Spaces:

josefchen
/

orchestration-is-allocation

Running

App Files Files Community

josefchen commited on Jun 26

Commit

b782ada

verified ·

1 Parent(s): 5db5468

Glyph-dither hero (Intuition-style) + full 67-model roster; data + reproducibility callouts

Browse files

Files changed (3) hide show

.gitattributes +1 -0
bloom.png +3 -0
index.html +94 -18

.gitattributes CHANGED Viewed

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text

 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+bloom.png filter=lfs diff=lfs merge=lfs -text

bloom.png ADDED Viewed

Git LFS Details

SHA256: 8b04002fc87958617f31de1c7917efb616505f302448875d9c89ce8ecf54d3cc
Pointer size: 131 Bytes
Size of remote file: 386 kB

index.html CHANGED Viewed

@@ -3,7 +3,7 @@
 <head>
 <meta charset="utf-8">
 <meta name="viewport" content="width=device-width, initial-scale=1">
-<script>/* match the user's OS theme before first paint; the toggle stays authoritative once clicked */try{if(matchMedia('(prefers-color-scheme:light)').matches){document.documentElement.setAttribute('data-theme','light');}}catch(e){}</script>
 <title>When Does Combining Language Models Help? — the co-failure ceiling</title>
 <meta name="description" content="Interactive companion to the paper “When Does Combining Language Models Help?” (Orchestration Is Allocation). The deployment-relevant object in multi-model LLM orchestration is β = P(all models wrong); a one-sample certificate bounds every router, vote, and cascade.">
 <meta property="og:title" content="When Does Combining Language Models Help?">
@@ -322,6 +322,20 @@ footer a:hover{border-bottom-color:var(--text)}
   input[type=range]::-webkit-slider-thumb{width:22px;height:22px;margin-top:-9px}
   input[type=range]::-moz-range-thumb{width:22px;height:22px}
 }
 @media(prefers-reduced-motion:reduce){*{animation:none!important;transition:none!important;scroll-behavior:auto}}
 </style>
 </head>
@@ -337,23 +351,19 @@ footer a:hover{border-bottom-color:var(--text)}
   </span>
 </nav>
-<header class="wrap">
-  <div class="eyebrow">Orchestration Is Allocation · interactive companion · Josef Chen / KAIKAKU</div>
-  <h1>The lever for orchestration is <em>failure-mode dispersion</em>, not model count.</h1>
-  <p class="lede">Adding models does not raise the ceiling. The binding rate is β — how often <em>every</em> model is wrong at once — and it caps every router, vote, and cascade. One query sample certifies it.</p>
-  <div class="herostat">
-    <div class="item lead">
-      <span class="big beta num" id="heroBeta" data-ticker="0.052">0.052</span>
-      <span class="hero-law num" aria-label="accuracy is at most one minus beta"><span class="gb">β</span> · MATH-500 &#160;&#8594;&#160; <span class="hero-law-rule">accuracy ≤ 1 − <span class="gb">β</span></span></span>
-      <span class="lbl">co-failure point estimate on the primary run. Whatever the policy, accuracy cannot clear <span class="num">1 − <span class="gb">β</span>&#770; = <span id="heroCeil">0.948</span></span> — the certified bound is wider.</span>
-    </div>
-    <div class="hero-aside">
-      <span class="hero-aside-k num">format flip</span>
-      <span class="hero-aside-v num"><span class="gb">β</span> 0 <span class="hero-aside-arr" aria-hidden="true">&#8594;</span> <span class="beta">0.127</span></span>
-      <span class="hero-aside-l">strip the options off the same GPQA questions and a tail opens where there was none — <a href="#flip">the flip ↓</a></span>
     </div>
   </div>
-  <div class="herospec num">67 frontier models · 21 providers · 4 structurally disjoint domains · temperature 0</div>
 </header>
 <section id="thesis" class="wrap">
@@ -487,14 +497,14 @@ footer a:hover{border-bottom-color:var(--text)}
   </div>
   <div class="poolstat"><span class="beta num">67</span> models &middot; <span class="beta num">21</span> providers &middot; priced live &middot; temperature 0 &middot; one 67&times;67 co-failure matrix</div>
   <div id="poolGrid" class="poolgrid"></div>
-  <div class="cap" style="max-width:64ch">Tier: <span class="tierdot f"></span> frontier &middot; <span class="tierdot m"></span> mid &middot; <span class="tierdot c"></span> cheap / open-weight. Every instrument above draws on this pool; the full roster with live prices ships in the paper's appendix.</div>
 </section>
 <footer class="wrap" id="paper">
   <div class="foot-grid">
     <div style="max-width:42ch">
       <div class="foot-title">Orchestration Is Allocation</div>
-      <p style="color:var(--dim);font-size:13.5px">Josef Chen · KAIKAKU. Every figure here recomputes from committed outcome matrices over 67 frontier models; the reproducibility bundle ships the scripts and data. Grading scope is stated honestly in the paper — the open-ended panel is five LLM judges, not human adjudication.</p>
       <p style="margin-top:16px"><span class="foot-pending">Read the paper (arXiv)</span> <span class="mono-note" style="display:inline">&#160;— link goes live when the arXiv id is assigned</span></p>
     </div>
     <div class="mono-note">
@@ -785,6 +795,72 @@ const ROSTER=[{n:"GPT-5.5",p:"OpenAI",t:"f"},{n:"Claude Opus 4.8",p:"Anthropic",
   }).join('');
 })();
 (function panHints(){
   const frames=[...document.querySelectorAll('.panx')];
   const sync=el=>el.classList.toggle('can-pan',el.scrollWidth-el.clientWidth>1);

 <head>
 <meta charset="utf-8">
 <meta name="viewport" content="width=device-width, initial-scale=1">
+<script>/* dark-first: the glyph-dither hero is black; the theme toggle stays authoritative once clicked */</script>
 <title>When Does Combining Language Models Help? — the co-failure ceiling</title>
 <meta name="description" content="Interactive companion to the paper “When Does Combining Language Models Help?” (Orchestration Is Allocation). The deployment-relevant object in multi-model LLM orchestration is β = P(all models wrong); a one-sample certificate bounds every router, vote, and cascade.">
 <meta property="og:title" content="When Does Combining Language Models Help?">
   input[type=range]::-webkit-slider-thumb{width:22px;height:22px;margin-top:-9px}
   input[type=range]::-moz-range-thumb{width:22px;height:22px}
 }
+/* ---- glyph-dither hero (Intuition-style pixel bloom) ---- */
+.hero-dither{position:relative;width:100%;height:clamp(560px,84vh,880px);background:#070809;overflow:hidden;border-bottom:.5px solid var(--hair);cursor:crosshair}
+#glyphCanvas{position:absolute;inset:0;width:100%;height:100%;display:block}
+.hero-grad{position:absolute;inset:0;pointer-events:none;background:radial-gradient(130% 100% at 58% 48%,transparent 38%,rgba(7,8,9,.55) 100%)}
+.hd-chrome{position:absolute;inset:0;z-index:2;pointer-events:none}
+.hd-mark{position:absolute;top:clamp(26px,5vh,58px);left:clamp(20px,4vw,64px);margin:0;font-family:var(--serif);font-weight:340;font-size:clamp(2.8rem,9vw,6.6rem);line-height:.86;letter-spacing:-.03em;color:#F5F3F4;text-shadow:0 4px 60px rgba(0,0,0,.45)}
+.hd-tag{position:absolute;top:clamp(30px,5.4vh,66px);right:clamp(20px,4vw,64px);font-family:var(--mono);font-size:11.5px;letter-spacing:.09em;line-height:1.75;color:#CDD0D6;text-align:right;text-shadow:0 2px 20px rgba(0,0,0,.6)}
+.hd-date{position:absolute;left:clamp(8px,2.4vw,30px);top:50%;transform:translateY(-50%) rotate(180deg);writing-mode:vertical-rl;font-family:var(--mono);font-size:10.5px;letter-spacing:.16em;color:#7E848D}
+.hd-foot{position:absolute;left:clamp(20px,4vw,64px);bottom:clamp(24px,5vh,50px);display:flex;align-items:flex-end;gap:18px}
+.hd-foot .hd-beta{font-family:var(--mono);font-weight:300;font-size:clamp(30px,4.4vw,46px);letter-spacing:-.02em;line-height:.82;color:var(--accent)}
+.hd-foot .gb{color:var(--accent);font-family:var(--serif);font-style:italic}
+.hd-foot-l{font-family:var(--mono);font-size:10px;letter-spacing:.07em;line-height:1.65;color:#9298A1;padding-bottom:3px}
+.hd-hint{position:absolute;right:clamp(20px,4vw,64px);bottom:clamp(24px,5vh,50px);font-family:var(--mono);font-size:10px;letter-spacing:.06em;color:#6A707A;text-transform:uppercase}
+@media(max-width:640px){.hd-date{display:none}.hd-tag{font-size:9.5px}.hd-hint{display:none}.hd-mark{font-size:clamp(2.6rem,13vw,4rem)}}
 @media(prefers-reduced-motion:reduce){*{animation:none!important;transition:none!important;scroll-behavior:auto}}
 </style>
 </head>
   </span>
 </nav>
+<header class="hero-dither" id="top">
+  <canvas id="glyphCanvas" aria-hidden="true"></canvas>
+  <div class="hero-grad" aria-hidden="true"></div>
+  <div class="hd-chrome">
+    <h1 class="hd-mark">CO&#8202;&middot;&#8202;FAILURE</h1>
+    <div class="hd-tag">WHEN EVERY MODEL IS<br>WRONG AT ONCE, NO<br>ENSEMBLE CAN WIN</div>
+    <div class="hd-date">2026 &middot; 06 &middot; 15 &nbsp;&middot;&nbsp; OPENROUTER FRONTIER</div>
+    <div class="hd-foot">
+      <span class="hd-beta num" id="heroBeta" data-ticker="0.052">0.052</span>
+      <span class="hd-foot-l"><span class="gb">β</span> = P(ALL MODELS WRONG) &middot; MATH-500<br>17 OF 330 QUERIES &middot; 67 MODELS, 21 PROVIDERS</span>
     </div>
+    <div class="hd-hint">move the cursor &mdash; the bloom scatters</div>
   </div>
 </header>
 <section id="thesis" class="wrap">
   </div>
   <div class="poolstat"><span class="beta num">67</span> models &middot; <span class="beta num">21</span> providers &middot; priced live &middot; temperature 0 &middot; one 67&times;67 co-failure matrix</div>
   <div id="poolGrid" class="poolgrid"></div>
+  <div class="cap" style="max-width:64ch">Tier: <span class="tierdot f"></span> frontier &middot; <span class="tierdot m"></span> mid &middot; <span class="tierdot c"></span> cheap / open-weight. Every instrument above draws on this pool, and all of it is released to replicate: the full roster with live prices, the 67&times;67 outcome matrices, the grading, and the analysis code. Every number on this page regenerates offline.</div>
 </section>
 <footer class="wrap" id="paper">
   <div class="foot-grid">
     <div style="max-width:42ch">
       <div class="foot-title">Orchestration Is Allocation</div>
+      <p style="color:var(--dim);font-size:13.5px">Josef Chen · KAIKAKU. Every figure here recomputes from committed outcome matrices over 67 frontier models. The full reproducibility bundle is uploaded: the outcome matrices, graders, and analysis code that regenerate every number in the paper offline. Grading scope is stated honestly: the open-ended panel is five LLM judges, not human adjudication.</p>
       <p style="margin-top:16px"><span class="foot-pending">Read the paper (arXiv)</span> <span class="mono-note" style="display:inline">&#160;— link goes live when the arXiv id is assigned</span></p>
     </div>
     <div class="mono-note">
   }).join('');
 })();
+/* ---------- glyph-dither hero: bloom.png sampled to a pixel grid, scattered by a fluid cursor trail ---------- */
+(function glyphDither(){
+  const cv=document.getElementById('glyphCanvas'); if(!cv||!cv.getContext) return;
+  const ctx=cv.getContext('2d',{alpha:false});
+  const reduce=matchMedia('(prefers-reduced-motion:reduce)').matches;
+  const img=new Image();
+  let cells=[],cell=10,cols=0,rows=0,W=0,H=0,dpr=1,ready=false;
+  function layout(){
+    const rect=cv.getBoundingClientRect(); W=Math.max(1,rect.width); H=Math.max(1,rect.height);
+    dpr=Math.min(2,window.devicePixelRatio||1);
+    cv.width=Math.round(W*dpr); cv.height=Math.round(H*dpr); ctx.setTransform(dpr,0,0,dpr,0,0);
+    cell=Math.max(7,Math.round(W/128));            // grid density (≈ glyph scale)
+    cols=Math.ceil(W/cell); rows=Math.ceil(H/cell); sample();
+  }
+  function sample(){
+    if(!img.complete||!img.naturalWidth) return;
+    const off=document.createElement('canvas'); off.width=cols; off.height=rows;
+    const o=off.getContext('2d');
+    const ir=img.naturalWidth/img.naturalHeight, cr=cols/rows; let sw,sh,sx,sy;   // cover-fit
+    if(ir>cr){ sh=img.naturalHeight; sw=sh*cr; sx=(img.naturalWidth-sw)/2; sy=0; }
+    else { sw=img.naturalWidth; sh=sw/cr; sx=0; sy=(img.naturalHeight-sh)/2; }
+    o.drawImage(img,sx,sy,sw,sh,0,0,cols,rows);
+    const d=o.getImageData(0,0,cols,rows).data; cells.length=0;
+    for(let j=0;j<rows;j++)for(let i=0;i<cols;i++){
+      const k=(j*cols+i)*4, R=d[k],G=d[k+1],B=d[k+2];
+      const l=(0.299*R+0.587*G+0.114*B)/255;
+      if(l<0.05) continue;                          // skip near-black: cleaner dither + faster
+      cells.push({bx:i*cell+cell*0.5, by:j*cell+cell*0.5, r:R,g:G,b:B,l, ox:0,oy:0});
+    }
+    ready=true;
+  }
+  let mx=-1e4,my=-1e4;
+  const pt=e=>{const rect=cv.getBoundingClientRect(); mx=e.clientX-rect.left; my=e.clientY-rect.top;};
+  const out=()=>{mx=-1e4;my=-1e4;};
+  const hdr=cv.closest('.hero-dither')||cv;
+  hdr.addEventListener('pointermove',pt,{passive:true}); hdr.addEventListener('pointerleave',out);
+  const RAD=130, STR=30;
+  function frame(t){
+    if(ready){
+      ctx.fillStyle='#070809'; ctx.fillRect(0,0,W,H);
+      for(let n=0;n<cells.length;n++){ const c=cells[n]; let jit=0;
+        if(!reduce){
+          const dx=c.bx-mx, dy=c.by-my, dist=Math.sqrt(dx*dx+dy*dy); let tx=0,ty=0;
+          if(dist<RAD){ const f=1-dist/RAD; jit=f;
+            const ang=Math.atan2(dy,dx)+Math.sin(c.bx*0.045+c.by*0.03+t*0.0021)*1.5;   // noise-displaced direction
+            const push=f*f*STR; tx=Math.cos(ang)*push; ty=Math.sin(ang)*push-push*0.25;
+          }
+          c.ox+=(tx-c.ox)*0.16; c.oy+=(ty-c.oy)*0.16;                                   // ease in/out = trail + dissipation
+        }
+        const px=c.bx+c.ox, py=c.by+c.oy, s=cell*(0.42+0.74*c.l), h=s*0.5;
+        if(jit>0.22){                                                                   // chromatic split while scattered
+          const o2=jit*3.2; ctx.globalAlpha=0.5;
+          ctx.fillStyle='rgb('+c.r+',46,52)'; ctx.fillRect(px-o2-h,py-h,s,s);
+          ctx.fillStyle='rgb(54,'+c.g+',190)'; ctx.fillRect(px+o2-h,py-h,s,s);
+          ctx.globalAlpha=1;
+        }
+        ctx.fillStyle='rgb('+c.r+','+c.g+','+c.b+')'; ctx.fillRect(px-h,py-h,s,s);
+      }
+    }
+    requestAnimationFrame(frame);
+  }
+  img.onload=()=>{ layout(); requestAnimationFrame(frame); };
+  img.src='bloom.png';
+  let rt; addEventListener('resize',()=>{clearTimeout(rt);rt=setTimeout(layout,160);},{passive:true});
+})();
 (function panHints(){
   const frames=[...document.querySelectorAll('.panx')];
   const sync=el=>el.classList.toggle('can-pan',el.scrollWidth-el.clientWidth>1);