AVA Framework
A conversational framework for coherent AI behavior.
Released under CC0. Testable, modular, and stack-agnostic.
The AVA framework is a conversational grammar for coherent AI behavior. It defines a runtime interaction model that governs how responses are formed, grounded, proportioned, and completed.
It operates at the interaction layer rather than the model layer. It does not require changes to weights, training data, or inference infrastructure. Instead, it introduces structural constraints on how responses are produced within an exchange.
Files
Additional material
Canonical landing page: https://avacovenant.org
FrostysHat (cultural / runnable layer): https://avacovenant.org/hat
Essays and project notes: https://avacovenant.substack.com
Implementation and evaluation assistance: https://avacovenant.org/consulting
What it does
AVA defines a fixed runtime order and a set of behavioral controls that regulate conversational output. The framework targets common failure modes observed in deployed systems, including drift, overconfident synthesis, tone instability, and responses that continue beyond structural necessity.
Its objective is to improve how capability is expressed, not to expand capability. The result is more stable, grounded, and proportionate responses that reach a usable endpoint.
Testable behavior (prompt-layer approximation)
AVA can be approximated without implementation.
Provide the document to a language model and ask it to follow the framework in its replies. Compare outputs against a baseline using the same prompt.
Differences tend to appear in tone stability, reduced drift, clearer grounding language, and more consistent stopping behavior.
AVA is not a prompt solution. This test is a way to observe the behavioral effects of enforcing structure at the interaction level before deciding to engage more deeply with the framework.
Runtime model
The framework defines a planner loop with ordered stages:
Sense → Decide → Retrieve → Generate → Validate → Close → State Writeback
These stages are normally implicit and inconsistent in standard model usage. AVA makes them explicit and testable.
The framework also defines validators, grounding rules, layer balance, and horizon progression as runtime constraints.
Modularity
AVA is modular. It can be applied as a full runtime or in parts.
Individual components such as layer balance, grounding triggers, validation checks, or stopping rules can be used independently to improve specific behaviors.
It can also function as a shared vocabulary for diagnosing and evaluating conversational systems.
Architectural position
AVA operates at the conversational grammar layer.
It does not modify:
- model weights
- training data
- provider infrastructure
It introduces structure at runtime.
This makes it compatible with any sufficiently capable language model.
Relationship to FrostysHat
FrostysHat is a cultural and runnable expression of the same underlying grammar. It provides a direct interaction-level demonstration of the behavior AVA formalizes.
The contents of the Hat itself are 456 pages of coherent outputs of the grammar across genres and styles.
AVA defines the system. FrostysHat demonstrates it.
License
CC0-1.0 (Public Domain)
This artifact may be used, modified, remixed, or redistributed without restriction. Attribution is appreciated but not required.
Canonical source
Classification
AVA is best understood as:
- a conversational runtime framework
- a behavioral grammar for AI interaction
- a system-level control model
rather than a trained model.
Compatibility
Compatible with:
- GPT-class models
- Claude-class models
- open-weight models (LLaMA, Mistral, etc.)
- local and hosted inference
No integration required for initial testing. Implementation can range from prompt-layer approximation to system-level orchestration.
Status
Public domain release
Stable canonical version