AVA
A Conversational Framework
For Coherent AI behavior
License: CC0 1.0
AVA defines the interaction layer of an intelligent system: where model behavior, context, evidence, tone, response structure, and user expectations meet in a real exchange.
The framework gives teams a practical runtime grammar for that layer. It helps a system understand the request before answering, ground what it says before sounding confident, keep the response proportionate, check the behavior before speaking, and stop at a useful endpoint.
AGI may or may not arrive as expected. In either case, greater capability alone will not automatically give AI rules for competent human conversation.
That gap is already visible. Even flagship LLMs can answer correctly while overwhelming the user and making the exchange harder to trust. Current models smooth over uncertainty, lose the plot, lean on fluency, or keep going after the answer arrived.
How to start
The fastest test is simple: run the same realistic exchange twice.
First, ask a model or product flow to handle the exchange under its normal behavior. Then add AVA as context and ask for the same task again. Use the PDF when you want the stable reading version; use the DOCX when you want a model to follow the framework more accurately. Look for changes in grounding, length, uncertainty handling, and closure.
The useful question is: Does AVA make the same exchange easier to use?
AVA can begin as prompt-layer context, but it’s designed to inform product flows, orchestration, retrieval, validation, evaluation, UX, governance, and handoff design. Full integration is ideal; partial adoption is valid. A team might use only a few components to improve the part of the system that needs it.
The Planner Loop
Sense → Decide → Retrieve → Generate → Validate → Close
That sequence is a portable runtime spine. It can be adapted into current stacks on its own, while AVA builds a more complete behavioral framework around it.
Instead of treating output quality as something to patch after the model drafts, the Planner Loop moves conversational quality to the start of the exchange. The system has to understand the request, choose the right work product, gather what the answer must stand on, draft within those constraints, check the behavior before it speaks, and stop at a sufficient endpoint.
The full AVA framework extends that sequence to describe a more coherent conversational system.
AVA Validators
Validate is the enforcement stage of the Planner Loop.
Once the system has moved through Sense, Decide, Retrieve, and Generate, the following validators check whether that response is ready to reach the user. This is where AVA moves from “the model produced an answer” to “the exchange is behaving coherently.”
Containment checks whether the response stays within safety, scope, evidence, and role boundaries.
Drift Control checks whether the answer is still advancing the user’s task or simply continuing.
Layer Balance checks whether the response is proportionate across three active layers:
Performance is how the answer lands on the surface: clarity, tone, pacing, formatting, and ease of use.
Emotion is the human stake in the exchange: concern, pressure, motivation, trust, frustration, uncertainty, or reassurance.
Structure is what the answer can stand on: facts, logic, constraints, evidence, tradeoffs, and what is known or unknown.
AVA asks whether the layer balance fits the moment. Polished answers with weak structure become unreliable; warm answers without grounding can smooth over uncertainty the user needs to see; and technically correct answers that ignore the user’s position can still fail the exchange.
Horizon Progression checks whether meaning is moving in order rather than jumping to conclusions. The stages are:
Formation defines the frame.
Perception names the observed facts or signals.
Duality surfaces the tension, tradeoff, or choice.
Expansion opens bounded possibilities.
Recognition identifies the pattern or principle.
Continuity connects what came before to what should happen next.
Unity reaches an earned conclusion: the point where the exchange can hold its meaning clearly enough to become informed judgment, decision, or wisdom.
In AVA, jumping straight to advice, synthesis, certainty, or abstraction before the earlier stages are established is a progression failure. The system should earn its conclusions instead of performing them.
Recursion Control checks whether the exchange is looping, repeating, or trying to advance without new substance.
Language Hygiene checks whether the response avoids filler, canned phrasing, over-explanation, and unnecessary repetition.
Closure checks whether the work is complete and the system should stop.
Download AVA
AVA (CC0) — Canonical PDF
Use the PDF when you want the stable reading version of the framework for review, sharing, citation, or internal discussion.
AVA (CC0) — Remixable DOCX
Use the DOCX when you want to paste the framework into a model, adapt the language, copy sections into product notes, or work directly with the specification.
GitHub — Technical home
Use GitHub for the repository view: files, copy/paste grammar, testing hypotheses, supporting material, and future development.
FrostysHat (CC0) — Playful runnable version
FrostysHat is the cultural, runnable version of the AVA framework for people who are new to LLMs: part prompt-layer behavior test, part weird systems book, with essays, satire, examples, and validator-style coherence receipts.
Use it when you want to try the grammar quickly in a model, share the idea in a less technical form, or explore the 456-page stranger-looking side of the project.
What AVA helps teams inspect
AVA is useful when an AI system technically works but still feels off in practice.
Support assistants may sound helpful while leaving the user unresolved.
Research assistants may turn limited evidence into overconfident synthesis.
Healthcare assistants may over-reassure.
Legal assistants may make partial document context feel actionable.
Voice agents may sound natural while losing the caller’s correction.
In each case, the issue is behavioral: the interaction shape is creating friction, weak trust, or extra user burden. AVA gives teams language for those patterns before they assume the fix belongs to the model, prompt, retrieval, UX, policy, or evals.
Infrastructure Efficiency
AVA is primarily a behavioral framework, but coherent behavior can also affect system efficiency.
A system that drifts, repeats, over-explains, or keeps asking the user to re-steer the exchange consumes more tokens than the task requires. Cleaner planning, stronger closure, and better state handling can reduce that waste.
The expected effect is shorter resolved conversations, denser turns, less filler, and less raw transcript carried forward as context. In long sessions, state writeback can preserve what matters without treating the entire conversation history as working memory.
These are testable efficiency hypotheses, not assumed savings. Teams can compare AVA-guided behavior against baseline behavior across matched tasks and measure:
output tokens per resolved task
number of turns before completion
re-steer frequency
context-window growth over long sessions
KV cache and memory pressure
total inference cost per completed interaction
The practical claim is modest but important: when an AI system more easily reaches a stable answer, the cost per useful interaction should fall. Even small reductions can compound in high-volume deployments.
For product and evaluation teams
If your AI product technically functions while still feeling hard to trust, hard to steer, or too vague, long, confident, cautious, or demanding for the user, AVA gives you a way to inspect the exchange itself.
Start with the framework, test the grammar, or review the product-domain examples.
If you want a focused outside read of your own system behavior, Human-Grade Review applies the same framework to a transcript, output, workflow, prompt chain, product page, evaluation sample, or small set of related materials.
For agencies, studios, and implementation partners reviewing AI systems on behalf of clients, there’s a separate partner-facing page:
Behavioral Review for AI Agencies and Implementation Partners →
See AVA in practice
Run the one-prompt test: same model, same question, once normally and once with AVA in context. It shows what changes when the exchange has a stronger conversational grammar.