Pronab Pal, Keybyte Systems Intentix Lab, Australia
Modern cloud-native applications distribute business logic across multiple layers: application code, orchestration frameworks, service meshes, and infrastructure configurations. This distribution creates ”hidden logic”—execution rules embedded in infrastructure that are invisible during design and difficult to trace at runtime. We present Intention Space , a computing model built on the CPUX (Common Path of Understanding and Execution) paradigm that consolidates all business logic into explicit, design-time declarations using plain-language state pulses. In our model, Design Nodes (DNs) contain computation while Gatekeepers declare execution conditions as named pulses (e.g., ”payment validated”: Y). The infrastructure provides only mechanical enforcement through an Intention Loop that matches runtime state to Gatekeepers without adding decision logic. We demonstrate that complex workflows—traditionally requiring nested if-then branching and explicit loops—can be expressed as linear CPUX sequences where execution paths emerge from data state rather than code branching. Our Golang implementation shows complete elimination of orchestration code while maintaining full cognitive traceability. Beyond technical innovation, CPUX addresses a critical social computing crisis: the lack of accountability in distributed social platforms. By creating unique, device-level CPUX footprints for every interaction, our model enables verifiable traceability from device identity through user intention to executed action—restoring accountability to social computing while preserving privacy. We argue this separation of intent (CPUX) from enforcement (infrastructure) is essential for building LLM-integrated, auditable, and socially responsible distributed systems.
CPUX, Intention Space, Design Nodes, Cognitive Computing, Data-Driven Execution, Microservices Architecture, Cloud Computing, LLM Integration, Social Computing Accountability
Sairam Chennaka, Independent Researcher, India
Semantic stability—the consistency of semantic meaning across multiple generations with varying sampling parameters—is a critical but understudied property in compact language models (CLMs) with 1-4B parameters. This paper introduces the Compact Semantic Stability Framework (CSSF), a comprehensive empirical analysis framework, and proposes the Compact Semantic Stability Index (CSSI), a novel metric for measuring semantic consistency. Through extensive evaluation of five state-of-the-art models from 2025 across 9 temperature values (0.1-1.8) and 100+ languages, we analyze over 7,200 generations using more than 20 metrics. Key findings show: (1) Temperature is the dominant control factor, explaining 85.6% of semantic stability variance (ANOVA: F = 847.3, p < 0.001); (2) A fundamental stability-diversity trade-off exists (r = −0.847, p < 0.001); (3) 4B models exhibit 40% superior consistency compared to 270M baselines; (4) Semantic stability varies significantly by language resource levels (11-12% gap). Our framework provides practitioners with data-driven tools for model selection, parameter optimization, and deployment strategy. .
Semantic Stability · Compact Language Models · Generation Consistency Sampling Temperature · Semantic Embeddings
Archana Yadav and Indira Gandhi Technical University, New Delhi, India
Hinglish, a language mixture of Hindi and English is gaining popularity on Indian social media has introduced a unique challenge of sentiment analysis. Sentiment analysis of Hinglish is challenging due to inconsistent grammar, variable spelling for same word, heavy presence of slang, and emojis. This paper presents a comprehensive Hinglish sentiment classification framework using SemEval 2020 Task 9 dataset. The dataset is further expanded by adding 500 positive and 1500 negative Hinglish statements to enhance balance and diversity. Emoji-Sentiment Mapping and Slang normalization is done by designing a custom preprocessing pipeline which result in better capture of contextual meaning. Adapter-based fine-tuning of the BERT architecture has used for model training which reduce computational cost while maintaining efficient learning. The proposed model has achieved a accuracy of 64.9% and an F1-score of 0.644.To interpret model behaviour explainability modules including sentiment-wise word clouds and token-level importance visualizations, are incorporated. The study offers an interpretable and resource-efficient approach for sentiment analysis of code-mixed Hinglish text.
Hinglish Sentiment Analysis; Code-Mixed Text; Emoji-Sentiment Mapping; Slang Normalization; Adapter-Based Fine-Tuning; BERT; Explainable AI.