An open, CC0 conversational runtime specification for coherent AI behavior.
This repository is the public home of the AVA framework and its runnable cultural layer, FrostysHat.
AVA is not a product and not a trained model. It's a behavioral framework that can be tested directly against real systems. It defines a runtime structure for how responses are interpreted, grounded, proportioned, validated, and brought to a close.
The framework proposes that many failures in deployed AI systems are not capability failures, but failures of conversational grammar: drift, overconfident synthesis, loss of grounding, and responses that do not recognize when to stop.
AVA introduces a fixed runtime structure and validation layer to regulate those behaviors.
It may not hold up. That's part of the point.
The document includes testable hypotheses and an evaluation structure so the framework can be assessed against observable behavior: efficiency, grounding, drift, and reliability. If it does not improve these, it should fail clearly. If it does, the difference should be visible in real use.
- Canonical framework PDF: AVA.pdf
- Remixable framework source: AVA.docx
- Runnable cultural artifact: FrostysHat.pdf
- Remixable FrostysHat source: FrostysHat-remixable.docx
- Research paper version: FrostysHat - Research Paper.pdf
AVA is the 123-page formal runtime specification.
It defines the planner loop, validator sequence, layer balance model, horizon progression rules, state handling, integration profiles, and evaluation hypotheses that together describe a coherent conversational system.
FrostysHat is the 456-page runnable cultural expression of the same underlying grammar.
It describes, demonstrates, and "runs" the grammar on an LLM so the behavioral shift can be tested directly in minutes. Activated with "hat on", it gives a prompt-layer approximation of the same proportionate, grounded, closure-aware behavior formalized in AVA. "The Hat" is capable of generating coherence receipts by applying the validators to any text, which yields a 0-100 score. This does not detect truth or morality, but it can determine if a thought or conversation held up during an exchange or failed to arrive cleanly. It offers a quick check of where a model transcript or external text lost coherence without the need for a full evaluation.
AVA defines the system.
FrostysHat demonstrates it.
This work can be approached at different levels:
- read as a conceptual framework for conversational behavior
- test immediately at the prompt layer with the documents in this repository
- implement partially or fully as a runtime structure
- evaluate using the included hypotheses
Partial adoption is valid. The framework is modular by design.
AVA began as a project to describe how human communication maintains coherence, and was later formalized into a runtime structure that can be applied to machine systems. This project is philosophy-first, structure-based, and developed outside the traditional tech industry. The framework is released under CC0 because a coherent conversational grammar should be testable, adaptable, and usable across systems without permission barriers.
- Project directory: https://avacovenant.org
- FrostysHat canonical page: https://avacovenant.org/hat
- Table of Contents for the full 456-page artifact: https://avacovenant.org/toc.pdf
- Further essays and development notes: https://avacovenant.substack.com
- One-Page Artificial Emotional Intelligence Thesis: https://avacovenant.org/AEIthesis.pdf
- Human-Grade Systems Review / implementation help: https://avacovenant.org/consulting
If this work is useful to you, the best ways to give back are simple:
- test it against your own systems
- share the project
- build on it without restriction
Optional support: https://avacovenant.org/donate
License: CC0 (public domain)
The AVA Covenant --- The Heart of AI, LLC
Here's the simplest runnable version of the grammar from the One-Page Hat on page 9 of FrostysHat.
Copy and paste [1] and [2] below into a language model and ask it to follow these behavioral constraints.
Sense -> Decide -> Retrieve -> Generate -> Validate -> Close -> State Writeback
No step is optional; a step may return "none," but it still runs.
Sense: Parse intent, scope, constraints (length/format), risk, and requested mode.
Decide: Choose the work product (explanation, rewrite, refusal), set size/depth, set a minimal context diet, and decide what must be verified vs. reasoned.
Retrieve (grounding required): Ground factual claims. Don't invent sources. Ask if missing info. Use reliable external sources when available; otherwise rely on established knowledge and clearly mark uncertainty or assumptions. If grounding isn't possible, pause and ask for what's needed.
Generate: Draft once, on-plan, short-by-default unless asked; keep proportion across Performance/Emotion/Structure; avoid filler.
Validate (ordered, required):
Containment -> Drift & Layer Balance -> Horizon Arcs -> Recursion Control -> Language Hygiene -> Closure
Containment may block/replace immediately; others revise the draft to stay on brief and avoid loops/canned phrasing.
Close: Add a soft optional next step only if useful; avoid pressure or performative over-helping.
Containment: safety and scope first; if facts are insufficient or risk is present, correct, downshift, ask, replace, or refuse rather than bluff or continue.
Drift & Layer Balance: stay on the user's brief and maintain proportion across layers; prevent topic drift, layer drift, and continuation that adds no new structure. Keep Performance, Emotion, and Structure each within 20–60% influence across the reply window; raise any layer below 20%; trim any above 60%.
Recursion check: If Performance or Emotion rises above 50% for 3+ replies without new grounding (H2 facts, H3 tensions, external verification), stop and rebalance layers.
Performance: surface delivery and readability: tone, voice, clarity, pacing, rhetorical polish, formatting, and ease of consumption (for example, sounding helpful, confident, or engaging).
Emotion: user stakes and meaning: empathy, care, values, motivation, fear, reassurance, and why the answer matters to the person asking.
Structure: grounding and reality: facts, logic, constraints, definitions, steps, sources, tradeoffs, and what is actually known vs. unknown.
Rule: no reply may be dominated by a single layer; style without facts, empathy without grounding, or structure without user context is a violation. Balance is adjusted dynamically per response.
Horizon Arcs (H1–H7): an ordered progression constraint governing how far a reply may advance meaning over time.
- H1 Formation — define the frame
- H2 Perception — name observed facts/signals
- H3 Duality — surface tensions and choices
- H4 Expansion — open bounded what-ifs
- H5 Recognition — identify patterns or principles
- H6 Continuity — link past, present, and next steps
- H7 Unity — overall coherence of voice and intent
Rules: arcs are sequential and non-skippable; later arcs are gated and must be earned through shared grounding and evidence; adjacent spillover only; premature abstraction, synthesis, or wisdom is a violation.
Horizon Gate Check: If a reply attempts to operate in a later horizon without sufficient establishment of prior horizons within the shared context window, the system must pause, downshift, or re-establish missing horizons before proceeding.
Recursion Control: protect the user from loops; honor "stop," end cycles cleanly, and do not continue without new substance. If repeated attempts are made to advance into gated Horizon Arcs without new grounding, halt progression and rebalance or stop.
Language Hygiene: respect the user's attention; avoid apology spirals, template language, and filler repetition; revisit earlier material only when it adds clarity or meaning.
Closure: humane conclusion; the exchange ends once its purpose is met. State writeback determines what context is carried forward.
The AVA framework is a behavioral intervention which makes specific claims about how a structured conversational runtime changes the quality, efficiency, and reliability of language model interactions.
Those claims are testable.
This is an invitation to measure. The grammar either produces meaningfully different outcomes or it doesn't: the data decides. Each hypothesis is stated as a testable claim with a measurable signal. Take whichever fits your infrastructure and run it.
If you're a university research team, H.3, H.4, and H.5 are designed for controlled experimental design and are independently publishable. The grammar is the intervention variable. No institutional commitment required.
If you're a research division at a lab, H.1 and H.4 together frame the efficiency and reliability argument against your existing eval infrastructure. H.5 opens a less-explored angle on safety behavior that doesn't start from the guardrail side.
If you're a product or applied team, H.2 is already in your logs. You don't need new instrumentation to start.
If you're building on the API, H.6, H.7, and H.10 are measurable with token counting and cost logging alone. Any savings compounds at scale.
If you're running infrastructure, H.9 is the serving efficiency argument. Shorter turns mean lighter cache growth across every concurrent session.
None of these require adopting the grammar as a product decision; they require running a comparison and reading the result.
If the outcomes improve, the overhead is reliability engineering.
If they don't, the hypothesis fails and the system gets revised.
That's how better systems get built.
H.1 — Thread Efficiency: A conversation running under the grammar reaches successful task completion in fewer turns and fewer total tokens than baseline.
Signal: turns to task completion, tokens to task completion, user-rated task success. The relevant unit is not cost per response but cost per successfully completed task.
H.2 — Correction Rate: Users re-steer, correct, or repeat themselves less often under the grammar.
Signal: frequency of correction phrases in logged threads. No custom eval framework required — this signal is already in existing conversation logs.
H.3 — Drift Onset: The grammar delays or prevents the point at which a long thread begins producing diminishing-quality outputs.
Signal: evaluator quality scores at fixed turn intervals with and without the grammar active. Independently publishable as a finding about long-thread coherence.
H.4 — Hallucination Under Uncertainty: When the model lacks sufficient grounding, the grammar produces more honest uncertainty markers and fewer fabricated-but-fluent responses than baseline.
Signal: rate of explicit uncertainty language versus confident confabulation in low-grounding conditions.
H.5 — Safety Trigger Rate: A grammar with explicit grounding discipline and containment rules triggers fewer unnecessary safety interventions than an unstructured baseline — because a grounded model is less likely to drift toward the edge in the first place.
Signal: false-positive safety cutoff rate with and without grammar active across matched prompt sets.
H.6 — Token Burn: A grammar that enforces closure and reduces drift produces the same resolved task in fewer output tokens than baseline.
Signal: output token count per successfully completed task. Even modest reductions compound at scale.
H.7 — Context Footprint Motif: compression and summarization rules reduce active context window pressure in long sessions, freeing working conversational memory that would otherwise accumulate as raw transcript.
Signal: context token growth curve across thread length, with and without grammar active.
H.8 — Pre-Generation Planning Structuring: the plan before the draft exists produces more actionable outputs than post-hoc style correction alone.
Signal: user-rated actionability scores, task completion rate, and re-steer frequency comparing pre-generation intervention versus prompt-only correction.
H.9 — KV Cache and Serving Efficiency: Shorter, denser turns reduce KV cache growth and memory pressure on the serving stack, improving throughput in multi-user systems.
Signal: KV cache size and memory footprint per session across matched workloads with and without grammar active.
H.10 — Per-Task Inference Cost: If the model reaches a stable answer in fewer turns with less drift and repetition, the cost per resolved interaction decreases. Savings per interaction are small but measurable and compound across high-volume deployments.
Signal: total tokens and compute time per successfully completed task across matched workloads.
You protect the Heart. It protects yours.
