line>
+<text class="ts" x="355.3" y="214" text-anchor="middle">2k</text>
+<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
+<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
+<text class="ts" x="510" y="224" text-anchor="end">step</text>
+<path class="vw v-full" d="M46.0 26.5 L49.1 26.0 L52.2 28.0 L55.3 32.8 L58.4 38.5 L61.5 44.3 L64.6 49.8 L67.7 55.2 L70.7 60.0 L73.8 64.6 L76.9 68.9 L80.0 72.8 L83.1 76.6 L86.2 79.8 L89.3 82.9 L92.4 85.6 L95.5 88.2 L98.6 90.6 L101.7 92.8 L104.8 94.7 L107.9 96.4 L111.0 98.0 L114.1 99.5 L117.1 100.9 L120.2 102.3 L123.3 103.5 L126.4 104.7 L129.5 105.7 L132.6 106.8 L135.7 107.9 L138.8 109.3 L141.9 110.6 L145.0 111.5 L148.1 112.8 L151.2 114.3 L154.3 115.8 L157.4 116.9 L160.5 118.1 L163.5 119.3 L166.6 120.9 L169.7 122.3 L172.8 123.5 L175.9 125.0 L179.0 126.2 L182.1 127.8 L185.2 129.2 L188.3 130.7 L191.4 131.7 L194.5 133.7 L197.6 135.2 L200.7 137.2 L203.8 139.4 L206.9 140.8 L209.9 141.8 L213.0 144.1 L216.1 146.2 L219.2 147.8 L222.3 149.4 L225.4 150.5 L228.5 152.0 L231.6 153.7 L234.7 154.8 L237.8 155.7 L240.9 157.5 L244.0 158.7 L247.1 160.0 L250.2 161.5 L253.3 161.1 L256.3 161.1 L259.4 161.7 L262.5 163.0 L265.6 164.3 L268.7 165.4 L271.8 165.6 L274.9 166.5 L278.0 168.2 L281.1 169.9 L284.2 170.6 L287.3 170.9 L290.4 172.1 L293.5 171.9 L296.6 172.1 L299.7 173.0 L302.7 174.2 L305.8 176.0 L308.9 177.1 L312.0 177.5 L315.1 177.8 L318.2 179.0 L321.3 179.9 L324.4 181.0 L327.5 181.5 L330.6 181.8 L333.7 183.7 L336.8 185.5 L339.9 184.7 L343.0 183.4 L346.1 183.9 L349.1 183.0 L352.2 184.5 L355.3 184.1 L358.4 184.9 L361.5 184.2 L364.6 184.3 L367.7 184.7 L370.8 187.0 L373.9 187.2 L377.0 188.9 L380.1 189.1 L383.2 187.6 L386.3 188.4 L389.4 189.2 L392.5 190.8 L395.5 191.1 L398.6 189.5 L401.7 187.6 L404.8 188.7 L407.9 190.3 L411.0 190.7 L414.1 191.6 L417.2 191.4 L420.3 189.8 L423.4 190.5 L426.5 191.2 L429.6 191.7 L432.7 192.6 L435.8 193.2 L438.9 193.9 L441.9 192.9 L445.0 194.0 L448.1 195.0 L451.2 195.3 L454.3 195.7 L457.4 196.1 L460.5 197.1 L463.6 195.4 L466.7 195.5 L469.8 197.0 L472.9 196.7 L476.0 196.7 L479.1 196.5 L482.2 196.1 L485.3 196.2 L488.3 195.9 L491.4 197.2 L494.5 197.2 L497.6 196.9 L500.7 194.4 L503.8 196.5 L506.9 199.6 L508.5 198.1"></path>
+<circle class="vdot v-full" cx="508.5" cy="198.1" r="3.5"></circle>
+<path class="vw v-notrack" d="M46.0 30.7 L49.1 30.4 L52.2 32.6 L55.3 37.6 L58.4 43.3 L61.5 49.0 L64.6 54.4 L67.7 59.6 L70.7 64.1 L73.8 68.5 L76.9 72.6 L80.0 76.2 L83.1 79.7 L86.2 82.8 L89.3 85.7 L92.4 88.1 L95.5 90.4 L98.6 92.7 L101.7 94.6 L104.8 96.4 L107.9 98.0 L111.0 99.6 L114.1 101.0 L117.1 102.4 L120.2 103.7 L123.3 104.9 L126.4 106.2 L129.5 107.4 L132.6 108.6 L135.7 110.0 L138.8 110.9 L141.9 111.9 L145.0 113.2 L148.1 114.6 L151.2 115.8 L154.3 117.1 L157.4 118.4 L160.5 120.0 L163.5 121.1 L166.6 122.1 L169.7 123.5 L172.8 125.8 L175.9 127.2 L179.0 128.5 L182.1 130.0 L185.2 131.4 L188.3 132.8 L191.4 134.4 L194.5 135.5 L197.6 136.5 L200.7 138.1 L203.8 139.8 L206.9 141.5 L209.9 142.3 L213.0 143.6 L216.1 145.9 L219.2 148.1 L222.3 149.2 L225.4 151.1 L228.5 152.1 L231.6 152.9 L234.7 154.2 L237.8 155.7 L240.9 157.2 L244.0 159.3 L247.1 160.4 L250.2 162.7 L253.3 165.1 L256.3 166.4 L259.4 167.1 L262.5 168.0 L265.6 169.0 L268.7 169.6 L271.8 171.4 L274.9 172.0 L278.0 172.7 L281.1 174.5 L284.2 175.0 L287.3 176.4 L290.4 175.0 L293.5 176.0 L296.6 176.5 L299.7 177.5 L302.7 178.4 L305.8 180.5 L308.9 181.0 L312.0 182.2 L315.1 182.0 L318.2 184.0 L321.3 183.9 L324.4 182.7 L327.5 183.2 L330.6 183.9 L333.7 185.0 L336.8 186.3 L339.9 187.0 L343.0 188.0 L346.1 188.2 L349.1 186.5 L352.2 187.1 L355.3 187.9 L358.4 187.8 L361.5 189.6 L364.6 189.3 L367.7 192.2 L370.8 191.2 L373.9 191.5 L377.0 192.8 L380.1 193.0 L383.2 194.9 L386.3 195.4 L389.4 196.5 L392.5 196.9 L395.5 197.0 L398.6 194.9 L401.7 195.6 L404.8 195.7 L407.9 195.8 L411.0 196.6 L414.1 197.3 L417.2 196.3 L420.3 198.0 L423.4 197.2 L426.5 197.7 L429.6 197.8 L432.7 197.2 L435.8 196.2 L438.9 197.0 L441.9 195.6 L445.0 195.5 L448.1 194.6 L451.2 194.0 L454.3 193.7 L457.4 194.0 L460.5 194.2 L463.6 195.3 L466.7 194.2 L469.8 193.7 L472.9 192.2 L476.0 194.7 L479.1 194.5 L482.2 194.4 L485.3 196.0 L488.3 196.9 L491.4 194.9 L494.5 197.3 L497.6 200.0 L500.7 199.8 L503.8 199.9 L506.9 198.8 L508.5 198.2"></path>
+<circle class="vdot v-notrack" cx="508.5" cy="198.2" r="3.5"></circle>
+<path class="vw v-noforce" d="M46.0 26.5 L47.5 26.4 L49.1 26.2 L50.6 26.4 L52.2 27.9 L53.7 30.2 L55.3 32.7 L56.8 35.6 L58.4 38.5 L59.9 41.3 L61.5 44.3 L63.0 47.3 L64.6 49.9 L66.1 52.8 L67.7 55.5 L69.2 57.8 L70.7 60.2 L72.3 62.5 L73.8 64.7 L75.4 66.9 L76.9 69.0 L78.5 71.0 L80.0 72.9 L81.6 74.8 L83.1 76.8 L84.7 78.3 L86.2 80.0 L87.8 81.6 L89.3 83.0 L90.9 84.4 L92.4 85.8 L93.9 87.2 L95.5 88.4 L97.0 89.7 L98.6 90.8 L100.1 91.9 L101.7 93.0 L103.2 94.0 L104.8 94.9 L106.3 95.9 L107.9 96.7 L109.4 97.6 L111.0 98.4 L112.5 99.0 L114.1 99.7 L115.6 100.5 L117.1 101.2 L118.7 101.9 L120.2 102.6 L121.8 103.1 L123.3 103.6 L124.9 104.4 L126.4 104.9 L128.0 105.4 L129.5 106.1 L131.1 106.6 L132.6 107.2 L134.2 107.9 L135.7 108.5 L137.3 109.1 L138.8 109.8 L140.3 110.5 L141.9 111.2 L143.4 111.6 L145.0 112.3 L146.5 112.8 L148.1 113.4 L149.6 113.9 L151.2 114.5 L152.7 115.1 L154.3 115.5 L155.8 116.3 L157.4 117.0 L158.9 117.5 L160.5 118.6 L162.0 119.3 L163.5 119.9 L165.1 120.7 L166.6 121.5 L168.2 122.4 L169.7 123.2 L171.3 123.9 L172.8 124.6 L174.4 125.1 L175.9 126.0 L177.5 126.3 L179.0 127.3 L180.6 128.0 L182.1 128.7 L183.7 129.4 L185.2 130.1 L186.7 130.7"></path>
+<circle class="vdot v-noforce" cx="186.7" cy="130.7" r="3.5"></circle>
+</svg></div>
+<div><svg viewBox="0 0 520 230" role="img" aria-label="dynamics (video flow) MSE">
+<text class="th" x="46" y="15">dynamics (video flow) MSE</text>
+<line class="grid-cell" x1="46" y1="154.8" x2="510" y2="154.8"></line>
+<text class="ts" x="40" y="157.8" text-anchor="end">0.5</text>
+<line class="grid-cell" x1="46" y1="100.1" x2="510" y2="100.1"></line>
+<text class="ts" x="40" y="103.1" text-anchor="end">1</text>
+<line class="grid-cell" x1="46" y1="45.5" x2="510" y2="45.5"></line>
+<text class="ts" x="40" y="48.5" text-anchor="end">2</text>
+<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
+<text class="ts" x="46.0" y="214" text-anchor="middle">0k</text>
+<line class="grid-cell" x1="200.7" y1="26" x2="200.7" y2="200"></line>
+<text class="ts" x="200.7" y="214" text-anchor="middle">1k</text>
+<line class="grid-cell" x1="355.3" y1="26" x2="355.3" y2="200"></line>
+<text class="ts" x="355.3" y="214" text-anchor="middle">2k</text>
+<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
+<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
+<text class="ts" x="510" y="224" text-anchor="end">step</text>
+<path class="vw v-full" d="M46.0 26.0 L49.1 26.7 L52.2 27.7 L55.3 29.9 L58.4 32.4 L61.5 35.6 L64.6 39.2 L67.7 42.6 L70.7 46.5 L73.8 51.2 L76.9 56.4 L80.0 62.9 L83.1 70.2 L86.2 77.0 L89.3 83.9 L92.4 90.3 L95.5 96.4 L98.6 102.3 L101.7 107.9 L104.8 113.9 L107.9 118.0 L111.0 121.5 L114.1 125.7 L117.1 129.8 L120.2 133.8 L123.3 136.2 L126.4 139.4 L129.5 141.8 L132.6 144.5 L135.7 146.7 L138.8 149.2 L141.9 152.2 L145.0 152.3 L148.1 154.4 L151.2 156.8 L154.3 159.6 L157.4 159.8 L160.5 160.6 L163.5 162.0 L166.6 163.5 L169.7 164.5 L172.8 164.5 L175.9 165.1 L179.0 165.2 L182.1 166.1 L185.2 166.5 L188.3 167.2 L191.4 166.8 L194.5 168.7 L197.6 169.3 L200.7 170.8 L203.8 172.3 L206.9 172.5 L209.9 171.8 L213.0 173.8 L216.1 175.3 L219.2 175.5 L222.3 176.0 L225.4 175.6 L228.5 175.9 L231.6 176.9 L234.7 176.7 L237.8 176.8 L240.9 178.0 L244.0 178.0 L247.1 178.6 L250.2 179.2 L253.3 177.2 L256.3 175.7 L259.4 175.8 L262.5 175.9 L265.6 176.0 L268.7 176.5 L271.8 176.0 L274.9 175.8 L278.0 176.9 L281.1 177.5 L284.2 177.8 L287.3 177.3 L290.4 178.2 L293.5 177.7 L296.6 177.4 L299.7 177.8 L302.7 178.3 L305.8 179.6 L308.9 180.2 L312.0 180.6 L315.1 180.1 L318.2 180.8 L321.3 181.5 L324.4 181.8 L327.5 181.8 L330.6 181.7 L333.7 183.1 L336.8 184.2 L339.9 183.1 L343.0 181.1 L346.1 181.6 L349.1 180.4 L352.2 181.5 L355.3 180.7 L358.4 181.1 L361.5 179.9 L364.6 179.7 L367.7 179.6 L370.8 181.7 L373.9 181.6 L377.0 182.8 L380.1 182.8 L383.2 181.4 L386.3 182.1 L389.4 182.7 L392.5 183.9 L395.5 184.0 L398.6 182.6 L401.7 180.5 L404.8 181.2 L407.9 182.2 L411.0 182.6 L414.1 183.2 L417.2 183.2 L420.3 181.7 L423.4 182.0 L426.5 182.6 L429.6 182.9 L432.7 183.7 L435.8 184.2 L438.9 184.5 L441.9 183.4 L445.0 184.8 L448.1 185.4 L451.2 185.7 L454.3 186.0 L457.4 186.3 L460.5 187.1 L463.6 185.6 L466.7 185.4 L469.8 186.6 L472.9 186.3 L476.0 186.3 L479.1 185.9 L482.2 185.4 L485.3 185.8 L488.3 185.8 L491.4 186.8 L494.5 186.5 L497.6 186.5 L500.7 184.5 L503.8 186.0 L506.9 188.3 L508.5 187.2"></path>
+<circle class="vdot v-full" cx="508.5" cy="187.2" r="3.5"></circle>
+<path class="vw v-notrack" d="M46.0 28.4 L49.1 29.7 L52.2 30.3 L55.3 32.5 L58.4 35.4 L61.5 38.5 L64.6 43.1 L67.7 47.8 L70.7 53.1 L73.8 59.4 L76.9 66.2 L80.0 73.0 L83.1 80.1 L86.2 87.1 L89.3 94.4 L92.4 101.1 L95.5 107.6 L98.6 113.1 L101.7 117.8 L104.8 122.6 L107.9 127.3 L111.0 131.2 L114.1 135.5 L117.1 139.6 L120.2 143.0 L123.3 146.5 L126.4 150.4 L129.5 153.2 L132.6 155.2 L135.7 158.0 L138.8 158.5 L141.9 159.2 L145.0 160.8 L148.1 163.4 L151.2 164.5 L154.3 166.7 L157.4 168.3 L160.5 169.4 L163.5 169.8 L166.6 169.6 L169.7 170.2 L172.8 173.7 L175.9 174.5 L179.0 175.2 L182.1 175.1 L185.2 175.9 L188.3 176.4 L191.4 177.2 L194.5 177.1 L197.6 176.4 L200.7 177.3 L203.8 178.1 L206.9 178.1 L209.9 178.1 L213.0 178.2 L216.1 180.4 L219.2 181.8 L222.3 181.4 L225.4 183.0 L228.5 182.8 L231.6 182.4 L234.7 182.6 L237.8 183.6 L240.9 184.7 L244.0 186.5 L247.1 187.0 L250.2 188.7 L253.3 190.4 L256.3 190.6 L259.4 191.0 L262.5 190.9 L265.6 190.8 L268.7 190.5 L271.8 192.1 L274.9 191.2 L278.0 190.8 L281.1 191.9 L284.2 191.9 L287.3 192.5 L290.4 190.3 L293.5 190.5 L296.6 190.5 L299.7 190.8 L302.7 191.3 L305.8 193.0 L308.9 192.8 L312.0 193.9 L315.1 192.7 L318.2 194.6 L321.3 194.2 L324.4 192.5 L327.5 192.5 L330.6 192.8 L333.7 193.5 L336.8 194.1 L339.9 194.2 L343.0 194.7 L346.1 194.7 L349.1 193.1 L352.2 193.3 L355.3 193.9 L358.4 193.6 L361.5 194.6 L364.6 193.7 L367.7 196.1 L370.8 195.1 L373.9 195.4 L377.0 196.2 L380.1 196.5 L383.2 198.2 L386.3 198.5 L389.4 199.1 L392.5 199.1 L395.5 199.4 L398.6 197.6 L401.7 197.8 L404.8 197.6 L407.9 197.2 L411.0 197.9 L414.1 198.4 L417.2 197.6 L420.3 199.4 L423.4 198.6 L426.5 198.6 L429.6 198.5 L432.7 197.4 L435.8 196.6 L438.9 197.4 L441.9 196.3 L445.0 196.4 L448.1 195.3 L451.2 195.0 L454.3 194.8 L457.4 195.0 L460.5 195.3 L463.6 196.2 L466.7 194.8 L469.8 194.1 L472.9 192.5 L476.0 194.7 L479.1 194.5 L482.2 194.2 L485.3 196.0 L488.3 197.3 L491.4 195.2 L494.5 197.2 L497.6 200.0 L500.7 199.8 L503.8 199.8 L506.9 198.7 L508.5 198.1"></path>
+<circle class="vdot v-notrack" cx="508.5" cy="198.1" r="3.5"></circle>
+<path class="vw v-noforce" d="M46.0 26.0 L47.5 26.3 L49.1 26.9 L50.6 27.3 L52.2 27.9 L53.7 28.7 L55.3 29.9 L56.8 31.1 L58.4 32.4 L59.9 33.5 L61.5 35.1 L63.0 36.7 L64.6 38.6 L66.1 40.5 L67.7 42.5 L69.2 44.6 L70.7 46.4 L72.3 48.7 L73.8 51.0 L75.4 53.4 L76.9 56.5 L78.5 59.5 L80.0 62.9 L81.6 66.4 L83.1 70.1 L84.7 73.7 L86.2 77.1 L87.8 80.4 L89.3 83.5 L90.9 86.9 L92.4 90.3 L93.9 93.3 L95.5 96.4 L97.0 99.7 L98.6 102.5 L100.1 105.2 L101.7 108.4 L103.2 111.7 L104.8 114.6 L106.3 116.8 L107.9 118.7 L109.4 121.1 L111.0 123.2 L112.5 125.2 L114.1 127.2 L115.6 129.2 L117.1 131.5 L118.7 133.2 L120.2 134.9 L121.8 136.4 L123.3 137.3 L124.9 139.6 L126.4 141.5 L128.0 142.1 L129.5 144.0 L131.1 144.8 L132.6 147.0 L134.2 148.2 L135.7 148.9 L137.3 149.9 L138.8 151.6 L140.3 152.6 L141.9 153.9 L143.4 154.4 L145.0 155.6 L146.5 156.4 L148.1 157.6 L149.6 157.8 L151.2 159.1 L152.7 159.4 L154.3 159.3 L155.8 160.5 L157.4 161.1 L158.9 161.4 L160.5 162.6 L162.0 163.7 L163.5 163.8 L165.1 164.4 L166.6 165.6 L168.2 166.8 L169.7 167.5 L171.3 167.6 L172.8 167.6 L174.4 167.3 L175.9 167.7 L177.5 166.9 L179.0 167.3 L180.6 167.3 L182.1 167.2 L183.7 167.2 L185.2 167.2 L186.7 166.9"></path>
+<circle class="vdot v-noforce" cx="186.7" cy="166.9" r="3.5"></circle>
+</svg></div>
+<div><svg viewBox="0 0 520 230" role="img" aria-label="track loss (weighted sum)">
+<text class="th" x="46" y="15">track loss (weighted sum)</text>
+<line class="grid-cell" x1="46" y1="185.5" x2="510" y2="185.5"></line>
+<text class="ts" x="40" y="188.5" text-anchor="end">0.2</text>
+<line class="grid-cell" x1="46" y1="145.2" x2="510" y2="145.2"></line>
+<text class="ts" x="40" y="148.2" text-anchor="end">0.5</text>
+<line class="grid-cell" x1="46" y1="114.6" x2="510" y2="114.6"></line>
+<text class="ts" x="40" y="117.6" text-anchor="end">1</text>
+<line class="grid-cell" x1="46" y1="84.1" x2="510" y2="84.1"></line>
+<text class="ts" x="40" y="87.1" text-anchor="end">2</text>
+<line class="grid-cell" x1="46" y1="43.8" x2="510" y2="43.8"></line>
+<text class="ts" x="40" y="46.8" text-anchor="end">5</text>
+<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
+<text class="ts" x="46.0" y="214" text-anchor="middle">0k</text>
+<line class="grid-cell" x1="200.7" y1="26" x2="200.7" y2="200"></line>
+<text class="ts" x="200.7" y="214" text-anchor="middle">1k</text>
+<line class="grid-cell" x1="355.3" y1="26" x2="355.3" y2="200"></line>
+<text class="ts" x="355.3" y="214" text-anchor="middle">2k</text>
+<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
+<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
+<text class="ts" x="510" y="224" text-anchor="end">step</text>
+<path class="vw v-full" d="M46.0 38.6 L49.1 34.8 L52.2 32.1 L55.3 30.1 L58.4 28.2 L61.5 26.9 L64.6 26.0 L67.7 26.6 L70.7 27.8 L73.8 29.5 L76.9 31.8 L80.0 33.6 L83.1 35.1 L86.2 37.4 L89.3 39.6 L92.4 42.0 L95.5 44.7 L98.6 47.2 L101.7 49.7 L104.8 52.1 L107.9 55.1 L111.0 58.3 L114.1 61.1 L117.1 64.0 L120.2 66.8 L123.3 69.8 L126.4 72.7 L129.5 75.6 L132.6 78.6 L135.7 81.3 L138.8 83.9 L141.9 86.5 L145.0 89.3 L148.1 91.9 L151.2 94.4 L154.3 96.9 L157.4 99.3 L160.5 101.9 L163.5 104.4 L166.6 107.0 L169.7 109.4 L172.8 111.6 L175.9 113.7 L179.0 115.9 L182.1 118.3 L185.2 120.7 L188.3 122.8 L191.4 125.0 L194.5 127.3 L197.6 129.7 L200.7 131.8 L203.8 134.3 L206.9 136.5 L209.9 138.2 L213.0 140.5 L216.1 142.4 L219.2 144.1 L222.3 146.1 L225.4 147.6 L228.5 149.2 L231.6 150.9 L234.7 152.5 L237.8 154.1 L240.9 156.0 L244.0 157.5 L247.1 158.8 L250.2 160.1 L253.3 160.4 L256.3 160.9 L259.4 162.1 L262.5 163.5 L265.6 164.4 L268.7 165.7 L271.8 166.0 L274.9 166.8 L278.0 168.6 L281.1 169.7 L284.2 170.9 L287.3 171.4 L290.4 172.6 L293.5 173.4 L296.6 173.3 L299.7 174.8 L302.7 176.2 L305.8 177.8 L308.9 178.5 L312.0 179.2 L315.1 179.8 L318.2 181.1 L321.3 182.0 L324.4 183.0 L327.5 183.6 L330.6 184.1 L333.7 185.3 L336.8 186.6 L339.9 186.1 L343.0 184.9 L346.1 186.4 L349.1 186.4 L352.2 187.6 L355.3 188.0 L358.4 188.8 L361.5 189.0 L364.6 189.7 L367.7 190.0 L370.8 191.2 L373.9 190.9 L377.0 192.6 L380.1 192.8 L383.2 191.3 L386.3 192.8 L389.4 194.1 L392.5 195.4 L395.5 195.3 L398.6 194.1 L401.7 193.1 L404.8 194.3 L407.9 195.3 L411.0 195.4 L414.1 196.2 L417.2 195.0 L420.3 193.4 L423.4 193.8 L426.5 195.1 L429.6 195.1 L432.7 196.7 L435.8 196.6 L438.9 197.1 L441.9 195.2 L445.0 196.6 L448.1 198.0 L451.2 197.6 L454.3 197.1 L457.4 197.3 L460.5 198.9 L463.6 197.9 L466.7 197.5 L469.8 199.4 L472.9 199.1 L476.0 199.1 L479.1 199.2 L482.2 198.7 L485.3 198.1 L488.3 198.5 L491.4 199.3 L494.5 199.3 L497.6 199.2 L500.7 196.9 L503.8 198.4 L506.9 200.0 L508.5 198.2"></path>
+<circle class="vdot v-full" cx="508.5" cy="198.2" r="3.5"></circle>
+<path class="vw v-noforce" d="M46.0 38.6 L47.5 36.8 L49.1 34.3 L50.6 32.6 L52.2 31.6 L53.7 30.7 L55.3 29.7 L56.8 28.9 L58.4 28.5 L59.9 28.1 L61.5 27.7 L63.0 27.4 L64.6 27.1 L66.1 26.9 L67.7 27.1 L69.2 27.0 L70.7 27.9 L72.3 28.7 L73.8 29.7 L75.4 30.8 L76.9 31.3 L78.5 32.5 L80.0 33.5 L81.6 34.5 L83.1 35.5 L84.7 36.7 L86.2 37.8 L87.8 38.9 L89.3 40.4 L90.9 41.5 L92.4 42.6 L93.9 44.0 L95.5 45.3 L97.0 46.5 L98.6 47.7 L100.1 49.2 L101.7 50.2 L103.2 51.2 L104.8 52.4 L106.3 53.9 L107.9 55.4 L109.4 56.9 L111.0 58.3 L112.5 59.7 L114.1 61.2 L115.6 62.7 L117.1 64.1 L118.7 65.6 L120.2 67.1 L121.8 68.5 L123.3 70.0 L124.9 71.4 L126.4 72.8 L128.0 74.3 L129.5 75.7 L131.1 77.1 L132.6 78.5 L134.2 79.8 L135.7 81.1 L137.3 82.5 L138.8 83.8 L140.3 85.0 L141.9 86.3 L143.4 87.7 L145.0 88.9 L146.5 90.4 L148.1 91.8 L149.6 93.1 L151.2 94.4 L152.7 95.7 L154.3 96.9 L155.8 98.1 L157.4 99.4 L158.9 100.7 L160.5 101.9 L162.0 103.2 L163.5 104.5 L165.1 105.8 L166.6 107.1 L168.2 108.3 L169.7 109.7 L171.3 110.8 L172.8 111.9 L174.4 113.1 L175.9 114.2 L177.5 115.2 L179.0 116.6 L180.6 117.7 L182.1 118.9 L183.7 120.0 L185.2 121.1 L186.7 122.1"></path>
+<circle class="vdot v-noforce" cx="186.7" cy="122.1" r="3.5"></circle>
+</svg></div>
+<div><svg viewBox="0 0 520 230" role="img" aria-label="force flow MSE">
+<text class="th" x="46" y="15">force flow MSE</text>
+<line class="grid-cell" x1="46" y1="121.3" x2="510" y2="121.3"></line>
+<text class="ts" x="40" y="124.3" text-anchor="end">1</text>
+<line class="grid-cell" x1="46.0" y1="26" x2="46.0" y2="200"></line>
+<text class="ts" x="46.0" y="214" text-anchor="middle">0k</text>
+<line class="grid-cell" x1="200.7" y1="26" x2="200.7" y2="200"></line>
+<text class="ts" x="200.7" y="214" text-anchor="middle">1k</text>
+<line class="grid-cell" x1="355.3" y1="26" x2="355.3" y2="200"></line>
+<text class="ts" x="355.3" y="214" text-anchor="middle">2k</text>
+<line class="grid-cell" x1="510.0" y1="26" x2="510.0" y2="200"></line>
+<text class="ts" x="510.0" y="214" text-anchor="middle">3k</text>
+<text class="ts" x="510" y="224" text-anchor="end">step</text>
+<path class="vw v-full" d="M46.0 26.0 L49.1 27.5 L52.2 30.1 L55.3 35.3 L58.4 40.6 L61.5 45.6 L64.6 49.5 L67.7 54.3 L70.7 56.4 L73.8 59.1 L76.9 62.4 L80.0 65.3 L83.1 68.7 L86.2 71.7 L89.3 75.3 L92.4 78.3 L95.5 81.6 L98.6 85.1 L101.7 88.0 L104.8 90.1 L107.9 92.1 L111.0 94.3 L114.1 96.1 L117.1 97.7 L120.2 99.4 L123.3 100.6 L126.4 101.7 L129.5 102.8 L132.6 104.1 L135.7 105.2 L138.8 106.3 L141.9 107.0 L145.0 107.3 L148.1 107.9 L151.2 108.6 L154.3 109.3 L157.4 109.9 L160.5 110.7 L163.5 111.0 L166.6 111.5 L169.7 111.9 L172.8 112.2 L175.9 112.6 L179.0 113.1 L182.1 113.3 L185.2 113.7 L188.3 114.0 L191.4 114.4 L194.5 114.8 L197.6 115.2 L200.7 115.3 L203.8 115.8 L206.9 116.2 L209.9 116.4 L213.0 116.8 L216.1 117.4 L219.2 117.7 L222.3 118.3 L225.4 118.8 L228.5 119.2 L231.6 119.7 L234.7 120.1 L237.8 120.5 L240.9 121.0 L244.0 122.0 L247.1 122.8 L250.2 123.7 L253.3 124.4 L256.3 125.0 L259.4 125.5 L262.5 126.3 L265.6 127.1 L268.7 127.9 L271.8 129.2 L274.9 130.2 L278.0 131.3 L281.1 132.4 L284.2 133.2 L287.3 134.4 L290.4 135.6 L293.5 136.4 L296.6 138.3 L299.7 139.9 L302.7 140.9 L305.8 142.6 L308.9 144.1 L312.0 146.0 L315.1 148.0 L318.2 149.5 L321.3 151.1 L324.4 152.5 L327.5 154.0 L330.6 155.4 L333.7 157.0 L336.8 158.9 L339.9 160.2 L343.0 161.9 L346.1 163.4 L349.1 164.9 L352.2 166.2 L355.3 167.4 L358.4 168.9 L361.5 170.1 L364.6 172.0 L367.7 173.8 L370.8 174.1 L373.9 175.5 L377.0 176.8 L380.1 178.5 L383.2 180.1 L386.3 181.4 L389.4 182.5 L392.5 183.9 L395.5 185.0 L398.6 185.3 L401.7 185.6 L404.8 186.6 L407.9 187.8 L411.0 189.4 L414.1 189.8 L417.2 190.6 L420.3 190.7 L423.4 191.0 L426.5 192.1 L429.6 193.1 L432.7 194.0 L435.8 195.4 L438.9 195.9 L441.9 196.9 L445.0 196.9 L448.1 197.5 L451.2 197.9 L454.3 197.3 L457.4 198.0 L460.5 198.5 L463.6 198.3 L466.7 198.4 L469.8 198.9 L472.9 199.1 L476.0 199.4 L479.1 199.7 L482.2 199.4 L485.3 199.1 L488.3 199.3 L491.4 199.5 L494.5 199.9 L497.6 200.0 L500.7 198.8 L503.8 199.1 L506.9 198.9 L508.5 199.2"></path>
+<circle class="vdot v-full" cx="508.5" cy="199.2" r="3.5"></circle>
+<path class="vw v-notrack" d="M46.0 29.6 L49.1 29.0 L52.2 31.0 L55.3 36.5 L58.4 41.7 L61.5 47.0 L64.6 51.2 L67.7 56.5 L70.7 59.3 L73.8 63.0 L76.9 66.9 L80.0 71.1 L83.1 75.5 L86.2 79.0 L89.3 82.7 L92.4 85.3 L95.5 87.8 L98.6 90.8 L101.7 93.0 L104.8 94.7 L107.9 96.2 L111.0 97.7 L114.1 99.0 L117.1 100.6 L120.2 102.0 L123.3 102.6 L126.4 103.7 L129.5 104.6 L132.6 105.5 L135.7 106.4 L138.8 107.5 L141.9 108.0 L145.0 108.4 L148.1 108.9 L151.2 109.2 L154.3 109.8 L157.4 110.2 L160.5 110.7 L163.5 110.6 L166.6 111.2 L169.7 111.2 L172.8 111.4 L175.9 111.5 L179.0 111.9 L182.1 112.1 L185.2 112.5 L188.3 112.8 L191.4 113.5 L194.5 113.6 L197.6 113.4 L200.7 113.8 L203.8 114.3 L206.9 114.6 L209.9 114.7 L213.0 115.2 L216.1 116.1 L219.2 116.3 L222.3 116.5 L225.4 116.7 L228.5 117.0 L231.6 117.6 L234.7 117.3 L237.8 117.7 L240.9 118.0 L244.0 118.5 L247.1 118.7 L250.2 119.2 L253.3 119.7 L256.3 119.9 L259.4 120.3 L262.5 120.8 L265.6 121.5 L268.7 122.0 L271.8 122.6 L274.9 122.6 L278.0 123.2 L281.1 124.1 L284.2 124.4 L287.3 124.8 L290.4 125.3 L293.5 125.6 L296.6 126.3 L299.7 127.0 L302.7 127.7 L305.8 128.3 L308.9 128.9 L312.0 129.3 L315.1 129.9 L318.2 130.8 L321.3 131.7 L324.4 132.5 L327.5 133.4 L330.6 134.1 L333.7 135.1 L336.8 136.2 L339.9 137.1 L343.0 138.0 L346.1 138.6 L349.1 139.9 L352.2 140.4 L355.3 141.7 L358.4 142.7 L361.5 143.0 L364.6 143.8 L367.7 144.9 L370.8 145.6 L373.9 146.6 L377.0 147.9 L380.1 149.3 L383.2 149.9 L386.3 151.2 L389.4 152.1 L392.5 152.6 L395.5 153.1 L398.6 153.6 L401.7 154.3 L404.8 155.5 L407.9 156.2 L411.0 156.8 L414.1 157.5 L417.2 158.1 L420.3 158.7 L423.4 159.4 L426.5 159.7 L429.6 161.0 L432.7 161.5 L435.8 161.5 L438.9 161.7 L441.9 162.3 L445.0 162.5 L448.1 163.3 L451.2 164.2 L454.3 164.4 L457.4 164.8 L460.5 164.9 L463.6 165.2 L466.7 166.3 L469.8 167.1 L472.9 167.1 L476.0 167.7 L479.1 167.7 L482.2 168.1 L485.3 167.6 L488.3 167.1 L491.4 167.0 L494.5 166.9 L497.6 167.1 L500.7 167.3 L503.8 167.3 L506.9 167.5 L508.5 167.7"></path>
+<circle class="vdot v-notrack" cx="508.5" cy="167.7" r="3.5"></circle>
+</svg></div>
+    </div>
+    <figcaption><b>Figure A1 &mdash; Same-step training loss convergence.</b>
+    Identical data, LoRA budget, seed, and 3,000 single-GPU steps for all four
+    variants; curves are EMA-smoothed (half-life 80 steps), log-scaled y.
+    Track loss exists only where track tokens exist; force loss only where the
+    force stage exists.</figcaption>
+  </figure>
+
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>Variant (mean loss, steps 2,800&ndash;3,000)</th><th>action</th>
+        <th>dynamics</th><th>track</th><th>force</th></tr></thead>
+        <tbody><tr><td><span class="chip"><span class="vs vs-full"></span>&nbsp;A · full (track+force)</span></td><td class="shape">0.2819</td><td class="shape">0.3325</td><td class="shape">0.1501</td><td class="shape">0.6187</td></tr>
+<tr><td><span class="chip"><span class="vs vs-notrack"></span>&nbsp;B · no track</span></td><td class="shape">0.2749</td><td class="shape">0.2837</td><td class="shape">&mdash;</td><td class="shape">0.7535</td></tr>
+<tr><td><span class="chip"><span class="vs vs-noforce"></span>&nbsp;C · no force stage</span></td><td class="shape">&mdash;</td><td class="shape">&mdash;</td><td class="shape">&mdash;</td><td class="shape">&mdash;</td></tr></tbody>
+      </table>
+    </div>
+    <h3>Controlled evaluation on frozen checkpoints</h3>
+    <div class="prose"><p>All numbers below come from the same fixed
+      anchor sets, noise seeds, and prompts for every variant
+      (<code>scripts/eval/trex_ablation_eval.py</code>). Open-loop: each variant
+      runs its own full sampling loop from pure noise on 16 anchors and is scored
+      against the ground-truth normalized delta-base chunk.</p></div>
+  <figure><svg viewBox="0 0 560 138" role="img" aria-label="open-loop action MSE ↓ (16 anchors, normalized 62-D)">
+<text class="th" x="0" y="14">open-loop action MSE ↓ (16 anchors, normalized 62-D)</text>
+<text class="t" x="232" y="39" text-anchor="end">A · cascade (+force stage)</text>
+<rect class="v-full" x="240" y="26" width="133.1" height="16" rx="3"></rect>
+<text class="ts" x="379.1" y="39">0.205</text>
+<text class="t" x="232" y="67" text-anchor="end">A · coarse only</text>
+<rect class="v-full vfaded" x="240" y="54" width="250.0" height="16" rx="3"></rect>
+<text class="ts" x="496.0" y="67">0.386</text>
+<text class="t" x="232" y="95" text-anchor="end">B · cascade (+force stage)</text>
+<rect class="v-notrack" x="240" y="82" width="149.0" height="16" rx="3"></rect>
+<text class="ts" x="395.0" y="95">0.230</text>
+<text class="t" x="232" y="123" text-anchor="end">B · coarse only</text>
+<rect class="v-notrack vfaded" x="240" y="110" width="248.4" height="16" rx="3"></rect>
+<text class="ts" x="494.4" y="123">0.383</text>
+</svg><figcaption><b>Figure A2 &mdash; Open-loop reconstruction.</b> Faded bars: the coarse action read out at &tau;=0 without tactile refinement (for C this is its only mode). Solid bars: the full cascade.</figcaption></figure>
+
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>flow MSE @ fixed &tau;</th><th>&tau;=0.8</th><th>&tau;=0.6</th><th>&tau;=0.4</th><th>&tau;=0.2</th><th>&tau;=0.05</th></tr></thead>
+        <tbody><tr><td><span class="chip"><span class="vs vs-full"></span>&nbsp;A &middot; action</span></td><td class="shape">0.139</td><td class="shape">0.204</td><td class="shape">0.399</td><td class="shape">0.752</td><td class="shape">1.076</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span>&nbsp;B &middot; action</span></td><td class="shape">0.124</td><td class="shape">0.191</td><td class="shape">0.388</td><td class="shape">0.745</td><td class="shape">1.078</td></tr><tr><td><span class="chip"><span class="vs vs-full"></span>&nbsp;A &middot; dynamics</span></td><td class="shape">0.215</td><td class="shape">0.280</td><td class="shape">0.414</td><td class="shape">0.724</td><td class="shape">1.131</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span>&nbsp;B &middot; dynamics</span></td><td class="shape">0.178</td><td class="shape">0.229</td><td class="shape">0.341</td><td class="shape">0.628</td><td class="shape">1.064</td></tr></tbody>
+      </table>
+    </div>
+  <figure><svg viewBox="0 0 640 146" role="img" aria-label="Stage-1: where action queries attend (mass share)">
+<text class="th" x="0" y="14">Stage-1: where action queries attend (mass share)</text>
+<text class="t" x="102" y="39" text-anchor="end">A · layer 0</text>
+<rect class="sg-cond" x="110.0" y="26" width="12.6" height="18" rx="3"></rect>
+<rect class="sg-obs" x="124.6" y="26" width="53.4" height="18" rx="3"></rect>
+<text class="ts" x="152.3" y="39" text-anchor="middle">11%</text>
+<rect class="sg-act" x="180.0" y="26" width="29.9" height="18" rx="3"></rect>
+<rect class="sg-st" x="211.9" y="26" width="0.5" height="18" rx="3"></rect>
+<rect class="sg-trkp" x="212.7" y="26" width="233.6" height="18" rx="3"></rect>
+<text class="ts" x="330.5" y="39" text-anchor="middle">45%</text>
+<rect class="sg-trkf" x="448.3" y="26" width="177.7" height="18" rx="3"></rect>
+<text class="ts" x="538.2" y="39" text-anchor="middle">35%</text>
+<text class="t" x="102" y="67" text-anchor="end">A · layer 29</text>
+<rect class="sg-cond" x="110.0" y="54" width="25.6" height="18" rx="3"></rect>
+<rect class="sg-obs" x="137.6" y="54" width="49.5" height="18" rx="3"></rect>
+<text class="ts" x="163.4" y="67" text-anchor="middle">10%</text>
+<rect class="sg-act" x="189.1" y="54" width="95.2" height="18" rx="3"></rect>
+<text class="ts" x="237.7" y="67" text-anchor="middle">19%</text>
+<rect class="sg-st" x="286.3" y="54" width="0.5" height="18" rx="3"></rect>
+<rect class="sg-trkp" x="287.6" y="54" width="296.0" height="18" rx="3"></rect>
+<text class="ts" x="436.6" y="67" text-anchor="middle">58%</text>
+<rect class="sg-trkf" x="585.6" y="54" width="40.4" height="18" rx="3"></rect>
+<text class="ts" x="606.8" y="67" text-anchor="middle">8%</text>
+<text class="t" x="102" y="95" text-anchor="end">B · layer 0</text>
+<rect class="sg-cond" x="110.0" y="82" width="106.6" height="18" rx="3"></rect>
+<text class="ts" x="164.3" y="95" text-anchor="middle">21%</text>
+<rect class="sg-obs" x="218.6" y="82" width="260.7" height="18" rx="3"></rect>
+<text class="ts" x="349.9" y="95" text-anchor="middle">51%</text>
+<rect class="sg-act" x="481.3" y="82" width="138.5" height="18" rx="3"></rect>
+<text class="ts" x="551.5" y="95" text-anchor="middle">27%</text>
+<rect class="sg-st" x="621.8" y="82" width="4.2" height="18" rx="3"></rect>
+<text class="t" x="102" y="123" text-anchor="end">B · layer 29</text>
+<rect class="sg-cond" x="110.0" y="110" width="51.4" height="18" rx="3"></rect>
+<text class="ts" x="136.7" y="123" text-anchor="middle">10%</text>
+<rect class="sg-obs" x="163.4" y="110" width="124.6" height="18" rx="3"></rect>
+<text class="ts" x="226.7" y="123" text-anchor="middle">24%</text>
+<rect class="sg-act" x="290.0" y="110" width="331.7" height="18" rx="3"></rect>
+<text class="ts" x="456.9" y="123" text-anchor="middle">64%</text>
+<rect class="sg-st" x="623.7" y="110" width="2.3" height="18" rx="3"></rect>
+</svg><figcaption><b>Figure A3 &mdash; Stage-1 attention routing.</b> Share of self-attention mass from action queries onto each token group at &tau;=0.4 (mean over heads/blocks).</figcaption></figure>
+  <figure><svg viewBox="0 0 640 90" role="img" aria-label="Stage-2: where action queries attend (mass share)">
+<text class="th" x="0" y="14">Stage-2: where action queries attend (mass share)</text>
+<text class="t" x="102" y="39" text-anchor="end">A · L0–L5 mean</text>
+<rect class="sg-act" x="110.0" y="26" width="12.3" height="18" rx="3"></rect>
+<rect class="sg-f6" x="124.3" y="26" width="6.9" height="18" rx="3"></rect>
+<rect class="sg-vq" x="133.2" y="26" width="8.4" height="18" rx="3"></rect>
+<rect class="sg-mem" x="143.6" y="26" width="482.4" height="18" rx="3"></rect>
+<text class="ts" x="385.8" y="39" text-anchor="middle">94%</text>
+<text class="t" x="102" y="67" text-anchor="end">B · L0–L5 mean</text>
+<rect class="sg-act" x="110.0" y="54" width="105.4" height="18" rx="3"></rect>
+<text class="ts" x="163.7" y="67" text-anchor="middle">21%</text>
+<rect class="sg-f6" x="217.4" y="54" width="49.2" height="18" rx="3"></rect>
+<text class="ts" x="243.0" y="67" text-anchor="middle">10%</text>
+<rect class="sg-vq" x="268.6" y="54" width="51.1" height="18" rx="3"></rect>
+<text class="ts" x="295.1" y="67" text-anchor="middle">10%</text>
+<rect class="sg-mem" x="321.7" y="54" width="304.3" height="18" rx="3"></rect>
+<text class="ts" x="474.8" y="67" text-anchor="middle">59%</text>
+</svg><figcaption><b>Figure A4 &mdash; Stage-2 attention routing.</b> Action-query mass over F6 / VQ-history / deform / coarse-memory tokens during real cascaded refinement.</figcaption></figure>
+
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>Tactile sensitivity (normalized units)</th>
+        <th>mean |&Delta;a| tactile vs masked</th><th>max |&Delta;a|</th>
+        <th>mean |refinement|</th></tr></thead>
+        <tbody><tr><td><span class="chip"><span class="vs vs-full"></span>&nbsp;A · full (track+force)</span></td><td class="shape">0.0141</td><td class="shape">0.0664</td><td class="shape">0.2452</td></tr><tr><td><span class="chip"><span class="vs vs-notrack"></span>&nbsp;B · no track</span></td><td class="shape">0.0256</td><td class="shape">0.1309</td><td class="shape">0.2472</td></tr></tbody>
+      </table>
+    </div>
+
+  </section>

ablation_runs/eval_part1.json

@@ -0,0 +1,432 @@
+{
+ "full": {
+  "checkpoint": "checkpoints/ablate_full_3k/checkpoint-3000",
+  "use_track": true,
+  "use_force": true,
+  "use_deform_tactile": false,
+  "forward_tau_losses": {
+   "0.8": {
+    "action_loss": [
+     0.13885498046875,
+     0.021893762307867347
+    ],
+    "dynamics_loss": [
+     0.214599609375,
+     0.028883325232585192
+    ],
+    "track_flow_loss": [
+     0.01959228515625,
+     0.005815500664137739
+    ],
+    "force_loss": [
+     0.6427408854166666,
+     0.07174448889029805
+    ]
+   },
+   "0.6": {
+    "action_loss": [
+     0.20402018229166666,
+     0.027474107787100906
+    ],
+    "dynamics_loss": [
+     0.2801106770833333,
+     0.02749254221890766
+    ],
+    "track_flow_loss": [
+     0.025583902994791668,
+     0.009258714454018649
+    ],
+    "force_loss": [
+     0.6427408854166666,
+     0.07174448889029805
+    ]
+   },
+   "0.4": {
+    "action_loss": [
+     0.3986002604166667,
+     0.0390743661447396
+    ],
+    "dynamics_loss": [
+     0.4138997395833333,
+     0.025458871788911118
+    ],
+    "track_flow_loss": [
+     0.051778157552083336,
+     0.017947368985419505
+    ],
+    "force_loss": [
+     0.6427408854166666,
+     0.07174448889029805
+    ]
+   },
+   "0.2": {
+    "action_loss": [
+     0.7521158854166666,
+     0.048475014437045466
+    ],
+    "dynamics_loss": [
+     0.7236328125,
+     0.020585171554378095
+    ],
+    "track_flow_loss": [
+     0.270751953125,
+     0.02853870589963502
+    ],
+    "force_loss": [
+     0.6427408854166666,
+     0.07174448889029805
+    ]
+   },
+   "0.05": {
+    "action_loss": [
+     1.0758463541666667,
+     0.05334471104376999
+    ],
+    "dynamics_loss": [
+     1.1305338541666667,
+     0.01504103522415094
+    ],
+    "track_flow_loss": [
+     0.912109375,
+     0.026373214974647258
+    ],
+    "force_loss": [
+     0.6427408854166666,
+     0.07174448889029805
+    ]
+   }
+  },
+  "openloop_action_mse": {
+   "cascade": [
+    0.20543859526515007,
+    0.10553651615425272,
+    16
+   ],
+   "coarse_only": [
+    0.3859300408512354,
+    0.09367721956997326,
+    16
+   ]
+  },
+  "openloop_action_mse_split": {
+   "cascade_arm": 0.29095884039998055,
+   "cascade_hand": 0.1704530455172062,
+   "coarse_arm": 0.42187667451798916,
+   "coarse_hand": 0.3712246026843786
+  },
+  "stage1_attention": {
+   "0": {
+    "action": {
+     "cond_obs": 0.028233155608177185,
+     "obs": 0.10685919225215912,
+     "action": 0.06159074231982231,
+     "state": 0.001643829746171832,
+     "track_past": 0.4548226296901703,
+     "track_future": 0.3468504846096039
+    },
+    "obs": {
+     "cond_obs": 0.053328514099121094,
+     "obs": 0.5826746821403503,
+     "action": 0.009129117242991924,
+     "state": 0.0005130452918820083,
+     "track_past": 0.1784334033727646,
+     "track_future": 0.1759212762117386
+    }
+   },
+   "14": {
+    "action": {
+     "cond_obs": 0.04729510098695755,
+     "obs": 0.08689284324645996,
+     "action": 0.20308324694633484,
+     "state": 0.0032673801761120558,
+     "track_past": 0.6379822492599487,
+     "track_future": 0.021479224786162376
+    },
+    "obs": {
+     "cond_obs": 0.16180963814258575,
+     "obs": 0.6352365016937256,
+     "action": 0.013975701294839382,
+     "state": 0.006121945567429066,
+     "track_past": 0.167484849691391,
+     "track_future": 0.015371335670351982
+    }
+   },
+   "29": {
+    "action": {
+     "cond_obs": 0.05335542559623718,
+     "obs": 0.09940896928310394,
+     "action": 0.18759986758232117,
+     "state": 0.0025236550718545914,
+     "track_past": 0.5753425359725952,
+     "track_future": 0.08176963031291962
+    },
+    "obs": {
+     "cond_obs": 0.09543927758932114,
+     "obs": 0.7705966234207153,
+     "action": 0.01101355254650116,
+     "state": 0.0004360256134532392,
+     "track_past": 0.07644421607255936,
+     "track_future": 0.046070314943790436
+    }
+   }
+  },
+  "stage2_attention": [
+   {
+    "action": 0.033935546875,
+    "f6": 0.0230712890625,
+    "vq": 0.024169921875,
+    "memory": 0.91796875
+   },
+   {
+    "action": 0.0322265625,
+    "f6": 0.0208740234375,
+    "vq": 0.023681640625,
+    "memory": 0.921875
+   },
+   {
+    "action": 0.0274658203125,
+    "f6": 0.01708984375,
+    "vq": 0.0203857421875,
+    "memory": 0.93359375
+   },
+   {
+    "action": 0.0250244140625,
+    "f6": 0.01513671875,
+    "vq": 0.0185546875,
+    "memory": 0.94140625
+   },
+   {
+    "action": 0.024169921875,
+    "f6": 0.01416015625,
+    "vq": 0.017333984375,
+    "memory": 0.9453125
+   },
+   {
+    "action": 0.0228271484375,
+    "f6": 0.0130615234375,
+    "vq": 0.0157470703125,
+    "memory": 0.94921875
+   }
+  ],
+  "tactile_sensitivity": {
+   "mean_abs_delta_vs_masked": 0.014084354043006897,
+   "max_abs_delta_vs_masked": 0.06640625,
+   "mean_abs_refinement": 0.2452070713043213
+  }
+ },
+ "notrack": {
+  "checkpoint": "checkpoints/ablate_notrack_3k/checkpoint-3000",
+  "use_track": false,
+  "use_force": true,
+  "use_deform_tactile": false,
+  "forward_tau_losses": {
+   "0.8": {
+    "action_loss": [
+     0.12394205729166667,
+     0.02089633950925863
+    ],
+    "dynamics_loss": [
+     0.178466796875,
+     0.028673447679287072
+    ],
+    "track_flow_loss": [
+     0.0,
+     0.0
+    ],
+    "force_loss": [
+     0.7783203125,
+     0.043793139306208324
+    ]
+   },
+   "0.6": {
+    "action_loss": [
+     0.19099934895833334,
+     0.028310372080589903
+    ],
+    "dynamics_loss": [
+     0.22908528645833334,
+     0.028607776677988368
+    ],
+    "track_flow_loss": [
+     0.0,
+     0.0
+    ],
+    "force_loss": [
+     0.7783203125,
+     0.043793139306208324
+    ]
+   },
+   "0.4": {
+    "action_loss": [
+     0.3876953125,
+     0.035242046640337726
+    ],
+    "dynamics_loss": [
+     0.3413899739583333,
+     0.029684818436720592
+    ],
+    "track_flow_loss": [
+     0.0,
+     0.0
+    ],
+    "force_loss": [
+     0.7783203125,
+     0.043793139306208324
+    ]
+   },
+   "0.2": {
+    "action_loss": [
+     0.7447916666666666,
+     0.0426817276121973
+    ],
+    "dynamics_loss": [
+     0.6276041666666666,
+     0.024927330363381052
+    ],
+    "track_flow_loss": [
+     0.0,
+     0.0
+    ],
+    "force_loss": [
+     0.7783203125,
+     0.043793139306208324
+    ]
+   },
+   "0.05": {
+    "action_loss": [
+     1.0777994791666667,
+     0.054336633312902415
+    ],
+    "dynamics_loss": [
+     1.0641276041666667,
+     0.01625975747883096
+    ],
+    "track_flow_loss": [
+     0.0,
+     0.0
+    ],
+    "force_loss": [
+     0.7783203125,
+     0.043793139306208324
+    ]
+   }
+  },
+  "openloop_action_mse": {
+   "cascade": [
+    0.22997993603348732,
+    0.1032914996028219,
+    16
+   ],
+   "coarse_only": [
+    0.38339873775839806,
+    0.09786209841435693,
+    16
+   ]
+  },
+  "openloop_action_mse_split": {
+   "cascade_arm": 0.29843927826732397,
+   "cascade_hand": 0.2019738396629691,
+   "coarse_arm": 0.4064813628792763,
+   "coarse_hand": 0.373955849558115
+  },
+  "stage1_attention": {
+   "0": {
+    "action": {
+     "cond_obs": 0.20962195098400116,
+     "obs": 0.507138729095459,
+     "action": 0.27125489711761475,
+     "state": 0.011984430253505707,
+     "track_past": 0.0,
+     "track_future": 0.0
+    },
+    "obs": {
+     "cond_obs": 0.16306498646736145,
+     "obs": 0.7571282386779785,
+     "action": 0.07827480137348175,
+     "state": 0.001532059395685792,
+     "track_past": 0.0,
+     "track_future": 0.0
+    }
+   },
+   "14": {
+    "action": {
+     "cond_obs": 0.09229698032140732,
+     "obs": 0.18946343660354614,
+     "action": 0.6833343505859375,
+     "state": 0.03490526229143143,
+     "track_past": 0.0,
+     "track_future": 0.0
+    },
+    "obs": {
+     "cond_obs": 0.26676031947135925,
+     "obs": 0.6440349221229553,
+     "action": 0.054011691361665726,
+     "state": 0.03519308567047119,
+     "track_past": 0.0,
+     "track_future": 0.0
+    }
+   },
+   "29": {
+    "action": {
+     "cond_obs": 0.10308176279067993,
+     "obs": 0.24439243972301483,
+     "action": 0.6442670822143555,
+     "state": 0.008258762769401073,
+     "track_past": 0.0,
+     "track_future": 0.0
+    },
+    "obs": {
+     "cond_obs": 0.15114127099514008,
+     "obs": 0.8275548219680786,
+     "action": 0.019980238750576973,
+     "state": 0.0013237263774499297,
+     "track_past": 0.0,
+     "track_future": 0.0
+    }
+   }
+  },
+  "stage2_attention": [
+   {
+    "action": 0.1103515625,
+    "f6": 0.0625,
+    "vq": 0.06640625,
+    "memory": 0.76171875
+   },
+   {
+    "action": 0.1904296875,
+    "f6": 0.09814453125,
+    "vq": 0.0966796875,
+    "memory": 0.61328125
+   },
+   {
+    "action": 0.3359375,
+    "f6": 0.146484375,
+    "vq": 0.1513671875,
+    "memory": 0.365234375
+   },
+   {
+    "action": 0.283203125,
+    "f6": 0.12353515625,
+    "vq": 0.1259765625,
+    "memory": 0.466796875
+   },
+   {
+    "action": 0.193359375,
+    "f6": 0.08984375,
+    "vq": 0.09716796875,
+    "memory": 0.62109375
+   },
+   {
+    "action": 0.1298828125,
+    "f6": 0.072265625,
+    "vq": 0.0771484375,
+    "memory": 0.71875
+   }
+  ],
+  "tactile_sensitivity": {
+   "mean_abs_delta_vs_masked": 0.025584066286683083,
+   "max_abs_delta_vs_masked": 0.130859375,
+   "mean_abs_refinement": 0.24716320633888245
+  }
+ }
+}

ablation_runs/tools/make_ablation_section.py

@@ -0,0 +1,332 @@
+"""Assemble the ablation section HTML for the Track-Force Cascade artifact."""
+import json
+import sys
+from pathlib import Path
+
+TOOLS = Path(__file__).parent
+sys.path.insert(0, str(TOOLS))
+from chartlib import line_chart, stacked_hbar, hbar, VARIANTS, VAR_LABEL  # noqa: E402
+from loss_series import series_for, load_components  # noqa: E402
+
+RUNS = {v: f"/scratch1/home/zhicao/dreamzero/checkpoints/ablate_{v}_3k/loss_log.jsonl"
+        for v in VARIANTS}
+
+EXTRA_CSS = """
+  /* ---- ablation additions ---- */
+  :root {
+    --v-full: #3c7ec0;
+    --v-notrack: #c26a1e;
+    --v-noforce: #6f66bd;
+    --v-deform: #b54e7f;
+  }
+  @media (prefers-color-scheme: dark) {
+    :root:not([data-theme="light"]) {
+      --v-full: #5895d1;
+      --v-notrack: #c98136;
+      --v-noforce: #8d7fd6;
+      --v-deform: #d76a94;
+    }
+  }
+  :root[data-theme="dark"] {
+    --v-full: #5895d1;
+    --v-notrack: #c98136;
+    --v-noforce: #8d7fd6;
+    --v-deform: #d76a94;
+  }
+  .vw { fill: none; stroke-width: 2; stroke-linejoin: round; stroke-linecap: round; }
+  .vdot { stroke: var(--surface); stroke-width: 2; }
+  .vw.v-full { stroke: var(--v-full); }
+  .vw.v-notrack { stroke: var(--v-notrack); }
+  .vw.v-noforce { stroke: var(--v-noforce); }
+  .vw.v-deform { stroke: var(--v-deform); }
+  .vdot.v-full, rect.v-full { fill: var(--v-full); }
+  .vdot.v-notrack, rect.v-notrack { fill: var(--v-notrack); }
+  .vdot.v-noforce, rect.v-noforce { fill: var(--v-noforce); }
+  .vdot.v-deform, rect.v-deform { fill: var(--v-deform); }
+  .vfaded { opacity: 0.45; }
+  .vs { display: inline-block; width: 13px; height: 3px; border-radius: 2px; }
+  .vs-full { background: var(--v-full); }
+  .vs-notrack { background: var(--v-notrack); }
+  .vs-noforce { background: var(--v-noforce); }
+  .vs-deform { background: var(--v-deform); }
+  .sg-cond { fill: var(--c-obs); opacity: 0.45; }
+  .sg-obs { fill: var(--c-obs); }
+  .sg-act { fill: var(--c-act); }
+  .sg-st { fill: var(--c-st); }
+  .sg-trkp { fill: var(--c-trk); }
+  .sg-trkf { fill: var(--c-trk); opacity: 0.45; }
+  .sg-f6 { fill: var(--c-frc); }
+  .sg-vq { fill: var(--c-frc); opacity: 0.45; }
+  .sg-deform { fill: var(--c-frc); opacity: 0.7; }
+  .sg-mem { fill: var(--c-mem); }
+  .sw-solid-trk { background: var(--c-trk); border-color: var(--c-trk); }
+  .quad { display: grid; grid-template-columns: 1fr 1fr; gap: 18px; }
+  @media (max-width: 900px) { .quad { grid-template-columns: 1fr; } }
+"""
+
+STAGE1_GROUPS = ["cond_obs", "obs", "action", "state", "track_past", "track_future"]
+STAGE1_CLASS = {"cond_obs": "sg-cond", "obs": "sg-obs", "action": "sg-act",
+                "state": "sg-st", "track_past": "sg-trkp", "track_future": "sg-trkf"}
+STAGE2_GROUPS = ["action", "f6", "vq", "deform", "memory"]
+STAGE2_CLASS = {"action": "sg-act", "f6": "sg-f6", "vq": "sg-vq",
+                "deform": "sg-deform", "memory": "sg-mem"}
+SHORT = {"full": "A", "notrack": "B", "noforce": "C", "deform": "D"}
+
+VARIANT_LEGEND = """
+  <div class="legend" role="list" aria-label="Ablation variant legend">
+    <span class="k" role="listitem"><span class="vs vs-full"></span>A &middot; full (track + force)</span>
+    <span class="k" role="listitem"><span class="vs vs-notrack"></span>B &middot; no track (pre-track structure)</span>
+    <span class="k" role="listitem"><span class="vs vs-noforce"></span>C &middot; no force stage</span>
+    <span class="k" role="listitem"><span class="vs vs-deform"></span>D &middot; deform tactile + force</span>
+  </div>
+"""
+
+PROTOCOL_PROSE = """
+    <div class="prose">
+      <p>Four variants of the cascade are trained from the same Wan2.2 foundation
+      weights on the <code>trex_mini_force</code> episode &mdash; identical anchors,
+      LoRA budget (r=4 on q,k,v,o,k_img,v_img,ffn), learning rate, batch size,
+      seed, and 3,000 single-GPU steps &mdash; so every difference below is
+      attributable to the ablated component, not to training budget.
+      <b>B</b> removes the track pathway structurally (no track tokens, encoder,
+      decoder, or loss): the packed block shrinks from 617 to 117 tokens &mdash;
+      exactly the pre-track architecture with the force stage kept. <b>C</b>
+      removes the force stage (no Stage-2 transformer, no tactile input, no
+      force/VQ loss); at inference its coarse model carries the flow all the way
+      to &tau;=0. <b>D</b> keeps the full model and adds ten per-finger
+      deformation-camera tokens (96&times;96 crops &rarr; 4-layer CNN &rarr; d 768)
+      beside the F6 and VQ-history tokens in Stage 2.</p>
+    </div>"""
+
+VARIANT_TABLE = """
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>Variant</th><th>Stage-1 tokens / block</th>
+        <th>Stage-2 tokens</th><th>Losses optimized</th></tr></thead>
+        <tbody>
+          <tr><td><span class="chip"><span class="vs vs-full"></span>&nbsp;A &middot; full</span></td>
+            <td class="shape">100 obs + 16 act + 1 st + 500 trk = 617</td>
+            <td class="shape">16 act + 10 F6 + 10 VQ + mem</td>
+            <td>dynamics &middot; action &middot; 20&middot;track &middot; force &middot; VQ</td></tr>
+          <tr><td><span class="chip"><span class="vs vs-notrack"></span>&nbsp;B &middot; no track</span></td>
+            <td class="shape">100 obs + 16 act + 1 st = 117</td>
+            <td class="shape">16 act + 10 F6 + 10 VQ + mem</td>
+            <td>dynamics &middot; action &middot; force &middot; VQ</td></tr>
+          <tr><td><span class="chip"><span class="vs vs-noforce"></span>&nbsp;C &middot; no force</span></td>
+            <td class="shape">617</td>
+            <td class="shape">&mdash; (coarse runs &tau;&rarr;0)</td>
+            <td>dynamics &middot; action &middot; 20&middot;track</td></tr>
+          <tr><td><span class="chip"><span class="vs vs-deform"></span>&nbsp;D &middot; deform + force</span></td>
+            <td class="shape">617</td>
+            <td class="shape">16 act + 10 F6 + 10 VQ + <b>10 deform</b> + mem</td>
+            <td>dynamics &middot; action &middot; 20&middot;track &middot; force &middot; VQ</td></tr>
+        </tbody>
+      </table>
+    </div>"""
+
+
+def loss_figure():
+    charts = {}
+    for variant, path in RUNS.items():
+        if Path(path).exists():
+            charts[variant] = series_for(
+                path, ["action_loss", "dynamics_loss", "track_loss", "force_loss"]
+            )
+    panels = []
+    for component, title in (
+        ("action_loss", "action flow MSE"),
+        ("dynamics_loss", "dynamics (video flow) MSE"),
+        ("track_loss", "track loss (weighted sum)"),
+        ("force_loss", "force flow MSE"),
+    ):
+        series = {
+            v: charts[v][component]
+            for v in VARIANTS
+            if v in charts and component in charts[v]
+            and max(x[1] for x in charts[v][component]) > 1e-8
+        }
+        if series:
+            panels.append(line_chart(series, title=title, width=520, height=230))
+    grid = "\n".join(f"<div>{p}</div>" for p in panels)
+    return f"""
+  <figure>
+    <div class="quad">
+{grid}
+    </div>
+    <figcaption><b>Figure A1 &mdash; Same-step training loss convergence.</b>
+    Identical data, LoRA budget, seed, and 3,000 single-GPU steps for all four
+    variants; curves are EMA-smoothed (half-life 80 steps), log-scaled y.
+    Track loss exists only where track tokens exist; force loss only where the
+    force stage exists.</figcaption>
+  </figure>"""
+
+
+def final_loss_table(last_n=200):
+    rows = []
+    for variant in VARIANTS:
+        path = RUNS[variant]
+        if not Path(path).exists():
+            continue
+        raw = load_components(path)
+        cells = {}
+        for component in ("action_loss", "dynamics_loss", "track_loss", "force_loss"):
+            pts = raw.get(component, [])
+            tail = [v for s, v in pts if s >= 3000 - last_n]
+            cells[component] = (sum(tail) / len(tail)) if tail else None
+        rows.append((variant, cells))
+
+    def fmt(v):
+        return "&mdash;" if v is None or v < 1e-9 else f"{v:.4f}"
+
+    body = "\n".join(
+        f'<tr><td><span class="chip"><span class="vs vs-{v}"></span>&nbsp;{VAR_LABEL[v]}</span></td>'
+        + "".join(
+            f'<td class="shape">{fmt(c[k])}</td>'
+            for k in ("action_loss", "dynamics_loss", "track_loss", "force_loss")
+        )
+        + "</tr>"
+        for v, c in rows
+    )
+    return f"""
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>Variant (mean loss, steps 2,800&ndash;3,000)</th><th>action</th>
+        <th>dynamics</th><th>track</th><th>force</th></tr></thead>
+        <tbody>{body}</tbody>
+      </table>
+    </div>"""
+
+
+def tau_table(eval_data):
+    taus = ["0.8", "0.6", "0.4", "0.2", "0.05"]
+    head = "".join(f"<th>&tau;={t}</th>" for t in taus)
+    body_rows = []
+    for metric, label in (("action_loss", "action"), ("dynamics_loss", "dynamics")):
+        for variant in VARIANTS:
+            r = eval_data.get(variant)
+            if not r:
+                continue
+            cells = "".join(
+                f'<td class="shape">{r["forward_tau_losses"][t][metric][0]:.3f}</td>'
+                for t in taus
+            )
+            body_rows.append(
+                f'<tr><td><span class="chip"><span class="vs vs-{variant}"></span>'
+                f"&nbsp;{SHORT[variant]} &middot; {label}</span></td>{cells}</tr>"
+            )
+    return f"""
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>flow MSE @ fixed &tau;</th>{head}</tr></thead>
+        <tbody>{"".join(body_rows)}</tbody>
+      </table>
+    </div>"""
+
+
+def openloop_figure(eval_data):
+    rows = []
+    for variant in VARIANTS:
+        r = eval_data.get(variant)
+        if not r:
+            continue
+        ol = r["openloop_action_mse"]
+        if "cascade" in ol:
+            rows.append((f"{SHORT[variant]} · cascade (+force stage)",
+                         ol["cascade"][0], f"v-{variant}"))
+        rows.append((f"{SHORT[variant]} · coarse only",
+                     ol["coarse_only"][0], f"v-{variant} vfaded"))
+    return hbar(rows, width=560, left=240, title="open-loop action MSE ↓ (16 anchors, normalized 62-D)",
+                fmt=lambda v: f"{v:.3f}")
+
+
+def stage1_attention_figure(eval_data, layers=("0", "29")):
+    rows = []
+    for variant in VARIANTS:
+        r = eval_data.get(variant)
+        if not r or "stage1_attention" not in r:
+            continue
+        for layer in layers:
+            masses = r["stage1_attention"].get(layer, {}).get("action")
+            if masses:
+                rows.append((f"{SHORT[variant]} · layer {layer}", masses))
+    return stacked_hbar(rows, STAGE1_GROUPS, STAGE1_CLASS, width=640, left=110,
+                        title="Stage-1: where action queries attend (mass share)")
+
+
+def stage2_attention_figure(eval_data):
+    rows = []
+    for variant in VARIANTS:
+        r = eval_data.get(variant)
+        if not r or not r.get("stage2_attention"):
+            continue
+        layers = r["stage2_attention"]
+        mean = {}
+        for key in STAGE2_GROUPS:
+            vals = [layer.get(key, 0.0) for layer in layers]
+            mean[key] = sum(vals) / len(vals)
+        rows.append((f"{SHORT[variant]} · L0–L5 mean", mean))
+    return stacked_hbar(rows, STAGE2_GROUPS, STAGE2_CLASS, width=640, left=110,
+                        title="Stage-2: where action queries attend (mass share)")
+
+
+def sensitivity_table(eval_data):
+    rows = []
+    for variant in VARIANTS:
+        r = eval_data.get(variant)
+        if not r or "tactile_sensitivity" not in r:
+            continue
+        t = r["tactile_sensitivity"]
+        rows.append(
+            f'<tr><td><span class="chip"><span class="vs vs-{variant}"></span>'
+            f'&nbsp;{VAR_LABEL[variant]}</span></td>'
+            f'<td class="shape">{t["mean_abs_delta_vs_masked"]:.4f}</td>'
+            f'<td class="shape">{t["max_abs_delta_vs_masked"]:.4f}</td>'
+            f'<td class="shape">{t["mean_abs_refinement"]:.4f}</td></tr>'
+        )
+    return f"""
+    <div class="tbl-wrap">
+      <table>
+        <thead><tr><th>Tactile sensitivity (normalized units)</th>
+        <th>mean |&Delta;a| tactile vs masked</th><th>max |&Delta;a|</th>
+        <th>mean |refinement|</th></tr></thead>
+        <tbody>{"".join(rows)}</tbody>
+      </table>
+    </div>"""
+
+
+def build_section(eval_data, findings_html=""):
+    parts = [
+        '\n  <!-- ======================= ABLATION ======================= -->',
+        '  <section id="ablation">',
+        '    <p class="sec-eyebrow">ablation</p>',
+        '    <h2>Ablation: what track, force, and deformation tactile each buy</h2>',
+        PROTOCOL_PROSE,
+        VARIANT_LEGEND,
+        VARIANT_TABLE,
+        loss_figure(),
+        final_loss_table(),
+        '    <h3>Controlled evaluation on frozen checkpoints</h3>',
+        '''    <div class="prose"><p>All numbers below come from the same fixed
+      anchor sets, noise seeds, and prompts for every variant
+      (<code>scripts/eval/trex_ablation_eval.py</code>). Open-loop: each variant
+      runs its own full sampling loop from pure noise on 16 anchors and is scored
+      against the ground-truth normalized delta-base chunk.</p></div>''',
+        f'  <figure>{openloop_figure(eval_data)}<figcaption><b>Figure A2 &mdash; Open-loop reconstruction.</b> Faded bars: the coarse action read out at &tau;=0 without tactile refinement (for C this is its only mode). Solid bars: the full cascade.</figcaption></figure>',
+        tau_table(eval_data),
+        f'  <figure>{stage1_attention_figure(eval_data)}<figcaption><b>Figure A3 &mdash; Stage-1 attention routing.</b> Share of self-attention mass from action queries onto each token group at &tau;=0.4 (mean over heads/blocks).</figcaption></figure>',
+        f'  <figure>{stage2_attention_figure(eval_data)}<figcaption><b>Figure A4 &mdash; Stage-2 attention routing.</b> Action-query mass over F6 / VQ-history / deform / coarse-memory tokens during real cascaded refinement.</figcaption></figure>',
+        sensitivity_table(eval_data),
+        findings_html,
+        '  </section>\n',
+    ]
+    return "\n".join(parts)
+
+
+if __name__ == "__main__":
+    eval_path = sys.argv[1] if len(sys.argv) > 1 else \
+        "/scratch1/home/zhicao/dreamzero/ablation_runs/eval_part1.json"
+    eval_data = json.load(open(eval_path))
+    section = build_section(eval_data)
+    out = TOOLS.parent / "ablation_section_preview.html"
+    out.write_text(section)
+    print(f"wrote {out} ({len(section)} chars)")

scripts/eval/trex_ablation_eval.py

@@ -562,6 +562,7 @@ def main() -> None:
             key: [float(np.mean(v)), float(np.std(v)), len(v)]
             for key, v in openloop.items()
         }
+        record["openloop_action_mse_raw"] = openloop
         record["openloop_action_mse_split"] = {
             key: float(np.mean(v)) for key, v in split_metrics.items()
         }