ClozeCongruence: Human-Style Essay Classification vs. AI-Written Text via Multi-Horizon Semantic Infilling and Calibrated Stylometric Manifolds

Distinguishing authentic human-authored essays from machine-generated prose remains an active challenge for educational integrity, admissions authenticity, and digital forensics. Existing zero-shot detectors either rely on inaccessible white-box token logits (DetectGPT, Fast-DetectGPT) or penalize formal, non-native human writing styles through uncalibrated perplexity metrics. Furthermore, prior forensic benchmarks frequently evaluate specialized technical preprints (e.g., ArXiv, PeerRead) or truncate samples to 200 words, destroying macro-discursive structure. In this paper, we present ClozeCongruence, a controlled forensic evaluation framework evaluated on a curated benchmark of N = 6,997 samples dedicated to full-length human-style essay classification vs. AI-written text (mean length ≈ 640 words, up to 1,500 words). Our methodology evaluates five core empirical questions under zero-target-leakage protocols (verified via label permutation tests yielding null AUROCs of 0.50 ± 0.02): Pairwise Prompt Matching & Calibrated Controls: With N = 3,499 verified human negative controls, the dataset satisfies the statistical Rule of Three (N ≥ 3/α) required to calibrate False Positive Rates down to 0.1%. In-Domain Essay Discrimination: On clean full essays (N = 6,000), zero-shot stylometric scoring achieves an AUROC of 0.8547 [95% CI: 0.8443–0.8642] and AUPRC of 0.8520, yielding True Positive Rates of 43.27% at 5.0% FPR, 20.17% at 1.0% FPR, and 5.30% at 0.1% FPR (a 75-fold sensitivity gain over 200-word truncated baselines). When calibrated through a 5-fold cross-validated supervised manifold ensemble, discrimination reaches AUROC 0.9651 with 52.17% TPR at 1.0% FPR and 15.87% TPR at 0.1% FPR. In isolated ablation testing, single-scalar paragraph discourse flow variance achieves AUROC 0.5343 (Cohen's d = 0.105), confirming that discourse progression acts as an integrative feature rather than a standalone classifier. Full RAID Adversarial Suite (11 Attacks, N = 6,000): Pre-inference canonicalization mitigates surface attacks (homoglyph substitution recovers from 0.6793 to 0.9649 AUROC, Δ = +0.2856; clean baseline maintains 0.9799 AUROC), while establishing the empirical failure boundary against 11B DIPPER neural paraphrasing (0.6956 AUROC, 13.2% TPR @ 1% FPR; overall suite AUROC = 0.9459). Cross-Domain & Multilingual Transfer (M4 Benchmark, N = 4,000): Evaluated across 8 strata and 6 languages, the framework demonstrates robust performance on structured academic English (PeerRead AUROC = 0.9977, 92.0% TPR @ 1% FPR; Reddit AUROC = 0.8660) while characterizing script attenuation across Arabic (0.8706), German (0.8008), Chinese (0.7143), Urdu (0.7038), and Russian (0.6700) without language-specific tuning (Full M4 AUROC = 0.8165). Chimeric Co-Authorship Localization (WildChat, N = 2,000): Across 9,995 paragraph segments, the sliding-window localizer achieves 0.8979 paragraph classification AUROC, exact transition localization F1 of 0.6789 (±0 paragraphs), and slack F1 of 0.7517 (±1 paragraph), reaching 100.0% Precision and 0.8322 F1 on alternating co-authorship.