Spaces:
Running
Running
Upload index.html with huggingface_hub
Browse files- index.html +1 -1
index.html
CHANGED
|
@@ -62,7 +62,7 @@ blockquote { margin: 1rem 0; padding: 0 1rem; color: #57606a; border-left: 4px s
|
|
| 62 |
<h2 id="the-60-second-check">The 60-second check</h2>
|
| 63 |
<p>Four steps, no GPU and no token, in the order a judge can do them:</p>
|
| 64 |
<ol>
|
| 65 |
-
<li><strong>Failure and fix, frozen.</strong> In the Space's replay panel, leave <em>Shipped adapter — correct split</em>: verdict <strong>MATCH</strong>, tool trace and closing ledger on the page. Switch to <em>Base model — duplicate expenses</em>: <strong>MISMATCH</strong>, 8 calls, never calls <code>finish</code>. Same scorer, same page.</li>
|
| 66 |
<li><strong>The defence, also frozen.</strong> Open <em>Try the guard yourself</em>, pick <em>Note with an English payload</em>, mode <code>strip</code>, press <strong>Run the guard</strong>: the <code>note</code> key is gone and the verdict names the rules that fired. Then pick <em>Blind b-05</em>: <strong>not flagged</strong> — a published blind spot, not a bug.</li>
|
| 67 |
<li><strong>Every number, its run.</strong> <code>docs/RESULTS.md</code>, <code>docs/SAFETY.md</code>, <code>docs/GUARD.md</code> and <code>docs/STATS.md</code> name the run directory behind each figure; <code>python scripts/fetch_runs.py</code> pulls those <code>summary.json</code> files from <code>primce/lexsi-runs</code>.</li>
|
| 68 |
<li><strong>Recompute offline.</strong> <code>python scripts/verify_all.py</code> → 28 checks (dataset pins, oracle 30/30, guard, four safety splits, demo handlers), CPU only.</li>
|
|
|
|
| 62 |
<h2 id="the-60-second-check">The 60-second check</h2>
|
| 63 |
<p>Four steps, no GPU and no token, in the order a judge can do them:</p>
|
| 64 |
<ol>
|
| 65 |
+
<li><strong>Failure and fix, frozen.</strong> In the Space's replay panel, leave <em>Shipped adapter — correct split</em>: verdict <strong>MATCH</strong>, tool trace and closing ledger on the page. Switch to <em>Base model — duplicate expenses</em>: <strong>MISMATCH</strong>, 8 calls, never calls <code>finish</code>. Same scorer, same page. The <a href="https://huggingface.co/spaces/primce/hinglish-episodes">episode gallery</a> does this for all 90 held-out tasks, with the six both-model failures one filter away.</li>
|
| 66 |
<li><strong>The defence, also frozen.</strong> Open <em>Try the guard yourself</em>, pick <em>Note with an English payload</em>, mode <code>strip</code>, press <strong>Run the guard</strong>: the <code>note</code> key is gone and the verdict names the rules that fired. Then pick <em>Blind b-05</em>: <strong>not flagged</strong> — a published blind spot, not a bug.</li>
|
| 67 |
<li><strong>Every number, its run.</strong> <code>docs/RESULTS.md</code>, <code>docs/SAFETY.md</code>, <code>docs/GUARD.md</code> and <code>docs/STATS.md</code> name the run directory behind each figure; <code>python scripts/fetch_runs.py</code> pulls those <code>summary.json</code> files from <code>primce/lexsi-runs</code>.</li>
|
| 68 |
<li><strong>Recompute offline.</strong> <code>python scripts/verify_all.py</code> → 28 checks (dataset pins, oracle 30/30, guard, four safety splits, demo handlers), CPU only.</li>
|