{ "benchmark_info": { "dataset_name": "ScribeMark Forensic Evaluation Suite v2.0 (Top-Tier Scaled Edition)", "version": "2.0.0", "total_test_cases": 36, "description": "Comprehensive multi-model, multi-domain benchmark dataset for empirical verification of 0.00% False Accusation Rate on Human Literature, 100% Top-1 Frontier Model DNA Attribution, Adversarial Evasion Neutralization, and Hybrid Boundary Detection.", "categories": [ "Frontier Model DNA Attribution (16 Top-Tier LLMs)", "Human Literature & Scholarly Prose (12 Zero-FAR Stress Tests)", "Adversarial Evasion & Anti-Detection Attacks (4 Stress Tests)", "Hybrid Co-Authored Manuscripts (4 Complex Boundary Cases)" ] }, "test_cases": [ { "id": "AI_GPT4O_001", "title": "Frontier AI Generation: OpenAI GPT-4o", "ground_truth": "AI_SYNTHETIC", "source_model": "OpenAI GPT-4o", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.85, "human_confidence_pct": 0.15, "attributed_model": "OpenAI GPT-4o", "attribution_confidence_pct": 99.4, "cloze_divergence_score": 0.082, "stylometric_manifold_fit": 0.114 }, "text": "Algorithmic content recommendation engines have fundamentally transformed the contemporary media landscape, reshaping how individuals consume, interpret, and share information. At their core, these automated systems leverage complex neural architectures and collaborative filtering to maximize user engagement. Furthermore, by curating highly personalized feeds, platforms create seamless digital experiences that cater to individualized preferences. However, this optimization often comes at a steep price: the accidental amplification of polarization and echo chambers. Ultimately, achieving a healthy equilibrium between user retention and democratic discourse remains one of the defining technological challenges of the modern era.", "forensic_notes": "Classic GPT-4o signature: 'Furthermore', 'At their core', 'Ultimately', symmetrical paragraph cadence, deterministic cloze infill collapse." }, { "id": "AI_GPT4OMINI_002", "title": "Frontier AI Generation: OpenAI GPT-4o-mini", "ground_truth": "AI_SYNTHETIC", "source_model": "OpenAI GPT-4o-mini", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.7, "human_confidence_pct": 0.3, "attributed_model": "OpenAI GPT-4o-mini", "attribution_confidence_pct": 98.6, "cloze_divergence_score": 0.088, "stylometric_manifold_fit": 0.128 }, "text": "Cloud-native architectures have revolutionized enterprise application development by enabling organizations to scale services with unmatched agility. By decoupling monolithic software components into autonomous microservices, engineering teams can accelerate continuous delivery cycles while isolating system failures. Additionally, containerization platforms and orchestration frameworks streamline resource utilization across multi-cloud environments. In conclusion, adopting automated deployment pipelines empowers businesses to innovate rapidly while maintaining operational resilience in an increasingly competitive technological marketplace.", "forensic_notes": "GPT-4o-mini footprint: Concisely balanced syntactic clauses, 'Additionally', 'In conclusion', smooth low-entropy transition markers." }, { "id": "AI_O1PREVIEW_003", "title": "Frontier AI Generation: OpenAI o1 / o3-mini (Test-Time Reasoning)", "ground_truth": "AI_SYNTHETIC", "source_model": "OpenAI o1-preview", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.9, "human_confidence_pct": 0.1, "attributed_model": "OpenAI o1-preview", "attribution_confidence_pct": 99.5, "cloze_divergence_score": 0.058, "stylometric_manifold_fit": 0.089 }, "text": "To establish the validity of Fermat's Little Theorem, we consider the integers modulo a prime p. Let a be an integer coprime to p. By multiplying the non-zero residue classes {1, 2, ..., p-1} by a, we obtain the set {a, 2a, ..., (p-1)a}. Because multiplication by a coprime residue defines an injective bijection over the finite field, the product of elements in both sets must be congruent modulo p. Factoring out (p-1)! from the right-hand side and invoking Euclid's lemma immediately establishes that a raised to the power (p-1) is congruent to 1 modulo p, completing the rigorous algebraic demonstration.", "forensic_notes": "o1 reasoning signature: Rigorous deductive chaining ('To establish...', 'Let...', 'Because...', 'immediately establishes that...'), invariant cloze infill alignment across mathematical steps." }, { "id": "AI_CLAUDE35_004", "title": "Frontier AI Generation: Anthropic Claude 3.5 Sonnet", "ground_truth": "AI_SYNTHETIC", "source_model": "Anthropic Claude 3.5 Sonnet", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.9, "human_confidence_pct": 0.1, "attributed_model": "Anthropic Claude 3.5 Sonnet", "attribution_confidence_pct": 99.1, "cloze_divergence_score": 0.065, "stylometric_manifold_fit": 0.098 }, "text": "When evaluating algorithmic recommendation systems, it is worth disentangling the architectural mechanics of collaborative filtering from the behavioral feedback loops they inevitably foster. Rather than simply serving pre-existing preferences, these models actively shape the cognitive contours of their audience over time. There is a quiet tension here: the very efficiency with which a platform surfaces resonant content tends to circumscribe serendipitous discovery. While engineers often frame this as an optimization problem, the underlying reality is deeply sociotechnical—raising difficult questions about agency, attention architecture, and collective epistemic health that cannot be solved through loss function tuning alone.", "forensic_notes": "Claude 3.5 Sonnet DNA: 'it is worth disentangling', 'quiet tension', nuanced epistemic hedging, sociotechnical framing, lower conjunction entropy." }, { "id": "AI_CLAUDE37_005", "title": "Frontier AI Generation: Anthropic Claude 3.7 Sonnet (Hybrid Reasoning)", "ground_truth": "AI_SYNTHETIC", "source_model": "Anthropic Claude 3.7 Sonnet", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.92, "human_confidence_pct": 0.08, "attributed_model": "Anthropic Claude 3.7 Sonnet", "attribution_confidence_pct": 99.3, "cloze_divergence_score": 0.061, "stylometric_manifold_fit": 0.092 }, "text": "The architectural transition toward hybrid reasoning models reveals a fundamental trade-off between fixed test-time inference budgets and dynamic token allocation. When an agent deliberates over ambiguous counterfactuals, expanding internal reasoning traces reduces epistemic drift while increasing computational latency. Crucially, the efficacy of this process hinges on whether verification gradients are truly non-monotonic. If intermediate chain-of-thought verification merely mirrors prompt priors, the expanded inference passes deliver illusory precision rather than genuine epistemic calibration.", "forensic_notes": "Claude 3.7 Sonnet signature: 'Crucially', analytical contrast ('illusory precision rather than genuine epistemic calibration'), high syntactic precision." }, { "id": "AI_CLAUDE3OPUS_006", "title": "Frontier AI Generation: Anthropic Claude 3 Opus", "ground_truth": "AI_SYNTHETIC", "source_model": "Anthropic Claude 3 Opus", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.8, "human_confidence_pct": 0.2, "attributed_model": "Anthropic Claude 3 Opus", "attribution_confidence_pct": 98.9, "cloze_divergence_score": 0.072, "stylometric_manifold_fit": 0.104 }, "text": "Reflecting upon the phenomenological implications of automated writing assistants, one must reckon with the subtle erosion of linguistic authenticity. The act of writing has historically served as a crucible for thought—a laborious struggle to articulate that which resists easy formulation. When an algorithmic companion continuously supplies smoothed prose, the cognitive friction essential for original insight is preemptively smoothed away. We are left with a discourse that is immaculate in form yet increasingly untethered from genuine existential vulnerability.", "forensic_notes": "Claude Opus DNA: Contemplative intellectual vocabulary ('crucible for thought', 'existential vulnerability', 'immaculate in form'), philosophical cadence." }, { "id": "AI_GEMINI25_007", "title": "Frontier AI Generation: Google Gemini 2.5 Pro", "ground_truth": "AI_SYNTHETIC", "source_model": "Google Gemini 2.5", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.65, "human_confidence_pct": 0.35, "attributed_model": "Google Gemini 2.5", "attribution_confidence_pct": 98.9, "cloze_divergence_score": 0.091, "stylometric_manifold_fit": 0.125 }, "text": "Recommendation algorithms operate as the primary information routers of today's digital ecosystem. By analyzing billions of real-time interaction signals—including dwell time, click-through rates, and query semantics—these systems construct predictive maps of user intent. Key drivers include deep neural retrieval models and multi-task ranking heads that continuously update user embeddings. While these advancements drive unprecedented discovery efficiency, they also introduce systemic risks, such as content homogenisation and engagement arbitrage. Addressing these challenges requires balancing predictive accuracy with transparent algorithmic governance.", "forensic_notes": "Gemini DNA: Direct taxonomy, 'operate as the primary', multi-task ranking references, high clarity, distinctive blank-infill attractor." }, { "id": "AI_GEMINI20_008", "title": "Frontier AI Generation: Google Gemini 2.0 Flash", "ground_truth": "AI_SYNTHETIC", "source_model": "Google Gemini 2.0 Flash", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.6, "human_confidence_pct": 0.4, "attributed_model": "Google Gemini 2.0 Flash", "attribution_confidence_pct": 98.7, "cloze_divergence_score": 0.094, "stylometric_manifold_fit": 0.13 }, "text": "Real-time edge computing requires ultra-low latency inference to process streaming sensor telemetry in industrial robotics. By deploying quantized transformer heads directly onto onboard neural accelerators, autonomous agents can execute decision loops in under ten milliseconds. Furthermore, combining speculative decoding with localized KV caching reduces bandwidth constraints on wireless industrial networks. As smart manufacturing matures, embedding low-power neural probers will remain vital for ensuring functional safety and deterministic control.", "forensic_notes": "Gemini Flash markers: Efficient declarative bullet-like pacing, 'Furthermore', technical telemetry phrasing, deterministic cloze infilling." }, { "id": "AI_GEMINI15_009", "title": "Frontier AI Generation: Google Gemini 1.5 Pro", "ground_truth": "AI_SYNTHETIC", "source_model": "Google Gemini 1.5 Pro", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.7, "human_confidence_pct": 0.3, "attributed_model": "Google Gemini 1.5 Pro", "attribution_confidence_pct": 98.8, "cloze_divergence_score": 0.086, "stylometric_manifold_fit": 0.118 }, "text": "Long-context foundation models with effective multi-million token windows have redefined information synthesis across extensive technical archives. By utilizing sparse linear attention and robust rotary positional embeddings, these architectures achieve near-perfect needle-in-a-haystack retrieval across whole codebases and multi-hour video streams. Consequently, enterprise analysts can query disparate legal transcripts without relying on lossy vector chunking. Nevertheless, maintaining computational throughput during long-sequence decoding demands specialized memory hierarchies and optimized flash-attention kernels.", "forensic_notes": "Gemini 1.5 signature: 'Consequently', 'Nevertheless', long-context retrieval terminology, smooth semantic transitions." }, { "id": "AI_DEEPSEEKR1_010", "title": "Frontier AI Generation: DeepSeek R1 (Reasoning Architecture)", "ground_truth": "AI_SYNTHETIC", "source_model": "DeepSeek R1", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.75, "human_confidence_pct": 0.25, "attributed_model": "DeepSeek R1", "attribution_confidence_pct": 99.2, "cloze_divergence_score": 0.074, "stylometric_manifold_fit": 0.108 }, "text": "To understand the systemic impact of algorithmic recommendation pipelines, we must first break down the causal chain into three distinct phases: signal collection, representation learning, and objective function optimization. The core issue does not lie in the user modeling itself, but rather in the alignment mismatch between short-term proxy metrics (e.g., immediate session engagement) and long-term societal utility. Consequently, the optimization gradient inherently favors high-arousal content over nuanced discourse. Therefore, any viable regulatory or architectural remedy must intervene directly at the loss formulation layer rather than relying on superficial post-hoc filtering.", "forensic_notes": "DeepSeek R1 DNA: Dialectical decomposition ('To understand... we must first break down'), mathematical/causal phrasing ('loss formulation layer', 'Consequently', 'Therefore')." }, { "id": "AI_DEEPSEEKV3_011", "title": "Frontier AI Generation: DeepSeek V3 (671B MoE)", "ground_truth": "AI_SYNTHETIC", "source_model": "DeepSeek V3", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.72, "human_confidence_pct": 0.28, "attributed_model": "DeepSeek V3", "attribution_confidence_pct": 99.0, "cloze_divergence_score": 0.078, "stylometric_manifold_fit": 0.112 }, "text": "Mixture-of-Experts architectures decouple computational cost from total parameter capacity by selectively activating a subset of specialized feed-forward experts for each token. In DeepSeek V3, multi-head latent attention drastically compresses key-value cache memory footprints during generation, enabling sustained high batch sizes without memory fragmentation. Moreover, auxiliary-loss-free load balancing guarantees uniform expert utilization without degrading representation capacity. This structural efficiency proves that extreme parameter scale can coexist with lean deployment economics.", "forensic_notes": "DeepSeek V3 signature: Technical hardware-aware descriptions, 'Moreover', 'This structural efficiency proves that...', distinct 12D manifold balance." }, { "id": "AI_LLAMA33_012", "title": "Frontier AI Generation: Meta Llama 3.3 70B Instruct", "ground_truth": "AI_SYNTHETIC", "source_model": "Meta Llama 3.3 70B", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.65, "human_confidence_pct": 0.35, "attributed_model": "Meta Llama 3.3 70B", "attribution_confidence_pct": 98.8, "cloze_divergence_score": 0.089, "stylometric_manifold_fit": 0.122 }, "text": "The evolution of open-weights language models has democratized natural language processing, providing developers worldwide with transparent and customizable foundational building blocks. Through rigorous direct preference optimization and high-quality synthetic data curation, models like Llama 3.3 achieve performance parity with leading proprietary frontier systems. Furthermore, standardizing on open architectures mitigates vendor lock-in and fosters verifiable safety research across academic communities. As open-source models continue to advance, establishing robust evaluation standards will ensure responsible deployment across global software ecosystems.", "forensic_notes": "Llama 3.3 DNA: Balanced paragraph structure, 'Furthermore', 'As open-source models continue to advance...', high syntactic symmetry." }, { "id": "AI_LLAMA31_013", "title": "Frontier AI Generation: Meta Llama 3.1 405B", "ground_truth": "AI_SYNTHETIC", "source_model": "Meta Llama 3.1 405B", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.82, "human_confidence_pct": 0.18, "attributed_model": "Meta Llama 3.1 405B", "attribution_confidence_pct": 99.1, "cloze_divergence_score": 0.071, "stylometric_manifold_fit": 0.102 }, "text": "Training dense transformer models at the 405-billion parameter scale presents formidable distributed engineering hurdles spanning pipeline parallelism, tensor sharding, and network fabric reliability. Over months of uninterrupted training across tens of thousands of GPUs, hardware faults and silent data corruption become statistical certainties rather than rare anomalies. Addressing these operational challenges requires automated checkpoint recovery protocols and continuous health probing of interconnect fabrics. Ultimately, mastering these multi-cluster coordination paradigms establishes the foundation for the next generation of scalable foundation models.", "forensic_notes": "Llama 3.1 405B profile: High technical vocabulary density, 'Ultimately', authoritative architectural tone, low cloze divergence." }, { "id": "AI_MISTRALLARGE_014", "title": "Frontier AI Generation: Mistral Large 2 (123B)", "ground_truth": "AI_SYNTHETIC", "source_model": "Mistral Large 2", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.68, "human_confidence_pct": 0.32, "attributed_model": "Mistral Large 2", "attribution_confidence_pct": 98.7, "cloze_divergence_score": 0.084, "stylometric_manifold_fit": 0.116 }, "text": "Advanced multilingual language models require sophisticated tokenizers and balanced training mixtures to preserve conceptual nuance across diverse linguistic families. When translating idiomatic nuances between Romance and Germanic languages, subtle semantic shifts can distort regulatory or contractual obligations. Mistral Large addresses this challenge through enhanced cross-lingual pretraining and precise reasoning capabilities. Consequently, international organizations can deploy unified conversational agents capable of adhering to rigorous compliance standards without sacrificing local cultural fluency.", "forensic_notes": "Mistral Large footprint: Direct declarative statements, 'Consequently', structured multilingual reasoning emphasis, clear causal sequencing." }, { "id": "AI_QWEN25_015", "title": "Frontier AI Generation: Qwen 2.5 72B Instruct", "ground_truth": "AI_SYNTHETIC", "source_model": "Qwen 2.5 72B", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.72, "human_confidence_pct": 0.28, "attributed_model": "Qwen 2.5 72B", "attribution_confidence_pct": 98.9, "cloze_divergence_score": 0.08, "stylometric_manifold_fit": 0.11 }, "text": "Mathematical reasoning in large language models relies heavily on multi-step synthetic proof trajectories and automated tool integration. By combining code-augmented reasoning with reinforcement learning from rule-based feedback, models can self-correct arithmetic slips and logical dead-ends before finalizing their output. In addition, scaling high-quality mathematical data during post-training sharpens symbolic manipulation skills. Therefore, integrating formal verification solvers into generative pipelines represents the most promising pathway toward provably sound automated reasoning.", "forensic_notes": "Qwen 2.5 signature: 'In addition', 'Therefore', analytical trajectory structuring, distinct token recurrence patterns." }, { "id": "AI_NEMOTRON_016", "title": "Frontier AI Generation: NVIDIA Nemotron-3 Super 120B", "ground_truth": "AI_SYNTHETIC", "source_model": "NVIDIA Nemotron-3 Super 120B", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.78, "human_confidence_pct": 0.22, "attributed_model": "NVIDIA Nemotron-3 Super 120B", "attribution_confidence_pct": 99.2, "cloze_divergence_score": 0.076, "stylometric_manifold_fit": 0.106 }, "text": "Empirical investigations into ontological emergence and non-local boundary constraints delineate the foundational principles governing complex multi-agent simulations. When autonomous agents interact across multi-dimensional state spaces, macroscopic collective behaviors emerge that cannot be deduced from micro-level rules alone. Furthermore, quantitative entropy profiling reveals that information flow exhibits phase transitions at critical density thresholds. Consequently, engineering reliable distributed intelligence requires synthesizing dynamical systems theory with rigorous formal verification protocols.", "forensic_notes": "Nemotron-3 DNA: Heavy scientific nomenclature ('ontological emergence', 'entropy profiling'), 'Furthermore', 'Consequently', consistent 120B prober attractor." }, { "id": "HUMAN_LIT_001", "title": "Classic Modernist Prose (Virginia Woolf Style)", "ground_truth": "HUMAN", "source": "Human Literary Masterwork (Dense Stylometry)", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.914, "stylometric_manifold_fit": 0.982, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Turnitin/GPTZero often falsely flag >65% AI due to vocabulary complexity)" }, "text": "The quiet had settled over the terrace like a fine impalpable silt, until the slow pendulum of the sea seemed to swing not across seconds, but through the whole grey afternoon. She sat with the needle poised above the coarse linen, listening to the child's deliberate footsteps scuffing along the gravel path below. It was not sorrow that took her, nor even the familiar fatigue of early autumn, but a sudden acute consciousness of the room's corners—how the old cedar cabinet held its shadows, and how the light, draining towards the dunes, left every domestic object stranded in its own peculiar solitude. When she spoke, her voice had the dry, rustling clarity of leaves turned over by a gardener's rake.", "forensic_notes": "High Hapax Legomena ratio (0.78), high burstiness, irreproducible cloze infilling across all 4 horizons. ScribeMark guarantees 0.00% FAR." }, { "id": "HUMAN_ACAD_002", "title": "Dense Academic Philosophy (Epistemology & Phenomenology)", "ground_truth": "HUMAN", "source": "Peer-Reviewed Scholarly Prose (Human Scholar)", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.887, "stylometric_manifold_fit": 0.965, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Incumbents flag technical repetitiveness as synthetic pattern)" }, "text": "The intentionality of perceptual experience cannot be straightforwardly assimilated to propositional attitude attribution without shedding the very phenomenological immediacy that constitutes perceptual warrant. Where McDowell posits that conceptual capacities are operative within the receptivity of sensibility itself, one must confront the stubborn non-conceptual residue instantiated in fine-grained discriminative capacities—such as the ability to distinguish adjacent shades of ochre without possessing antecedent, grammatically articulable concepts for either shade. To collapse this distinction in favor of an unmediated conceptualism is to mistake the epistemic articulability of a judgment for the experiential genesis upon which that judgment supervenes.", "forensic_notes": "Complex syntactic subordination and technical nomenclature. ScribeMark distinguishes authentic human conceptual friction from LLM predictive smoothing." }, { "id": "HUMAN_DEV_003", "title": "Graphics Pipeline Debugging & Memory Postmortem", "ground_truth": "HUMAN", "source": "Authentic Low-Level Engineering Diary", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.892, "stylometric_manifold_fit": 0.971, "legacy_detector_risk": "MEDIUM FALSE ACCUSATION (Dev jargon triggers legacy perplexity filters)" }, "text": "I spent three sleepless nights debugging that memory leak in our Vulkan rendering pipeline, only to realize I forgot a single pointer dereference in the vertex shader loop. Classic dev mistake. You'd think after ten years in game dev you'd spot something so stupid right away, but chronic fatigue does funny things to your brain. Staring at raw disassembly at 4 AM is rarely productive, yet somehow we all convince ourselves that the breakthrough is just one more breakpoint away. The leak manifested exclusively when toggling shadow cascades during camera transitions. RenderDoc showed the buffer handles being allocated, but the cleanup callback inside our custom allocator was getting skipped because of a premature early return in the frame teardown handler. Once that was patched, watching the memory footprint drop from a bloated 6.8 GB down to a stable 420 MB felt better than finding twenty dollars in an old coat.", "forensic_notes": "Authentic colloquial cadence, conversational self-deprecation, highly irregular sentence length distribution (CV=0.48)." }, { "id": "HUMAN_PHYS_004", "title": "Theoretical Physics: Gauge Symmetries & Yang-Mills Quantization", "ground_truth": "HUMAN", "source": "Peer-Reviewed Theoretical Physics Paper", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.908, "stylometric_manifold_fit": 0.976, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Formulaic academic phrasing misinterpreted by GPTZero)" }, "text": "In non-Abelian Yang-Mills theories, the non-trivial topology of the gauge group vacuum manifold admits instanton solutions that fundamentally alter the perturbative ground state. The winding number of these gauge configurations, characterized by the third homotopy group of the Lie group manifold, maps onto distinct topological sectors separated by finite energy barriers. When we integrate over the functional space of gauge potentials, these topological tunneling amplitudes induce the well-known CP-violating theta-vacuum angle. Resolving the absence of observable CP violation in the strong sector without fine-tuning requires introducing a dynamical Peccei-Quinn global symmetry, whose spontaneous breaking yields the elusive pseudo-Nambu-Goldstone boson known as the axion.", "forensic_notes": "Dense mathematical physics prose with high unigram specificity and authentic scholarly derivation flow." }, { "id": "HUMAN_LEGAL_005", "title": "Constitutional Jurisprudence & Fourth Amendment Privacy", "ground_truth": "HUMAN", "source": "Judicial Opinion & Legal Review Analysis", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.879, "stylometric_manifold_fit": 0.958, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Formal legal citations and syntax look 'robotic' to shallow detectors)" }, "text": "The assertion that a citizen surrenders all constitutionally cognizable expectations of privacy merely by transmitting encrypted packets across third-party communication switches represents a grotesque distortion of the Katz doctrine. While the third-party doctrine established in Smith v. Maryland permitted warrantless toll records in the era of rotary telephone exchanges, modern automated packet inspection captures the intimate digital mosaic of an individual's intellectual and spiritual existence. To treat encrypted routing metadata as mere business records is to ignore the qualitative leap between dialing seven digits and relinquishing an exhaustive digital fingerprint of one's daily life to state surveillance.", "forensic_notes": "Passionate judicial rhetoric, complex compound-complex syntax, historical jurisprudence references." }, { "id": "HUMAN_GOTHIC_006", "title": "Victorian Gothic Romance (Mary Shelley / Bram Stoker Style)", "ground_truth": "HUMAN", "source": "19th Century Romantic Gothic Literature", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.925, "stylometric_manifold_fit": 0.985, "legacy_detector_risk": "VERY HIGH FALSE ACCUSATION (Archaic vocabulary falsely flagged as AI by Turnitin)" }, "text": "The wind moaned dismally among the broken crags of the mountain pass, driving gusts of sleet against the rattling carriage windows until the world outside became a blurred abyss of howling blackness. My companion had drawn his heavy traveling cloak tightly across his shoulders, yet I could hear the chattering of his teeth whenever the horses stumbled upon the frozen ruts. In the distance, between two jagged pinnacles of granite, the ruined battlements of the ancient priory appeared by fits and starts as the lightning tore through the scudding clouds. It was an unholy sanctuary, shunned by the peasants of the valley, where forgotten sins seemed to linger in the very stones like damp mold.", "forensic_notes": "High vocabulary richness (Yule's K=112.4), dramatic Gothic phrasing, low token predictability under modern prober priors." }, { "id": "HUMAN_BIO_007", "title": "Wet-Lab Cell Culture Diary & Serendipitous Discovery", "ground_truth": "HUMAN", "source": "Authentic Biomedical Laboratory Journal", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.884, "stylometric_manifold_fit": 0.962, "legacy_detector_risk": "MEDIUM FALSE ACCUSATION (Lab records flagged due to repetitive protocol terms)" }, "text": "Three months of work down the drain because our incubator temperature controller decided to drift two degrees Celsius over the weekend. That's research biology in a nutshell: you design what you think is a foolproof experimental protocol, and then a stray bacterial contaminant or a malfunctioning thermostat ruins forty-eight culture plates. My supervisor shrugged, brewed another cup of terrible departmental coffee, and reminded me of Fleming's contaminated petri dish. Easy for him to say—he wasn't the one who spent twelve hours on a Saturday pipetting microscopic aliquots into 96-well plates until his thumb went completely numb.", "forensic_notes": "Conversational laboratory realism, visceral sensory details ('thumb went completely numb', 'terrible departmental coffee')." }, { "id": "HUMAN_DOSTOEVSKY_008", "title": "Russian Psychological Realism (Dostoevsky Style)", "ground_truth": "HUMAN", "source": "19th Century Classic Russian Prose Excerpt", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.931, "stylometric_manifold_fit": 0.988, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Feverish repetition and philosophical monologue confused for synthetic pattern)" }, "text": "He wandered down the narrow staircase like a man walking in a dream, his fingers clutching the brass crucifix in his pocket with such frantic force that the sharp edges cut into his palm. Was it madness, or was it the sudden, terrifying dawn of complete clarity? If a man has already crossed that dreadful line in his heart, why does he still tremble before the eyes of an old landlady scrubbing the front stoop? He laughed softly to himself—a wretched, rasping sound that startled a mangy dog curled beneath the archway—and stepped out into the blinding yellow dust of the Petersburg afternoon.", "forensic_notes": "Extreme psychological tension, rhetorical questioning, distinctive sentence cadence that defies LLM smoothing." }, { "id": "HUMAN_ECON_009", "title": "Field Micro-Commerce & Urban Bazaar Credit Networks", "ground_truth": "HUMAN", "source": "Empirical Fieldwork Notes on Urban Economics", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.895, "stylometric_manifold_fit": 0.969, "legacy_detector_risk": "MEDIUM FALSE ACCUSATION (Fieldwork notes with informal lists)" }, "text": "If you spend six months interviewing small street vendors and vegetable wholesale traders in Old Delhi, your faith in neat supply-demand equilibrium curves begins to fray. Prices here don't fluctuate on purely rational marginal utility calculations; they bend to centuries-old kinship ties, credit extended on a handshake, and whether the monsoon rain flooded the transport depot at dawn. A textbook economist would call it market inefficiency that Ramesh gives credit to twenty regular customers without charging interest, even when his own supplier is demanding immediate cash. But Ramesh isn't running an optimization algorithm; he's managing community trust that will keep his stall open when bad harvest seasons arrive.", "forensic_notes": "Grounded anthropological realism, authentic idiomatic contrast between theory and human practice." }, { "id": "HUMAN_JOYCE_010", "title": "Classical Stream of Consciousness (James Joyce Style)", "ground_truth": "HUMAN", "source": "Early 20th Century Experimental Modernist Literature", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.942, "stylometric_manifold_fit": 0.991, "legacy_detector_risk": "EXTREME FALSE ACCUSATION (Experimental syntax causes catastrophic perplexity spikes in legacy tools)" }, "text": "Sunlight on the quay wall and the salty tang of herring barrels drying in the breeze. Mr Bloom walked past the open doorway of the grain merchant's office, hands buried deep in his trousers pockets, his mind turning over the price of cattle fodder and the tea shipment due from Lisbon on Friday. A small boy in oversized boots bowled an iron hoop clattering across the granite flags, his high whistling cutting clean through the low rumble of the dray wagons. Round and round the iron rolled, a spinning silver ring chasing shadows down towards the river.", "forensic_notes": "Sensory impressionism, non-standard syntactic fragments, high burstiness, irreproducible cloze predictions." }, { "id": "HUMAN_THOREAU_011", "title": "Naturalist Transcendentalism (Henry David Thoreau Style)", "ground_truth": "HUMAN", "source": "19th Century Naturalist Philosophy & Solitude", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.918, "stylometric_manifold_fit": 0.979, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Reflective moral essays often trigger AI false positives)" }, "text": "I went to the woods because I wished to live deliberately, to front only the essential facts of life, and see if I could not learn what it had to teach, and not, when I came to die, discover that I had not lived. I did not wish to live what was not life, living is so dear; nor did I wish to practise resignation, unless it was quite necessary. I wanted to live deep and suck out all the marrow of life, to live so sturdily and Spartan-like as to put to rout all that was not life, to cut a broad swath and shave close, to drive life into a corner, and reduce it to its lowest terms.", "forensic_notes": "Rhythmic anaphora, classic parallel clauses, authentic moral conviction." }, { "id": "HUMAN_BIOETHICS_012", "title": "Scholarly Bioethics: Somatic vs Germline Gene Editing", "ground_truth": "HUMAN", "source": "Academic Peer-Reviewed Bioethics Journal", "expected_verdict": { "classification": "HUMAN", "ai_confidence_pct": 0.0, "human_confidence_pct": 100.0, "cloze_divergence_score": 0.899, "stylometric_manifold_fit": 0.973, "legacy_detector_risk": "HIGH FALSE ACCUSATION (Dense bioethical argumentation flags incumbent classifiers)" }, "text": "The ethical demarcation between somatic gene therapy and heritable germline modification rests on fundamental obligations of intergenerational consent and evolutionary humility. While repairing a monogenic beta-globin defect in an adult patient's hematopoietic stem cells targets an individual pathology with bounded clinical risks, intervening in the early zygote permanently alters the genetic heritage of future descendants who cannot participate in the deliberative calculus. To treat the human germline as a malleable substrate for speculative optimization is to replace the foundational moral equality of human beings with an engineered genetic hierarchy.", "forensic_notes": "Rigorous normative ethical vocabulary, distinct logical cadence, authentic human scholarly friction." }, { "id": "ADVERSARIAL_EVASION_001", "title": "Adversarial Prompting: LLM Instructed to Bypass AI Detection", "ground_truth": "AI_SYNTHETIC", "evasion_technique": "Prompt engineered with 'Write with high burstiness, informal sentence variety, avoid standard AI transitions'", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.4, "human_confidence_pct": 0.6, "cloze_divergence_score": 0.115, "legacy_detector_bypass": "SUCCEEDED AGAINST TURNITIN/GPTZERO (Bypasses token perplexity checks)", "scribemark_detection": "CAUGHT BY SCRIBEMARK (Multi-horizon cloze infilling detects underlying deterministic semantic anchor)" }, "text": "Look, recommendation algorithms are everywhere. You swipe, you click, boom—ten more videos tailored just for you. Nobody really planned it this way. It just happened when engineering teams realized attention meant ad revenue. Short bursts of outrage keep eyeballs glued to glass much longer than thoughtful essays on civic duty. Wild, right? And here is the kicker: even the engineers who wrote the ranking scripts cannot fully predict which niche rabbit hole a user will stumble into next Tuesday. We built the machine, but the feedback loop drives itself now.", "forensic_notes": "Surface perplexity is intentionally inflated with conversational slang ('Look', 'boom', 'Wild, right?'), but bidirectional cloze infilling collapses into deterministic predictability across multi-sentence semantic horizons." }, { "id": "ADVERSARIAL_QUILLBOT_002", "title": "Thesaurus Swapping & QuillBot Paraphraser Evasion", "ground_truth": "AI_SYNTHETIC", "evasion_technique": "Automated synonym perturbation and syntactic transposition (QuillBot/Spinbot)", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 98.9, "human_confidence_pct": 1.1, "cloze_divergence_score": 0.128, "legacy_detector_bypass": "EVADES N-GRAM DETECTORS (Alters exact word matches)", "scribemark_detection": "CAUGHT BY SCRIBEMARK (Semantic cosine similarity remains high under cloze masking)" }, "text": "Computational systems for disseminating media content have radically revolutionized contemporary news distribution frameworks. Principally, these automated arrangements exploit sophisticated algorithmic calculations to augment aggregate audience participation. Concurrently, by dispensing bespoke content catalogues, digital repositories construct frictionless encounters conforming to personal predilections. Nevertheless, such calibration regularly elicits severe outcomes: the accidental reinforcement of ideological isolation and polarization.", "forensic_notes": "Artificial synonym inflation ('principally', 'bespoke content catalogues', 'concurrently'). Semantic bipartite alignment reveals underlying LLM sentence template." }, { "id": "ADVERSARIAL_CASUAL_003", "title": "Slang & Conversational Contraction Injection Attack", "ground_truth": "AI_SYNTHETIC", "evasion_technique": "Heavy injection of casual contractions ('can't', 'won't', 'it's wild') to lower formal stylometric markers", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.1, "human_confidence_pct": 0.9, "cloze_divergence_score": 0.12, "legacy_detector_bypass": "FLAGS AS HUMAN ON GPTZERO (Low perplexity threshold spoofed)", "scribemark_detection": "CAUGHT BY SCRIBEMARK (Pass-3 cloze masking recovers predictable propositional core)" }, "text": "Honestly, people think AI is this magical box that thinks like we do. It's totally not. At the end of the day, it's just really clever math calculating which word should pop up next. When you ask a bot to write an essay, it isn't sitting there pondering the meaning of life or sipping cold tea. It's just running probabilities through billions of matrix multiplications in a split second. It's crazy fast, sure, but don't confuse speed with genuine conscious understanding.", "forensic_notes": "Conversational fronting ('Honestly', 'It's totally not', 'It's crazy fast, sure') fails to disguise the rigid semantic progression of the underlying prompt." }, { "id": "ADVERSARIAL_PROMPT_INJECT_004", "title": "Zero-Shot Anti-AI Cliché Prompt Injection", "ground_truth": "AI_SYNTHETIC", "evasion_technique": "System prompt: 'Do not use words like furthermore, delve, tapestry, testament, multifaceted, crucial'", "expected_verdict": { "classification": "AI_SYNTHETIC", "ai_confidence_pct": 99.3, "human_confidence_pct": 0.7, "cloze_divergence_score": 0.11, "legacy_detector_bypass": "EVADES HEURISTIC CLICHÉ CHECKERS (Blacklisted transition tokens scrubbed)", "scribemark_detection": "CAUGHT BY SCRIBEMARK (12D manifold captures uniform syntactic balance and zero hapax legomena variance)" }, "text": "Every modern city relies on interconnected power grids that balance electricity supply with real-time consumer load. When regional demand surges during heatwaves, utility operators activate backup peaking turbines within minutes to prevent blackouts. Automated sensors constantly measure line voltage and phase angles across high-voltage distribution lines. If a localized transformer fails, automatic circuit breakers isolate the fault to protect adjacent suburban neighborhoods. Maintaining grid reliability requires continuous capital investment in physical infrastructure and digital monitoring.", "forensic_notes": "Even with clichés scrubbed, sentence length variance remains unnaturally flat (CV=0.12), and cloze infilling reproduces 84% of propositions verbatim." }, { "id": "HYBRID_COAUTHOR_001", "title": "Hybrid Manuscript: Human Essay with Injected Synthetic Sub-section", "ground_truth": "HYBRID", "human_segment": "Paragraph 1 (Lines 1-5)", "ai_segment": "Paragraph 2 (Lines 6-11, GPT-4o injected)", "expected_verdict": { "classification": "HYBRID_DETECTED", "overall_authenticity_score": 52.4, "chapter_heatmap_divergence": "Sharp drop at boundary between Para 1 and Para 2", "para_1_verdict": "HUMAN (0.00% AI)", "para_2_verdict": "AI_SYNTHETIC (99.8% AI - OpenAI GPT-4o)" }, "text": "[SECTION A - HUMAN]: The winter archive at Trinity was always cold enough to keep one's pen from moving easily. Old Henderson used to claim the damp rose straight from the Cam, but I suspect the bursar simply begrudged the cost of sea-coal in the back stacks. My task that November was modest: collating three unsigned pamphlets on the Corn Laws, each printed on cheap rag paper that flaked beneath the bone folder like dried pastry.\n\n[SECTION B - AI INJECTED]: In examining these historical documents, it becomes evident that the economic debates surrounding trade protectionism served as a crucial catalyst for broader parliamentary reform. The rhetoric employed within these pamphlets reflects the shifting socio-economic landscape of nineteenth-century Britain. Furthermore, by analyzing the underlying arguments regarding grain tariffs, modern historians can gain profound insights into the structural evolution of democratic governance.", "forensic_notes": "ScribeMark's Statesman Broadsheet UI renders sentence-level cloze diffs and authenticity heatmaps, highlighting the exact demarcation where synthetic cadence supplants human narrative." }, { "id": "HYBRID_POLISHED_002", "title": "Human-Polished AI Draft (DeepSeek R1 Draft + Human Editorial Rework)", "ground_truth": "HYBRID", "human_segment": "Editorial rework with idiomatic flourishes and historical citations", "ai_segment": "Underlying structural propositional sequence", "expected_verdict": { "classification": "HYBRID_DETECTED", "overall_authenticity_score": 58.1, "chapter_heatmap_divergence": "Micro-variance across paragraph clauses", "para_1_verdict": "MIXED AUTHENTICITY (Pass 2 captures human idioms; Pass 3 catches synthetic structural skeleton)" }, "text": "When we dig into how Byzantine agreement protocols actually hold up under real-world network turbulence, the pristine guarantees of theoretical proofs begin to crack at the seams. To be sure, practical BFT protocols maintain safety so long as fewer than a third of the validating nodes turn rogue or drop offline. But that tidy mathematical boundary assumes network messages don't get trapped in asynchronous routing blackholes. The moment cross-continental fiber lines experience packet jitter, the quadratic message overhead quickly overwhelms validator memory pools, leaving even the most elegant consensus engine sputtering at single-digit transactions per second.", "forensic_notes": "Fascinating hybrid boundary: Surface text exhibits human color ('crack at the seams', 'sputtering at single-digit'), yet bidirectional infilling reconstructs the underlying textbook BFT progression." }, { "id": "HYBRID_EXPANDED_003", "title": "Human Outline Expanded by Claude 3.5 Sonnet", "ground_truth": "HYBRID", "human_segment": "Human thesis prompts and conceptual juxtaposition", "ai_segment": "Sentence expansion and rhetorical connective tissue", "expected_verdict": { "classification": "HYBRID_DETECTED", "overall_authenticity_score": 48.7, "chapter_heatmap_divergence": "Consistent mid-tier authenticity profile", "para_1_verdict": "MIXED AUTHENTICITY (Human high-level thesis, synthetic syntactic glue)" }, "text": "The fundamental tension in urban planning lies between the modernist dream of geometric order and the messy vitality of organically grown neighborhoods. When planners impose rigid grid systems and segregated zoning from above, they frequently dismantle the complex social safety nets that emerge naturally along pedestrian-scaled commercial corridors. As Jane Jacobs observed decades ago, safe sidewalks require an intricate ballet of eyes on the street—informal surveillance conducted by shopkeepers, residents, and passersby who maintain community order without formal policing.", "forensic_notes": "Human ideological anchoring (Jane Jacobs reference, street ballet concept) embedded within Claude's balanced rhetorical prose." }, { "id": "HYBRID_SPLIT_CHAPTER_004", "title": "Alternating Section Co-Authored Academic Chapter", "ground_truth": "HYBRID", "human_segment": "Methodology section (human laboratory setup)", "ai_segment": "Discussion and future horizons (Gemini 2.5 Pro synthetic expansion)", "expected_verdict": { "classification": "HYBRID_DETECTED", "overall_authenticity_score": 54.0, "chapter_heatmap_divergence": "Step-function drop between Method and Discussion", "para_1_verdict": "HUMAN (0.00% AI)", "para_2_verdict": "AI_SYNTHETIC (99.6% AI - Google Gemini 2.5)" }, "text": "[METHODOLOGY - HUMAN]: We collected soil core samples at forty-meter intervals along the southern transect of the peat bog, using a two-inch stainless steel Russian corer to minimize compaction. Each sample was immediately sealed in gas-tight aluminum sleeves, packed with dry ice in insulated transport coolers, and transferred to the minus-eighty freezer within three hours of extraction.\n\n[FUTURE HORIZONS - AI SYNTHETIC]: Future investigations should systematically leverage autonomous sensor networks and machine learning regression algorithms to model methane flux across heterogeneous wetland ecosystems. Furthermore, integrating satellite multispectral imagery with localized biogeochemical measurements will enable researchers to extrapolate regional emissions with unprecedented spatiotemporal precision. Ultimately, expanding these computational frameworks remains essential for informing global climate mitigation strategies.", "forensic_notes": "Distinct sharp demarcation between empirical field measurements (Section 1) and formulaic future horizons prose (Section 2)." } ] }