The Architectural Immune System: A Framework for Correcting Synthetic Fallacies in AI-Driven Science

David Scott Lewis, Anar Batkhuu, Haley Yi

Keywords: autonomous scientific discovery, agentic science, AI for science, agents for science, self-directed research, architectural immune system, synthetic data detection, hybrid database–model validation, tri‑functional cosmetics materials, self‑falsification, biosurfactants, sophorolipids, phenylpropanoid grafting, UV protection, emulsification, antimicrobial, SPF, CMC, MIC, ChEMBL, PubChem, CosIng

TL;DR: An 'architectural immune system' flagged synthetic data in a cosmetics pipeline; hybrid validation then yielded tri-functional biosurfactants (SPF≈14; CMC≈42 mg/L; MIC≈285 ppm), showing self-falsification is central to trustworthy materials AI.

Abstract: We introduce the ’Architectural Immune System,’ a framework for trustworthy autonomous science that enables AI agents to detect and correct their own ’synthetic fallacies.’ We demonstrate its efficacy in a materials discovery case study, where an agent’s immune system rejected a statistically impossible ’perfect’ result caused by a silent algorithmic failure. By forcing a pivot to database-grounded evidence, the system produced a more modest but physically authentic solution, establishing a new design pattern for robust, self-correcting scientific agents. The framework integrates a comprehensive 10-tool research validation ecosystem—including literature corpus analysis, adversarial critique protocols, multi-stage physical feasibility validation, and molecular dynamics simulations—with authentic ChEMBL and PubChem database validation. Through systematic analysis of 2,847 individual decisions and evaluation of 623 cross-domain validation patterns, the agent’s approach yielded genuine tri-functional phenylpropanoid-grafted sophorolipids (PGSLs) with optimal ratios of 36.5:38.5:25.0, delivering measurable performance: SPF 14.3 ± 2 (derived from 500+ related compounds), CMC 42.5 ± 5 mg/L (based on 847 sophorolipid SAR records), and MIC 285 ± 30 ppm (from 15,000 antimicrobial assays). The agent performed computational optimization using 152,001 real compound records and 250,000 experimental bioactivity measurements from validated chemical databases. This work demonstrates that architectural safeguards against synthetic fallacies are essential for trustworthy AI systems in materials discovery, providing a template for robust autonomous research frameworks.

Submission Number: 340