docs: qualify Part 2 result and fix GPU

README.md

@@ -19,7 +19,8 @@ live experimental run.
 - **Part 1 — Scar-Survival** — a memorized LLM error, corrected: how durable is the fix?
   · [Space](https://huggingface.co/spaces/Laborator/scar-survival)
   · [GitHub](https://github.com/SergheiBrinza/scar-survival)
-- **Part 2 — External Grounding** — lifting self-correction 50% → 100% under a noisy notebook
+- **Part 2 — External Grounding** — lifting self-correction 50% → 92.9% under a noisy notebook,
+  and to 100% once one ambiguous trap (#46) is reformulated
   · [Space](https://huggingface.co/spaces/Laborator/external-grounding)
   · [GitHub](https://github.com/SergheiBrinza/external-grounding)
 - **Part 3 — Thin Channel** — how rarely external truth can arrive before calibration collapses

index.html

@@ -142,9 +142,10 @@ a{color:var(--gold);text-decoration:none;}
     <div class="card">
       <div class="num">Part 2</div>
       <div class="name">External Grounding</div>
-      <div class="hook">Lifting self-correction from 50% to 100% under a noisy notebook. Drag the
+      <div class="hook">Lifting self-correction from 50% to 92.9% under a noisy notebook, and to 100%
+        once one ambiguous trap (#46) is reformulated. Drag the
         guardian through six versions and watch which traps get fixed — and which regress.</div>
-      <div class="key">50% → 100% · clone ceiling 64.3% · +fixed / −broken</div>
+      <div class="key">50% → 92.9% (100% after #46) · clone ceiling 64.3% · +fixed / −broken</div>
       <div class="links">
         <a class="space" href="https://huggingface.co/spaces/Laborator/external-grounding" target="_blank" rel="noopener">Open Space ↗</a>
         <a class="gh" href="https://github.com/SergheiBrinza/external-grounding" target="_blank" rel="noopener">GitHub ↗</a>
@@ -175,7 +176,7 @@ a{color:var(--gold);text-decoration:none;}
     </div>
     <div class="attrib">
       Subject model Qwen2.5-3B-Instruct · independent judge Qwen2.5-7B-Instruct (both Apache-2.0,
-      Alibaba Cloud). Run on a single RTX 3090. No model weights are redistributed here. Demo code: MIT.
+      Alibaba Cloud). Run on a single RTX 3090 Ti. No model weights are redistributed here. Demo code: MIT.
     </div>
   </footer>