Datasets:
upload paper.html
Browse files- paper.html +11 -6
paper.html
CHANGED
|
@@ -4,7 +4,7 @@
|
|
| 4 |
<meta charset="utf-8" />
|
| 5 |
<meta name="generator" content="pandoc" />
|
| 6 |
<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes" />
|
| 7 |
-
<title>ContentOS Preprint v1.0.
|
| 8 |
<style>
|
| 9 |
/* Default styles provided by pandoc.
|
| 10 |
** See https://pandoc.org/MANUAL.html#variables-for-html for config info.
|
|
@@ -238,7 +238,7 @@
|
|
| 238 |
</head>
|
| 239 |
<body>
|
| 240 |
<header id="title-block-header">
|
| 241 |
-
<h1 class="title">ContentOS Preprint v1.0.
|
| 242 |
</header>
|
| 243 |
<h1
|
| 244 |
id="contentos-a-reproducible-bilingual-ai-text-detection-ensemble-with-adversarial-robustness-evaluation">ContentOS:
|
|
@@ -270,8 +270,12 @@ Wrong-rate of 4% and median latency of 1.2 seconds on commodity 8-vCPU
|
|
| 270 |
hardware. Earlier 44-text hand-curated smoke (v1.0 paper measurement)
|
| 271 |
reported 0.821 EN / 0.837 RU; the 4× expanded battery with proper class
|
| 272 |
balance per (lang, genre) cell stabilized the numbers upward.</p>
|
| 273 |
-
<p>On the 300-sample adversarial paired set
|
| 274 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 275 |
<p>The contribution of this work is <strong>field-leading
|
| 276 |
reproducibility</strong>, not state-of-the-art absolute AUROC. Anyone
|
| 277 |
can clone the repository, run the regression test in 0.05 seconds, and
|
|
@@ -323,8 +327,9 @@ hardware ($25/month) within 90 minutes:</p>
|
|
| 323 |
expanded smoke, 2026-04-29)</li>
|
| 324 |
<li><strong>Russian ensemble OOD AUROC: 0.846</strong> (176-sample
|
| 325 |
expanded smoke, 2026-04-29)</li>
|
| 326 |
-
<li><strong>English ensemble adversarial AUROC: 0.
|
| 327 |
-
300-sample paraphrase-paired set
|
|
|
|
| 328 |
<li><strong>English ensemble p50 latency: 1.2 seconds</strong> (8-core
|
| 329 |
CPU, no GPU)</li>
|
| 330 |
</ul>
|
|
|
|
| 4 |
<meta charset="utf-8" />
|
| 5 |
<meta name="generator" content="pandoc" />
|
| 6 |
<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes" />
|
| 7 |
+
<title>ContentOS Preprint v1.0.2</title>
|
| 8 |
<style>
|
| 9 |
/* Default styles provided by pandoc.
|
| 10 |
** See https://pandoc.org/MANUAL.html#variables-for-html for config info.
|
|
|
|
| 238 |
</head>
|
| 239 |
<body>
|
| 240 |
<header id="title-block-header">
|
| 241 |
+
<h1 class="title">ContentOS Preprint v1.0.2</h1>
|
| 242 |
</header>
|
| 243 |
<h1
|
| 244 |
id="contentos-a-reproducible-bilingual-ai-text-detection-ensemble-with-adversarial-robustness-evaluation">ContentOS:
|
|
|
|
| 270 |
hardware. Earlier 44-text hand-curated smoke (v1.0 paper measurement)
|
| 271 |
reported 0.821 EN / 0.837 RU; the 4× expanded battery with proper class
|
| 272 |
balance per (lang, genre) cell stabilized the numbers upward.</p>
|
| 273 |
+
<p>On the 300-sample adversarial paired set (cross-model paraphrasing
|
| 274 |
+
attack, OOD-augmented baseline), ensemble AUROC reaches
|
| 275 |
+
<strong>0.998</strong> (re-measured 2026-04-29 with current
|
| 276 |
+
calibration). Earlier v1.0 paper measurement was 0.985 — the slight
|
| 277 |
+
increase reflects the intervening calibration tuning between Gap-7 and
|
| 278 |
+
current state.</p>
|
| 279 |
<p>The contribution of this work is <strong>field-leading
|
| 280 |
reproducibility</strong>, not state-of-the-art absolute AUROC. Anyone
|
| 281 |
can clone the repository, run the regression test in 0.05 seconds, and
|
|
|
|
| 327 |
expanded smoke, 2026-04-29)</li>
|
| 328 |
<li><strong>Russian ensemble OOD AUROC: 0.846</strong> (176-sample
|
| 329 |
expanded smoke, 2026-04-29)</li>
|
| 330 |
+
<li><strong>English ensemble adversarial AUROC: 0.998</strong> on
|
| 331 |
+
300-sample paraphrase-paired OOD-augmented set (re-measured
|
| 332 |
+
2026-04-29)</li>
|
| 333 |
<li><strong>English ensemble p50 latency: 1.2 seconds</strong> (8-core
|
| 334 |
CPU, no GPU)</li>
|
| 335 |
</ul>
|