From e6beb61e623742f57970596724bb55cebd175327 Mon Sep 17 00:00:00 2001
From: Experimenting With Computers
 <1908752+makefunstuff@users.noreply.github.com>
Date: Sun, 20 Sep 2026 19:01:17 +0300
Subject: [PATCH] Clarify demo, classifier, and cost section copy

---
 site/index.html | 42 ++++++++++++++++++++++--------------------
 1 file changed, 22 insertions(+), 20 deletions(-)

diff --git a/site/index.html b/site/index.html
index bf6b643..e0e3b0c 100644
--- a/site/index.html
+++ b/site/index.html
@@ -103,16 +103,17 @@ <h1>Steer code with rules, not with token maxxing.</h1>
   <div class="shell">
     <header class="sec-head">
       <p class="eyebrow"><span class="eyebrow-dot" aria-hidden="true"></span>Demo</p>
-      <h2>Finding and fixing AI-slop commentary</h2>
-      <p class="sec-lead">Silent action loop across OpenCode, Neovim, and VS Code: paste an AI-slop comment, get a finding, apply the fix.</p>
+      <h2>Same <code>no-ai-slop</code> hit in three clients</h2>
+      <p class="sec-lead">OpenCode writes a matching comment and gets the diagnostic. Neovim shows it as virtual text. VS Code clears it with a code action. Silent, ~32s.</p>
     </header>
 
     <figure class="demo-frame">
       <video controls playsinline preload="metadata" width="100%"
-             aria-label="Silent demo: OpenCode, Neovim, and VS Code detect and fix an AI-slop comment">
+             poster="./assets/demo-slop-triptych-poster.jpg"
+             aria-label="Silent screencap: OpenCode, Neovim, and VS Code on the same no-ai-slop finding">
         <source src="./assets/demo-slop-triptych.mp4" type="video/mp4">
       </video>
-      <figcaption>OpenCode → Neovim → VS Code · ~32s · no voiceover</figcaption>
+      <figcaption>OpenCode → Neovim → VS Code · ordinary LSP diagnostic</figcaption>
     </figure>
   </div>
 </section>
@@ -169,28 +170,29 @@ <h2>Findings on the line: same surfaces your client already has</h2>
     <div class="split-copy">
       <p class="eyebrow"><span class="eyebrow-dot" aria-hidden="true"></span>Model shape</p>
       <h2>Classifier, not chat</h2>
+      <p class="sec-lead">Chat reviews emit prose about the whole file. Jev answers one rule question about one candidate line: a typed value and a confidence <code>p</code>.</p>
       <ul class="plain">
-        <li><strong>One question → one answer</strong>: pass/fail or typed value + probability floor.</li>
-        <li><strong>Rules in the repo</strong>: <code>.jev/rules/*.json</code>, versioned like code.</li>
-        <li><strong>Semi-deterministic</strong>: pattern exact; judgement must clear the floor before publish.</li>
+        <li><strong>One question → one answer</strong>: <code>pass</code>/<code>fail</code> (or another typed value), plus <code>p</code>.</li>
+        <li><strong>Rules in the repo</strong>: <code>.jev/rules/*.json</code>, reviewed like code.</li>
+        <li><strong>Gate before publish</strong>: regex picks the line; <code>p</code> must clear the rule <code>floor</code> or nothing is emitted.</li>
       </ul>
     </div>
     <div class="compare-stack">
       <div class="chat-reject">
-        <p class="compare-label">prose · whole file · every turn</p>
+        <p class="compare-label">chat review · reject</p>
         <p class="chat-blob">Here are several thoughts about this file… consider error handling on line 412, also naming, also whether the module boundary is right, also…</p>
       </div>
       <div class="win win-answer">
         <div class="win-chrome">
           <span class="traffic" aria-hidden="true"><i></i><i></i><i></i></span>
-          <span class="win-title">Jev rule answer</span>
-          <span class="win-meta">typed · one line</span>
+          <span class="win-title">Jev decide</span>
+          <span class="win-meta">one line · one rule</span>
         </div>
         <pre class="win-body">{
   <span class="k">"value"</span>: <span class="amb">"fail"</span>,
   <span class="k">"p"</span>: <span class="n">0.91</span>
 }</pre>
-        <div class="win-foot">typed value · one line · clears rule floor</div>
+        <div class="win-foot">publish only if p ≥ floor (e.g. 0.8)</div>
       </div>
     </div>
   </div>
@@ -231,8 +233,8 @@ <h2>The loop</h2>
   <div class="shell">
     <header class="sec-head">
       <p class="eyebrow"><span class="eyebrow-dot" aria-hidden="true"></span>Economics</p>
-      <h2>Why local / cheap models work here</h2>
-      <p class="sec-lead">Rules carry conventions; the model answers about one candidate line. No matching pattern → <strong>no model call</strong>.</p>
+      <h2>One line costs less than a whole-file chat</h2>
+      <p class="sec-lead">Patterns decide which lines run the model. No match → <strong>no call</strong>. A match asks about that line only, so a small/local decide model is enough.</p>
     </header>
 
     <div class="cost-card">
@@ -246,27 +248,27 @@ <h2>Why local / cheap models work here</h2>
           <thead>
             <tr>
               <th scope="col"></th>
-              <th scope="col">rules pass</th>
-              <th scope="col">chat review (same file)</th>
+              <th scope="col">jev rules pass</th>
+              <th scope="col">whole-file chat review</th>
             </tr>
           </thead>
           <tbody>
             <tr class="mark">
               <th scope="row"><code>server.rs</code> ~2.6k lines</th>
-              <td>~5.3k in / 345 out · 2 findings · <strong>~$0.00022</strong></td>
-              <td>~30.5k in / 9 out · 0 findings · <strong>~$0.00306</strong></td>
+              <td>2 findings · <strong>~$0.00022</strong><br><span class="dim">~5.3k in / 345 out</span></td>
+              <td>0 findings · <strong>~$0.00306</strong><br><span class="dim">~30.5k in / 9 out</span></td>
             </tr>
             <tr>
               <th scope="row">30 Rust docs</th>
-              <td><strong>~$0.0034</strong> · 20 calls</td>
+              <td><strong>~$0.0034</strong> · 20 decide calls</td>
               <td>—</td>
             </tr>
           </tbody>
         </table>
       </div>
-      <p class="cost-cap">decide <code>jev-1.13</code> vs chat review <code>gemini-2.5-flash-lite</code>. PoC: accuracy from steering, not token maxxing.</p>
+      <p class="cost-cap"><code>jev-1.13</code> decide vs <code>gemini-2.5-flash-lite</code> chat on the same file. Measured in this repo.</p>
     </div>
-    <p class="pull">Accuracy from steering, not from burning tokens.</p>
+    <p class="pull">Cheaper because most lines never call the model.</p>
   </div>
 </section>