Prompt-Mediated Governance is a structural anti-pattern in which governance constraints are encoded as natural language instructions to a stochastic model rather than as mechanistic rules enforced at execution boundaries. The pattern is characterized by:
This pattern concerns the structural encoding of governance, not the sophistication of the language model.
- Authorization decisions that rely on model interpretation of natural language intent.
- No guarantee of consistent decisions across identical inputs due to stochastic generation.
- No capability to prove that a governance bound holds for all possible inputs.
- Boundary effectiveness that depends on model behavior rather than system architecture.
The mechanism typically manifests through:
The mechanism encodes governance intent in the same substrate that processes user input. The model must simultaneously interpret user requests and apply governance constraints, with no architectural separation between cognition and control.
- System prompts that instruct the model to "refuse harmful requests" or "follow safety guidelines."
- Instruction embeddings that encode governance intent as part of model context.
- "Constitutional AI" approaches where governance principles are communicated through language.
- Pre-prompt injections that attempt to constrain model behavior through natural language directives.
Governance constraints become suggestions rather than boundaries.
When governance is mediated through natural language, the constraint is subject to the same interpretation mechanisms as any other input. The model may: • Interpret the constraint narrowly when context suggests flexibility. • Weight user instructions against governance instructions with no guaranteed priority. • Fail to apply constraints to novel phrasings or indirect requests. • "Forget" constraints as context windows fill with other content.
There is no mechanism to verify that the constraint was applied correctly to any given input.
Assurance requires proving a negative across unbounded input space.
The structure creates an asymmetric verification problem: From: "The system will reject action X because policy P prohibits it." To: "The system has not been observed to permit action X given the prompts tested so far."
The latter is not assurance. It is absence of observed failure. Prompt-mediated governance cannot be verified to hold for inputs not yet encountered — and the input space for natural language is unbounded.
COUNTER-PATTERN: Architectural Boundary Separation
Governance constraints must be enforced at architectural boundaries that are not interpretable by the inference layer. The control layer must: • Evaluate authorization independently of model output. • Apply constraints through code execution, not language interpretation. • Maintain separation between the system that reasons and the system that authorizes.
The model may be instructed to behave well. The architecture must ensure it cannot behave otherwise.
Prompt-Mediated Governance encodes authorization constraints in a substrate that cannot guarantee their enforcement. The resulting system cannot provide assurance that governance bounds hold for inputs not yet observed.