Title: Reproducibility as the New Copyleft: Establishing Standards for AGI-Oriented Reproducible Builds

Abstract:

The legal strategy of copyleft, exemplified by licenses like the GNU General Public License, functioned as a clever inversion of copyright law to secure user freedoms by mandating source code availability with every distribution. However, this approach relied on an unspoken technical assumption: that there exists a clear, auditable, and reproducible link between source code and its compiled object code. This foundational premise is fundamentally undermined by large language models and, potentially, Artificial General Intelligence (AGI) systems.

Reconstructing an AI model requires a complex ecosystem of artifacts—including code, datasets, model weights, hyperparameters, software toolchains, and hardware specifications. Each of these components faces distinct legal, technical, and economic barriers that existing open-source frameworks fail to adequately address. Furthermore, highly capable AI systems possess the ability to rewrite licensed source code into functionally identical derivatives, effectively stripping away original licensing obligations. This form of "laundering" exposes a critical weakness in traditional copyleft enforcement.

This paper contends that a functional equivalent of copyleft for AGI cannot rely on share-alike clauses applied to code. Instead, it must be anchored in reproducible builds, a methodology that ensures bit-exact reconstruction from specified inputs. We analyze the underlying logic of copyleft and critically evaluate Maffulli’s "Second Liberation" thesis, which posits that AI realizes Stallman’s original vision. We demonstrate that this argument holds true only if AGI systems themselves are reproducible.

By synthesizing insights from the Open Source AI Definition (OSAID), the Model Openness Framework (MOF), OpenMDW, and research into deterministic inference, we propose seven specific requirements for AGI-oriented reproducible builds. Additionally, we argue that the Model Context Protocol (MCP) and similar AI-to-AI coupling mechanisms represent a new dynamic linking layer. For this layer, traditional copyleft licensing is inadequate. Instead, we suggest that Masnick’s framework of "protocols, not platforms" provides a more effective governance model for the future of AI.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC