OpenAI Decisions: real span=6 number (0.825/.41), fix overwritten nutrient-document-decision row + models-benchmarked block

index.html

@@ -128,9 +128,9 @@
           <td class="g1"><span class="f1">0.917</span><span class="kap">.86</span></td>
         </tr>
         <tr>
-          <td class="model"><div class="modelname">nutrient-document-decision</div><div class="modelsub">v7 · Qwen3.5-4B LoRA decision model (general-purpose, shared with grounding/doc-cls) · chunk-based span=6</div><span class="chip flag">Commercial</span></td>
+          <td class="model"><div class="modelname">nutrient-document-decision</div><div class="modelsub">unified multi-task model (grounding + doc-cls + doc-split)</div><span class="chip flag">Commercial</span></td>
           <td class="g1"><span class="f1">0.962</span><span class="kap">.86</span></td>
-          <td class="g3"><span class="f1">0.683</span><span class="kap">.64</span></td>
+          <td class="g3"><span class="f1">0.683</span><span class="kap">.63</span></td>
           <td class="g1"><span class="f1">0.959</span><span class="kap">.95</span></td>
           <td class="g1"><span class="f1">0.960</span><span class="kap">.94</span></td>
           <td class="g1"><span class="f1">0.987</span><span class="kap">.97</span></td>
@@ -177,8 +177,8 @@
         </tr>
         <tr class="grouprow"><td colspan="7">Cloud Hosted System-One (text or image; only Jev is text-only)</td></tr>
         <tr>
-          <td class="model"><div class="modelname">OpenAI Decisions API</div><div class="modelsub">gpt-6-luna · image · pairwise span-2 boundary calls · our-200 only so far</div><span class="chip cloud">Cloud</span></td>
-          <td class="g3"><span class="f1">0.843</span><span class="kap">.54</span></td>
+          <td class="model"><div class="modelname">OpenAI Decisions API</div><div class="modelsub">gpt-6-luna · image</div><span class="chip cloud">Cloud</span></td>
+          <td class="g2"><span class="f1">0.825</span><span class="kap">.41</span></td>
           <td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td>
         </tr>
         <tr>
@@ -223,10 +223,20 @@
     <p><b>One balanced model vs two specialists.</b> No single OpenPSS model wins both slices — their short-specialist craters on long (0.50), their long-specialist on short (0.62). the v4 flagship does short + long with one model, and its long (0.891) tops even their long-specialist (0.83).</p>
     <p><b>Cloud VLMs can't ingest long streams.</b> Per-request image caps (Anthropic ~100 / OpenAI ~200 / Gemini ~500) force predict-none on streams over the cap, which dominates OpenPSS-long → best cloud 0.244 vs flagship 0.891, at ~20× the cost per page.</p>
     <p><b>* bert-pss</b> self-declares ~0.915 accuracy / 0.825 κ on Tobacco800, but collapses to a single class when actually run (κ ≈ 0) — the only released PSS specialist is non-functional off its exact serving harness.</p>
-    <p><b>OpenAI Decisions API (gpt-6-luna).</b> Public-beta <code>/v1/decisions</code> endpoint, one <code>predicate</code> call per adjacent page pair (image-only, span-2 pairwise — directly comparable to the flagship's span-2 pairwise-equivalent baseline of 0.944/.79). Scored on our-200 only so far: 0.843 F1 / .54 κ. Other five slices not yet run.</p>
+    <p><b>OpenAI Decisions API (gpt-6-luna).</b> Public-beta <code>/v1/decisions</code> endpoint, scored with the SAME chunk-based span=6/stitch=average methodology as the flagship and nutrient-document-decision (shared-state chunk images + one predicate question per boundary in a single request, not a simplified pairwise scheme). Scored on our-200 only so far: 0.825 F1 / .41 κ, vs flagship 0.944/.79 and nutrient-document-decision 0.962/.86. Other five slices not yet run.</p>
     <p><b>Jev (TypeSafe), text-only.</b> The hosted System-One model reads page OCR only (no image). On our-200 it scores 0.637 F1 / .29 κ — high precision, low recall (under-splits textually-similar form/receipt pages), and giving it the whole document instead of adjacent pages does not help (pairwise 0.650). Confirms split is image-native: text alone tops out ~0.64 F1 here, far below the image models (ours 0.944).</p>
     <p><b>Metric.</b> Ours &amp; cloud: micro boundary-F1 + κ over internal pages, identical harness. OpenPSS rows: their published per-stream page-F1 (ballpark-comparable). Tobacco incumbents report accuracy — compared via κ. AI-Lab-Splitter omitted: data gated, absolute F1 paywalled.</p>
   </div>
+
+  <p id="models-benchmarked" style="color:var(--muted);font-size:12.5px;margin-top:28px;max-width:88ch">
+    <strong>Models benchmarked:</strong>
+    <a href="https://huggingface.co/nutrientdocs/doc-split-v2">nutrientdocs/doc-split-v2</a> ·
+    <a href="https://huggingface.co/nutrientdocs/doc-split-v1">nutrientdocs/doc-split-v1</a> ·
+    <a href="https://huggingface.co/nutrientdocs/nutrient-document-decision">nutrientdocs/nutrient-document-decision</a> ·
+    <a href="https://huggingface.co/agiagoulas/bert-pss">agiagoulas/bert-pss</a>.
+    Cloud VLMs (no public HF page): Gemini 2.5 Pro / Flash · GPT-sol · Claude Opus · OpenAI Decisions API (gpt-6-luna).
+    OpenPSS SHORT/LONG specialists and AI-Lab-Splitter: published results, no public HF page.
+  </p>
 </div>
 </body>
 </html>