FORGEry: A Multi-Model Adversarial Research Methodology for Independent AI Researchers

We present FORGEry, a multi-model adversarial research methodology enabling independent researchers to produce credible, reproducible AI research results without institutional infrastructure. FORGEry structures collaboration between large language models in strictly separated roles — execution substrate and architectural interrogator — with the human researcher as the sole cross-model routing node. Role separation prevents inter-model collusion and preserves adversarial tension throughout the research loop. Milestone surprisal checks via third-model injection guard against primary-model convergence on shared blind spots.

We demonstrate validity through a concrete result: structural weight mapping from Mamba2-2.7B to Mamba3 (cross-entropy ratio 1.0016 on uniform random token sequences), prior to the Mamba team's reference model release. The nine-point weight mapping is open-sourced at https://github.com/Rta-Forge/heists-galore. The reference checkpoint is at https://huggingface.co/RtaForge/Mamba3-2.7B. Source baseline: https://huggingface.co/state-spaces/mamba2-2.7b.

The structural mapping was produced using the FORGEry adversarial methodology. All architectural decisions and validation were performed by the author; execution substrate and interrogator roles were filled by frontier LLMs under strict role separation. No institutional funding.