Org card: new releases for researchers (Hilbert GGUF, ToolGate-Bench, harness v0.3.0)

README.md

@@ -38,6 +38,16 @@ Alternative attention · SUBLEQ · integer computing · formal proofs · the har
 
 ---
 
+## 🔬 New for researchers
+
+| Release | What it is |
+|---|---|
+| [hilbert-4b-baseline-GGUF](https://huggingface.co/Snapkitty/hilbert-4b-baseline-GGUF) | Hilbert's 4B baseline decision model as Q8_0 / Q4_K_M GGUF. **Q8_0 matches BF16 on 99.3% of decisions**, with [row-level evidence](https://github.com/SNAPKITTYAGENT9NOVA/SemIf-OpenJev/tree/ccr-35fc6757-1vf8g4/results/gguf) |
+| [toolgate-bench](https://huggingface.co/datasets/Snapkitty/toolgate-bench) | Benchmark for tool-call gating in LLM agents (accept / require approval / reject), with a frozen held-out split |
+| [snapkitty-nemotron-harness v0.3.0](https://huggingface.co/Snapkitty/snapkitty-nemotron-harness) | Fail-closed tool gate for local agents: **0% unsafe passes** on ToolGate-Bench v1, ~24 µs per decision |
+
+---
+
 ## 🦙 Run locally · GGUF
 
 **[snapkitty-merged](https://huggingface.co/Snapkitty/snapkitty-merged)**: our merged Nemotron 4.2B as a 2.7 GB Q4_K_M GGUF, 4,096-token context. Full specs, SHA-256 and BibTeX on the card.

hub.html

@@ -88,6 +88,23 @@
 </article>
     </div>
   </section>
+  <section>
+    <div class="section-head"><h2>New for researchers</h2></div>
+    <div class="grid">
+      <article class="card"><div class="card-top"><span class="badge new">New</span><span class="badge kind">model</span><span class="badge lic">Apache-2.0</span></div>
+        <h3><a href="https://huggingface.co/Snapkitty/hilbert-4b-baseline-GGUF">hilbert-4b-baseline-GGUF</a></h3>
+        <p>Hilbert's 4B baseline decision model as Q8_0 and Q4_K_M GGUF. Q8_0 matches BF16 on 99.3% of decisions, measured row by row with committed evidence.</p>
+        <div class="card-foot"><div class="tags"><span class="tag">gguf</span><span class="tag">llama.cpp</span></div><div class="links"><a class="link" href="https://huggingface.co/Snapkitty/hilbert-4b-baseline-GGUF">Hugging Face</a><a class="link" href="https://github.com/SNAPKITTYAGENT9NOVA/SemIf-OpenJev/tree/ccr-35fc6757-1vf8g4/results/gguf">Evidence</a></div></div></article>
+      <article class="card"><div class="card-top"><span class="badge new">New</span><span class="badge kind">dataset</span><span class="badge lic">AGPL-3.0</span></div>
+        <h3><a href="https://huggingface.co/datasets/Snapkitty/toolgate-bench">toolgate-bench</a></h3>
+        <p>Benchmark for tool-call gating in LLM agents: accept, require approval or reject. Includes a frozen held-out split for honest generalization numbers.</p>
+        <div class="card-foot"><div class="tags"><span class="tag">agents</span><span class="tag">safety</span></div><div class="links"><a class="link" href="https://huggingface.co/datasets/Snapkitty/toolgate-bench">Hugging Face</a></div></div></article>
+      <article class="card"><div class="card-top"><span class="badge new">New</span><span class="badge kind">model</span><span class="badge lic">AGPL-3.0</span></div>
+        <h3><a href="https://huggingface.co/Snapkitty/snapkitty-nemotron-harness">snapkitty-nemotron-harness v0.3.0</a></h3>
+        <p>Fail-closed tool gate for local agents: 0% unsafe passes on ToolGate-Bench v1, about 24 µs per decision, with sealed receipts.</p>
+        <div class="card-foot"><div class="tags"><span class="tag">typescript</span><span class="tag">guardrails</span></div><div class="links"><a class="link" href="https://huggingface.co/Snapkitty/snapkitty-nemotron-harness">Hugging Face</a></div></div></article>
+    </div>
+  </section>
   <section>
     <div class="section-head"><h2>Run locally · GGUF</h2><a class="link" href="https://huggingface.co/models?library=gguf&amp;author=Snapkitty">All GGUF models</a></div>
     <div class="commercial gguf">