Semantic Artifact Harness for AI-Native Research: Reshaping the Paper Form Through Claims, Algorithms, Code, and Evidence

Abstract

**Problem.** Most research papers follow a linear narrative format designed for human reading. This format is mismatched with how papers are now consumed: AI systems search, summarize, compare, and reuse scientific claims before any human reads them. Conventional prose hides the links among claims, assumptions, algorithms, code, formulas, figures, citations, and implementation decisions. The resulting risk is an AI-to-AI contamination loop: AI writes, reads, and reviews papers; databases index them; later AI systems train on accumulated output. In that loop, errors do not disappear---they become more stable because they are repeated in a scholarly style. **Intervention.** This paper proposes the *Semantic Artifact Harness*: a paper form that treats the manuscript as a typed program with explicit interfaces rather than as a final prose report. The design applies software engineering norms---contracts, test harnesses, API specifications, separation of concerns, supply-chain integrity, and CI/CD-style verification---as explicit constraints for AI-native research artifacts. The goal is to make every claim auditable, every evidence link traceable, every algorithm inspectable, and every limitation machine-readable. **Framework.** The harness is organized into three layers---narrative, semantic evidence, and executable/code---following the separation-of-concerns principle. Its central components are research cards, claim-evidence contracts, dense literature matrices, formula passports, algorithm passports, code-artifact maps, claim-linked figures, AI-interaction traces, and an anti-contamination review protocol. A verification ladder establishes explicit norms: each research claim must pass through layers of verification (formal proof, code execution, simulation, experiment, embodied interaction, literature feedback) before being promoted. **Demonstration.** The format is demonstrated through a loaded tooth contact analysis (LTCA) program for cycloid-pin-housing systems in robotic RV reducers. The manuscript itself is used as a self-bootstrapping AI-assisted artifact: its code inspection, manuscript structure, tables, figures, literature matrices, BibTeX records, and correction traces were iteratively generated, reorganized, and audited through human--AI interaction. The case demonstrates that SE practices---contract-driven design, interface specification, regression testing, and supply-chain integrity---transfer naturally to computational research artifacts.

Keywords