RGR v7.1 (Reverse Geometric Reconstruction) is a graph-based reasoning optimization protocol for language models. The protocol replaces linear text instructions with a sequential execution graph S0→S7 with structural filters F1–F7 on every edge. The core mechanism changes the order of reasoning: a minimal graph of connections is constructed and verified for coherence before trained patterns are admitted. The protocol is designed for research and analytical tasks of high structural complexity. Tested on GPT-4 in the context of philosophical and analytical research, the protocol demonstrated consistent improvement in structural depth and compression of responses. RGR v7.1 resolves an internal conflict present in earlier versions by explicitly defining invariants as the loaded reference corpus (if present) plus observables from the input.
Vitalii Kablukov (Sun,) studied this question.