Upload index.html

index.html

@@ -226,9 +226,9 @@
     @keyframes pulse{70%{box-shadow:0 0 0 9px rgba(60,205,180,0)}100%{box-shadow:0 0 0 0 rgba(60,205,180,0)}}
 
     h1{
-      font-size:clamp(3.2rem,7.5vw,6.7rem);
+      font-size:clamp(2.8rem,6vw,5.35rem);
       line-height:.91;
-      letter-spacing:-.065em;
+      letter-spacing:-.045em;
       margin:22px 0 24px;
       max-width:810px;
     }
@@ -257,8 +257,8 @@
     }
     .btn.primary{
       color:white;
-      background:linear-gradient(135deg,#1788c1,#27b7c9 55%,#3cc7aa);
-      box-shadow:0 10px 26px rgba(35,152,186,.28);
+      background:#197fa9;
+      box-shadow:0 6px 16px rgba(35,152,186,.16);
     }
     .btn.secondary{background:rgba(255,255,255,.38)}
     body.dark .btn.secondary{background:rgba(255,255,255,.045)}
@@ -748,7 +748,11 @@
     @media (prefers-reduced-motion: reduce){
       *,*::before,*::after{animation:none!important;transition:none!important;scroll-behavior:auto!important}
     }
-  </style>
+  .hero-main,.hero-side{border-radius:24px!important;box-shadow:0 12px 32px rgba(39,88,112,.09)}
+.eyebrow{border-radius:10px}.stat-card{border-radius:16px}.glass-card,.callout,.table-wrap,.timeline-card{border-radius:18px}
+.btn{border-radius:11px}.arch-row{border-radius:12px}.arch-row .tag{border-radius:7px}.toc{border-radius:18px}
+.section-kicker{letter-spacing:.055em}.quote{font-weight:650}
+</style>
 </head>
 <body>
   <div class="progress" id="progress"></div>
@@ -781,15 +785,15 @@
     <div class="hero-grid">
       <div class="hero-main">
         <div class="eyebrow"><span class="pulse-dot"></span> Ongoing independent ML research</div>
-        <h1>Transformer 2:<br><span class="soft">what survived the ablations.</span></h1>
+        <h1>Transformer 2<br><span class="soft">Research Notes</span></h1>
         <p class="dek">
-          Orion began with a simple idea: maybe a language model should not force persistent knowledge,
-          reusable procedures, and temporary reasoning into the same machinery. After multiple generations,
-          some of that idea survived. Some of it did not.
+          Notes from the Orion / Project Prism experiments on separating persistent knowledge,
+          reusable procedures, temporary state, and recurrent computation. This page records the architecture,
+          measurements, negative results, and changes made between T2 generations.
         </p>
         <div class="hero-actions">
-          <a class="btn primary" href="#hypothesis">Read the research →</a>
-          <a class="btn secondary" href="#scale">Jump to the 2B run</a>
+          <a class="btn primary" href="#hypothesis">Start with the hypothesis</a>
+          <a class="btn secondary" href="#scale">Current 2B experiment</a>
         </div>
       </div>
 
@@ -866,7 +870,7 @@
 
     <main>
       <section id="hypothesis">
-        <div class="section-kicker">01 / Research premise</div>
+        <div class="section-kicker">01 / Original hypothesis</div>
         <h2>The original hypothesis</h2>
         <p class="lead">
           Orion / Project Prism is an experimental architecture family I call <strong>Transformer 2 (T2)</strong>.
@@ -960,7 +964,7 @@
       </section>
 
       <section id="mini-ablation">
-        <div class="section-kicker">03 / First causal probes</div>
+        <div class="section-kicker">03 / Mini ablations</div>
         <h2>Mini ablations changed the story</h2>
         <p class="lead">
           One early set of inference-time ablations made the Working State look much more important than the Vault
@@ -1092,7 +1096,7 @@
       </section>
 
       <section id="benchmark">
-        <div class="section-kicker">06 / One benchmark, not a universal verdict</div>
+        <div class="section-kicker">06 / Nano benchmark</div>
         <h2>Nano’s BananaMind result</h2>
         <p>
           On BananaMind Base Bench 1.1, a 350-example multiple-choice benchmark, Nano reached an overall Elo of
@@ -1124,7 +1128,7 @@
       </section>
 
       <section id="vault">
-        <div class="section-kicker">07 / A useful negative result</div>
+        <div class="section-kicker">07 / Knowledge Vault</div>
         <h2>The Knowledge Vault kept failing to earn its weights</h2>
         <p class="lead">
           Across Orion generations and scales, the explicit neural Knowledge Vault repeatedly produced much weaker
@@ -1258,7 +1262,7 @@
       </section>
 
       <section id="dynamics">
-        <div class="section-kicker">10 / Why this run is still early</div>
+        <div class="section-kicker">10 / Training dynamics</div>
         <h2>Working State may not have “arrived” yet</h2>
         <p class="lead">
           In previous Orion generations, Working State often became much more important only after the model had
@@ -1275,7 +1279,7 @@
       </section>
 
       <section id="t22">
-        <div class="section-kicker">11 / Emerging next generation</div>
+        <div class="section-kicker">11 / T2.2 design notes</div>
         <h2>T2.2 is becoming simpler, not more ornate</h2>
         <p>
           The provisional T2.2 direction follows the evidence instead of preserving every original idea:
@@ -1313,7 +1317,7 @@
       </section>
 
       <section id="vm">
-        <div class="section-kicker">12 / Future experiment, not a validated feature</div>
+        <div class="section-kicker">12 / Proposed exact-compute path</div>
         <h2>A possible built-in exact-compute path</h2>
         <p>
           One future T2.2 idea is to add a deterministic computation path — effectively a small built-in VM or
@@ -1351,7 +1355,7 @@
       </section>
 
       <section id="science">
-        <div class="section-kicker">13 / Scientific caution</div>
+        <div class="section-kicker">13 / Evidence and limitations</div>
         <h2>What the current evidence does — and does not — prove</h2>
         <p class="lead">
           Inference-time ablations answer an important question: <em>what did this particular trained checkpoint