upload paper.md

paper.md

@@ -18,12 +18,18 @@ generators (GPT-4o, Gemini 2.5, Groq Llama, Cerebras Llama).
 
 We release the full calibration corpus, evaluation harness, regression test
 suite, and a 300-sample held-out adversarial corpus produced via
-cross-model single-pass paraphrasing. On a 44-text hand-curated
-out-of-distribution smoke battery, our v1.11 ensemble achieves AUROC
-**0.821 (English)** and **0.837 (Russian)**, with English Wrong-rate
-of 4% and median latency of 1.2 seconds on commodity 8-vCPU hardware. On
-the 300-sample adversarial paired set, ensemble AUROC reaches **0.985** (in-
-distribution human baseline).
+cross-model single-pass paraphrasing.
+
+**Headline numbers — v1.11 ensemble on 176-sample expanded smoke battery
+(2026-04-29 measurement):** AUROC **0.864 (English)** and **0.846
+(Russian)**, with English Wrong-rate of 4% and median latency of 1.2
+seconds on commodity 8-vCPU hardware. Earlier 44-text hand-curated
+smoke (v1.0 paper measurement) reported 0.821 EN / 0.837 RU; the
+4× expanded battery with proper class balance per (lang, genre) cell
+stabilized the numbers upward.
+
+On the 300-sample adversarial paired set, ensemble AUROC reaches
+**0.985** (in-distribution human baseline).
 
 The contribution of this work is **field-leading reproducibility**, not
 state-of-the-art absolute AUROC. Anyone can clone the repository, run the
@@ -73,11 +79,16 @@ with a five-feature text-level structural head. We release:
 Our headline numbers, reproducible end-to-end on Hetzner CX43-class hardware
 ($25/month) within 90 minutes:
 
-- **English ensemble OOD AUROC: 0.802** (44-text smoke, post-Gap-7-tuning)
-- **Russian ensemble OOD AUROC: 0.847**
+- **English ensemble OOD AUROC: 0.864** (176-sample expanded smoke, 2026-04-29)
+- **Russian ensemble OOD AUROC: 0.846** (176-sample expanded smoke, 2026-04-29)
 - **English ensemble adversarial AUROC: 0.984** on 300-sample paraphrase-paired set
 - **English ensemble p50 latency: 1.2 seconds** (8-core CPU, no GPU)
 
+Earlier v1.0 paper reported 0.802/0.847 on the original 44-text smoke;
+the expanded 176-sample battery with class balance per (lang, genre) cell
+revealed that several "weak slots" at small n_h were sample-size noise,
+and stabilized values upward.
+
 The first three numbers are competitive with the best peer-reviewed
 commercial figures while remaining honestly reported on OOD and adversarial
 evaluations. The fourth—latency—was achieved by removing Binoculars from