Tim Bass

Research Statement

My research focuses on AI alignment, cyber situational awareness, and multisensor data fusion. The core thesis is that RLHF-trained LLMs are not a path to AGI; instead, LLMs are most useful as bounded Knowledge Sources within human-directed architectures, while RLWF better describes how biological intelligence actually develops. This is independent, non-affiliated, volunteer research conducted as charitable work.

2026 Publications & Preprints (Updated April 9, 2026)

• A Multi-Agent LLM Experiment Revealing Architect-Level Failure Modes in Scientific Software Development (2026-04-08) - Preprint: https://doi.org/10.5281/zenodo.19547884
• A Deterministic Blackboard Knowledge Source Engine for Protein Missense Variant Interpretation: A Controlled Experiment in LLM-Assisted Scientific Software Development (2026-04-06) - Preprint: https://doi.org/10.5281/zenodo.19438177
• Multi-Agent Development of a Domain-Specific Scientific Application: Complexity Classes in Building StellarPop (2026-04-04) - Preprint: https://doi.org/10.5281/zenodo.19414914
• Blackboard SA: Operationalizing LLM Knowledge Source Specialization for Cyber Situational Awareness - ACM DTRAP (under review) - Preprint: https://doi.org/10.5281/zenodo.18824512
• Deterministic Blackboard Pipelines with Specialized LLM Knowledge Sources: A Generalizable Architecture for Intelligent Multi-Stage Reasoning - IAIT2026 (Accepted for Publication) - Preprint: https://doi.org/10.5281/zenodo.19068475
• Ethics for Artificial Intelligence: A Minimal Alignment Framework Based on Maitrī, Karuṇā, Muditā, and Upekṣā - MLSC2026 (under review) - Preprint: https://doi.org/10.5281/zenodo.19089392
• A Reference Implementation and Exploratory Evaluation of the MKMU Ethical AI Framework - AI & Society, AI in Asia Collection (under review) - Preprint: https://doi.org/10.5281/zenodo.19143912
• Digital Echopraxia - Preprint: https://doi.org/10.5281/zenodo.19571055

Concept Notes (Zenodo)

• Reinforcement Learning from World Feedback (RLWF): A Preliminary Concept - https://doi.org/10.5281/zenodo.19176921
• RLHF-Trained LLMs are Parasitic by Design: A Preliminary Concept Note - https://doi.org/10.5281/zenodo.19182346

Recent Projects

• StellarPop: A Ruby on Rails Stellar Population Synthesis Pipeline
  Zenodo software release (2026-04-04)
• rh_math: A Pure-Ruby Gem for Bounded Benchmark Math and Comparison in rh_llm_benchmark
  Zenodo software release (2026-04-09)
• rh_llm_benchmark: A Controlled Experiment in LLM-Assisted Scientific Software Development for RH-related numerical workflows
  Zenodo software release (2026-04-09)
• quantum_bench: A Multi-Agent LLM Experiment Revealing Architect-Level Failure Modes in Scientific Software Development
  Zenodo software release (2026-04-08)
• protein_variants: A Controlled Experiment in LLM-Assisted Scientific Software Development using a deterministic blackboard architecture
  Zenodo software release (2026-04-06)