Empirical Research on Structural Limits of Generative AI Models: Context Pack, Token Pollution, and Multimodal Stabilization Protocols (Sora 1 / ChatGPT)

This report documents an independent R&D investigation conducted entirely on a mobile phone over two months (October–December 2025), without institutional affiliation, AI training access, or technical infrastructure.

The work focuses on the structural behavior of large language and multimodal generative models, including ChatGPT, Sora 1, Claude, Gemini, and Grok, through intensive, reproducible, multi-run testing across six languages (English, French, Japanese, Russian, Czech, Chinese).

Key contributions include: the identification and quantification of a "token pollution" phenomenon linked to contextual reconstruction; the development of the Context Pack, an external cognitive preset designed to reduce implicit context reconstruction, token waste, and drift; the design of a structured prompt architecture (Glasswraith Protocol) enabling multimodal stabilization in video generation, including simultaneous stabilization of up to 61 distinct objects in Sora 1; the identification of implicit physical and linguistic biases in generative models; and a conceptual proposal for a PDF Engine 2.0 oriented toward structured knowledge transmission.

All observations are visually documented and reproducible. The report also includes an applied analysis of the ecological and computational cost of token inefficiency at scale.

Primary language: French.

Executive summary available in English.