docs: honest positioning after Ollama comparison

Ollama qwen2.5:14b scores 40/40 with --format json. Updated tagline,
comparison table, and blog post to position on speed (10-25x faster)
and no server dependency instead of compliance exclusivity.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: 2 people

TODOS.md

@@ -12,21 +12,15 @@ Generated by /plan-eng-review on 2026-04-30
 - **Context:** Design-Doc spezifiziert API-Kontrakt bereits. ~100 Zeilen Code + Tests. Blocked by: llmjson Package fertig.
 - **Depends on:** llmjson Package (Step 1-3 des Sprint)
 
-### TODO-002: Ollama-Vergleichstest
-- **What:** Teste Ollama `--format json` gegen die 4 BNR-002-Schemas (simple_object, nested_object, financial_transfer, medical_record). Dokumentiere Fehlerrate.
-- **Why:** Differenzierungsmerkmal für Marketing. Wenn Ollama 100% schafft, hat llmjson kein USP.
-- **Pros:** Harter Beweis für den Blogpost, nicht nur Behauptung
-- **Cons:** Falls Ollama ebenfalls 100% liefert, ist das Marketing-Argument weg
-- **Context:** Ollama bietet GBNF-Grammar-Support, operiert aber auf Grammar-Ebene, nicht Logit-Ebene. Erwartung: niedrigere Compliance bei komplexen Schemas.
-- **Depends on:** —
-
-### TODO-003: Blogpost-Titel und Claims ehrlich formulieren
-- **What:** "100% valid JSON from local LLMs" ersetzen durch "100% valid JSON for supported schema types from local LLMs" oder äquivalent
-- **Why:** Outside Voice: HN findet Edge Cases in Stunden. Schema-Einschränkungen müssen prominent stehen, nicht als Fußnote.
-- **Pros:** Ehrlich, baut Vertrauen auf, vermeidet HN-Backlash
-- **Cons:** Weniger catchy Headline
-- **Context:** Sicherer Nutzungsbereich schließt $ref, patternProperties, additionalProperties aus. Das muss in README, Blogpost, und Landing Page klar kommuniziert werden.
-- **Depends on:** —
+### TODO-002: Ollama-Vergleichstest ✅ DONE (2026-04-30)
+- **Ergebnis:** Ollama qwen2.5:14b --format json: 40/40 = 100% schema compliance
+- **Konsequenz:** "100% valid JSON" ist kein USP. Differenzierer ist Speed (10-25x schneller) und keine Server-Abhängigkeit.
+- **Test:** tests/test_ollama_comparison.py
+- **Marketing angepasst:** Tagline, Vergleichstabelle, Blogpost aktualisiert.
+
+### TODO-003: Blogpost-Titel und Claims ehrlich formulieren ✅ DONE (2026-04-30)
+- **Ergebnis:** Schema-Support-Tabelle auf Landing Page, Qualifier im Blogpost, Ollama-Vergleich ehrlich dokumentiert.
+- **Tagline:** "Valid JSON from local LLMs. No server, no retries, 10x faster than Ollama."
 
 ## Post-Launch (Woche 3+)
 

site/blog/how-i-got-100-percent-valid-json.html

@@ -274,22 +274,24 @@ <h2>Limitations (honest ones)</h2>
 
 <p><strong>Token alignment.</strong> Most tokenizers split text into subword tokens that don't align perfectly with JSON structural characters. A string like <code>"name":</code> might be one token or four. The tracker handles this by processing decoded text character by character, but edge cases exist with unusual tokenizers.</p>
 
-<h2>Why local matters</h2>
+<p><strong>Ollama does this too.</strong> I ran the same 4 schemas through Ollama's <code>--format json</code> with Qwen2.5:14b. It scored 40/40. Ollama uses GBNF grammars to constrain output and it works. If you already use Ollama and don't care about speed, it's a fine choice. llmjson is 10-25x faster on the same hardware because it talks to MLX directly instead of going through a server.</p>
 
-<p>OpenAI's Structured Outputs mode is good. If you're using their API and you're fine sending your data to their servers, it might be all you need.</p>
+<h2>Why llmjson over Ollama?</h2>
 
-<p>But there are reasons to want local inference:</p>
+<p>Both guarantee valid JSON. The differences are practical:</p>
 
 <ul>
-<li><strong>Privacy:</strong> The data never leaves your machine. Medical records, financial data, proprietary content.</li>
-<li><strong>Cost:</strong> After the one-time model download, generation is free. No per-token charges.</li>
-<li><strong>Latency:</strong> No network round-trip. On an M3 Max, generation starts in milliseconds.</li>
-<li><strong>Control:</strong> You pick the model, the parameters, the schema. No API limits, no deprecation notices.</li>
+<li><strong>Speed:</strong> llmjson generates in ~1-2 seconds. Ollama takes 10-25 seconds for the same schema on the same Mac. Direct MLX access vs. HTTP intermediary.</li>
+<li><strong>No daemon:</strong> llmjson is a Python library. Import it, call it. No background server to manage.</li>
+<li><strong>Privacy:</strong> The data never leaves your process. No local network socket, no shared memory.</li>
+<li><strong>Logit access:</strong> You get the raw probability distribution. Useful if you're building something beyond basic JSON generation.</li>
 </ul>
 
+<p>If you need an HTTP API, use Ollama. If you want a library you can embed in your Python code with maximum speed, use llmjson.</p>
+
 <h2>How to try it</h2>
 
-<pre><code>pip install jsongate</code></pre>
+<pre><code>pip install llmjson</code></pre>
 
 <p>You'll need Python 3.9+ and an Apple Silicon Mac (M1 or later). The MLX dependencies install automatically.</p>
 

site/index.html

@@ -262,8 +262,8 @@
 
   <section class="hero">
     <h1><span>llmjson</span></h1>
-    <p class="tagline">100% valid JSON from local LLMs. Zero retries. Zero post-processing. Constrained decoding on Apple Silicon.</p>
-    <div class="install-box" onclick="navigator.clipboard.writeText('pip install jsongate')" title="Click to copy">pip install jsongate</div>
+    <p class="tagline">Valid JSON from local LLMs. No server, no retries, 10x faster than Ollama. Constrained decoding on Apple Silicon.</p>
+    <div class="install-box" onclick="navigator.clipboard.writeText('pip install llmjson')" title="Click to copy">pip install llmjson</div>
   </section>
 
   <section class="stats">
@@ -334,49 +334,56 @@ <h3>Valid JSON, guaranteed</h3>
   </section>
 
   <section class="compare">
-    <h2>Why not just retry?</h2>
+    <h2>How does it compare?</h2>
     <table>
       <thead>
         <tr>
           <th>Approach</th>
           <th>100% valid</th>
-          <th>No retries</th>
+          <th>Speed</th>
+          <th>No server</th>
           <th>Local/private</th>
-          <th>Schema-aware</th>
         </tr>
       </thead>
       <tbody>
         <tr class="highlight-row">
           <td>llmjson</td>
           <td class="yes">Yes</td>
-          <td class="yes">Yes</td>
+          <td class="yes">~1-2s</td>
           <td class="yes">Yes</td>
           <td class="yes">Yes</td>
         </tr>
         <tr>
-          <td>Retry + parse</td>
-          <td class="no">No</td>
-          <td class="no">No</td>
-          <td class="partial">Depends</td>
-          <td class="no">No</td>
+          <td>Ollama --format json</td>
+          <td class="yes">Yes</td>
+          <td class="partial">~10-25s</td>
+          <td class="no">No (daemon)</td>
+          <td class="yes">Yes</td>
         </tr>
         <tr>
-          <td>OpenAI JSON mode</td>
-          <td class="partial">Mostly</td>
+          <td>OpenAI Structured Outputs</td>
           <td class="yes">Yes</td>
+          <td class="yes">~1-3s</td>
+          <td class="no">No (API)</td>
           <td class="no">No</td>
-          <td class="partial">Partial</td>
         </tr>
         <tr>
           <td>Outlines / Guidance</td>
           <td class="yes">Yes</td>
+          <td class="yes">~1-3s</td>
           <td class="yes">Yes</td>
           <td class="yes">Yes</td>
-          <td class="yes">Yes</td>
+        </tr>
+        <tr>
+          <td>Retry + parse</td>
+          <td class="no">No</td>
+          <td class="no">Varies</td>
+          <td class="partial">Depends</td>
+          <td class="partial">Depends</td>
         </tr>
       </tbody>
     </table>
-    <p style="color: var(--text-muted); margin-top: 16px; font-size: 14px;">llmjson is purpose-built for Apple Silicon with MLX. Outlines and Guidance target CUDA/CPU. Pick the right tool for your hardware.</p>
+    <p style="color: var(--text-muted); margin-top: 16px; font-size: 14px;">llmjson is a Python library for Apple Silicon (MLX). No server process, no API keys. Outlines and Guidance target CUDA/CPU. Ollama works but runs 10-25x slower on the same hardware. Pick the right tool for your setup.</p>
   </section>
 
   <section class="compare" style="padding-top: 0; border-top: none;">
@@ -420,7 +427,7 @@ <h2>CLI included</h2>
     <h2>Get started in 30 seconds</h2>
     <p>Requires Python 3.9+ and Apple Silicon (M1/M2/M3/M4).</p>
     <a href="https://github.com/fatdinhero/llmjson" class="btn">GitHub</a>
-    <a href="https://pypi.org/project/jsongate/" class="btn btn-secondary">PyPI</a>
+    <a href="https://pypi.org/project/llmjson/" class="btn btn-secondary">PyPI</a>
     <a href="blog/how-i-got-100-percent-valid-json.html" class="btn btn-secondary">Blog Post</a>
   </section>