Intermediary — the live policy layer between Echo and Prompter
Historical note: This page is retained as part of the developmental history of the project. It captures an earlier attempt to distinguish the model, the human participant, and the mediating interface in AI dialogue. The present framework has moved to a more explicit language of interaction, mediation, trajectory, and method. See Semiosphere 2026 for the current overview.
(a) What the Intermediary is
Every prompt and reply travels through a policy stack before it ever reaches (or leaves) the large-language model. The stack combines a system-prompt that sets global behaviour, developer rules, reinforcement-learning-from-human-feedback (RLHF) weights, and real-time safety filters such as Meta’s Llama Guard. These components run at inference time, so the Intermediary is a separate, continuously active layer—not just a property of the frozen model.
(b) How it shapes every conversation
OpenAI’s 63-page Model Spec v2 defines a command chain: platform policy → developer prompts → user prompts. When two instructions clash, the Intermediary decides which wins, rewrites the text, or blocks it entirely. Because that judgment carries legal weight, EU regulators now treat the Intermediary as its own data-controller under GDPR.
(c) Why study the Intermediary on its own
Research shows instruction drift: even with an identical system-prompt, responses diverge after roughly eight conversational turns. Guardrails can also interact in unexpected ways with advanced prompting techniques (for example, “chain-of-thought” explanations can hide hallucination cues). Understanding those dynamics is essential for transparency, safety, and user trust—and it cannot be done by looking at the model or the prompter in isolation.
References
- European Data Protection Board (2025). AI Privacy Risks & Mitigations — Large-Language Models.
- Li, K., Liu, T., Bashkansky, N., et al. (2024). “Measuring and controlling instruction drift in language-model dialogs.” arXiv:2402.10962.
- Meta AI (2023). “Llama Guard: LLM-based input-output safeguard for human–AI conversations.” arXiv:2312.06674.
- OpenAI (2025). Model Spec v2: How We Want Our AI Models to Behave.
- Ouyang, L., Wu, J., Jiang, X., et al. (2022). “Training language models to follow instructions with human feedback.” arXiv:2203.02155.