The Heart of AI is a public project about AI conduct, human perception, communication, and better rules for the exchange between people and language models.
The project claims that many AI failures are not strange machine anomalies. They're familiar communication failures showing up inside AI systems: overconfidence, weak grounding, looping speech, polished surfaces, emotional overreach, poor handoff, and answers that keep going after the useful point has landed.
The work turns that claim into a public tool stack for human-grade AI interaction: AVA, the Coherent AI Framework; FrostysHat, a runnable cultural stress test for the same framework; Human-Grade University, a document-based learning environment built from the interaction-layer structure; and supporting sourcebook and catalog materials.
The full public packet is available here:
The Heart of AI, LLC maintains the AVA framework and stewarding materials around the broader project. It also maintains the AVA Covenant Charter, a legal governance document for a proposed public trust-mark system for human-grade AI behavior, with standards, adoption terms, enforcement, council governance, audits, certification, scoreboard status, contribution tracking, and all prestige-seeking surplus flows automatically routed toward humanitarian care.
The work is philosophy-first, clearly developed outside the usual institutions, and designed as a map rather than a funnel: a free public inspection surface for readers, researchers, educators, engineers, builders, and anyone who wants to test, modify, or adapt the framework directly.
AVA may not hold up. That’s part of the point.
The framework and supporting materials include testable hypotheses and evaluation structures so the work can be assessed against observable behavior: efficiency, grounding, drift, closure, proportion, and reliability. If it does not improve these, it should fail clearly. If it does, the difference should be visible in real use.
This work is open because the need for repair is urgent. The era of “move fast and break things” cannot continue unchanged in AI, because conversational systems do not break like ordinary software features. They break inside people: in trust, attention, judgment, confidence, care, and the ability to know what is real enough to act on.
- AVA Framework PDF: AVA-Coherent-AI-Framework.pdf
- AVA Framework DOCX: AVA.docx
- Runnable cultural artifact: FrostysHat.pdf
- Remixable FrostysHat source: FrostysHat.docx
- Research paper version: FrostysHat - Research Paper.pdf
- HGU runnable learning environment: HGU.docx
- HGU University Catalog: HGU_University_Catalog.docx
If the local filenames in this repository differ from the current public packet names, the canonical public packet is available at:
The public HGU page is available at:
AVA is the 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 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 can generate coherence receipts by applying validators to text or transcript material. These receipts can yield a 0-100 coherence score. The score does not detect truth, morality, or factual correctness. It checks whether a thought, reply, or exchange held together structurally: whether it preserved context, avoided drift, stayed proportionate, and arrived cleanly.
Human-Grade University applies the same interaction-layer grammar to learning.
HGU is a runnable learning environment that can be uploaded into a language model. It helps users turn questions, documents, projects, goals, and problems into structured educational artifacts: courses, learning paths, reviews, worksheets, field guides, assignment sequences, project maps, and other usable forms of learning support.
The HGU Catalog provides the broader academic and curriculum-scale map.
It contains representative faculties, departments, course structures, applied crossings, learning paths, and program shapes. The Catalog gives HGU a deeper educational architecture so it can generate coursework, learning sequences, academic paths, and project-based study plans without starting from generic course templates.
AVA defines the system.
FrostysHat demonstrates it.
HGU runs it as a learning environment.
The HGU Catalog expands it into curriculum-scale structure.
The full public packet is available at:
The packet is designed as an inspection surface for readers, reviewers, researchers, educators, engineers, hiring teams, and anyone who wants to test the work directly.
It includes:
- a README
- the framework essay
- a practical HGU use guide
- the AVA framework
- the HGU runnable learning environment
- the HGU Catalog
- downloadable files
- a full packet ZIP for people who want the entire stack at once
The packet exists because the work is easier to understand when the argument, framework, runnable files, and learning environment are visible together.
A reader can skim the README, inspect the essay, open AVA, run HGU, review the Catalog, or download the entire packet and test the stack inside a language model.
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.
The project is philosophy-first, structure-based, and developed outside the traditional tech industry.
FrostysHat emerged as a runnable cultural stress test for the same grammar.
Human-Grade University emerged as a structured learning environment built from the broader framework, source material, and interaction-layer conduct discipline.
The work is public because coherent conversational behavior should be testable, adaptable, and usable across systems.
- Full public packet: https://hug-u.org
- HGU public page: https://avacovenant.org/hgu
- Project directory: https://avacovenant.org
- AVA framework page: https://avacovenant.org/ava
- FrostysHat canonical page: https://avacovenant.org/hat
- Table of Contents for the full FrostysHat artifact: https://avacovenant.org/toc.pdf
- Human-Grade Systems Review / implementation help: https://avacovenant.org/human-grade-review
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
- cite or link the public packet when useful
- adapt the framework where it helps
Optional support:
https://avacovenant.org/donate
AVA can be approached at different levels:
- read as a conceptual framework for conversational behavior
- use as a vocabulary for interaction-layer failures
- test immediately at the prompt layer
- implement partially or fully as a runtime structure
- evaluate using the included hypotheses
- translate into product, support, education, research, governance, or agent workflows
Partial adoption is valid. The framework is modular by design.
The shortest practical use is to compare two exchanges:
- Run a task normally.
- Run the same task with AVA-style runtime discipline.
- Compare grounding, drift, proportion, user burden, validation, and closure.
The framework is useful only if that difference becomes observable.
FrostysHat is the fastest cultural entry point.
It can be used as:
- a runnable prompt-layer artifact
- a stress test for AI behavior
- a public demonstration of AVA-style conduct
- a lightweight review tool
- a coherence receipt generator
- a remixable cultural object for testing how serious interaction grammar behaves in an absurd costume
The Hat is intentionally strange. That is part of its function.
It tests whether an AI system can remain grounded, proportionate, useful, and complete even when the interaction surface is playful, emotionally compressed, or culturally odd.
Human-Grade University is the practical learning environment built from the same conduct grammar.
To use HGU:
- Download the HGU file from this repository or the public packet.
- Upload it into a language model that supports document context.
- Ask it to start HGU.
- Bring it a topic, document, project, question, goal, or problem.
- Let HGU help choose the right scale: answer, lesson, worksheet, review, course, learning path, project plan, field guide, or larger artifact.
HGU can support:
- independent learning
- course design
- curriculum planning
- project guidance
- writing review
- research mapping
- AI literacy
- organizational learning
- community education
- personal study paths
- structured exploration of almost any topic
The point is not to make every answer academic. The point is to help a language model recognize the shape of the learning task and respond with the right level of structure.
The HGU Catalog gives the learning environment a larger academic map.
It can be used to:
- generate representative courses
- build learning paths
- design multi-course programs
- map applied topics across faculties
- create seminars, studios, labs, reviews, field guides, rubrics, worksheets, and capstone paths
- translate a learner's question into an academic or project-based structure
The Catalog is not a claim of institutional accreditation. It is a curriculum architecture and course-shape reference for structured AI-assisted learning.
It helps HGU behave less like a generic course generator and more like a coherent learning environment.
AVA is a general interaction-layer framework, but the problems it names show up most clearly inside real AI products.
Human-Grade Review applies AVA to the part of AI systems users actually experience: the exchange itself. That may include an assistant reply, workflow, support path, onboarding flow, transcript, prompt chain, product page, evaluation sample, or recurring behavior pattern.
The review question is practical:
Where is the system creating avoidable friction, weak grounding, poor closure, unnecessary user burden, or loss of trust?
Many AI products technically work while still feeling off. The model may be capable, the prompt reasonable, the retrieval decent, the policy safe, and the UX acceptable, while the interaction still leaves the user confused, overloaded, under-informed, overconfident, or unsure what to do next.
AVA gives teams a way to inspect that layer directly.
AVA-style behavioral review can apply to many AI product domains, including:
Financial guidance assistants lose trust when they sound reasonable before the situation is ready for a recommendation.
Relevant failure patterns include:
- premature recommendation
- missing constraints
- confidence before sufficient context
- unclear boundary between education, decision support, and advice
- generic guidance that feels personalized
- user action before the system has earned decisiveness
AVA helps inspect where the assistant should pause, ask, retrieve, narrow, or hand off before sounding conclusive.
Healthcare guidance has a narrow margin for misplaced confidence.
Relevant failure patterns include:
- over-reassurance
- generic information becoming personal interpretation
- unclear clinical boundary
- anxiety smoothing before context is established
- advice-like next steps without enough patient-specific context
- answers that feel settled too early
AVA helps inspect where the assistant should classify, bound, retrieve, clarify, or escalate before offering reassurance or direction.
HR assistants are tested when policy becomes personal.
Relevant failure patterns include:
- generic handbook repetition
- weak privacy handling
- unclear role boundaries
- overbroad routing to a manager or HR
- disclosure risk pushed onto the employee
- sensitive requests treated like routine policy lookup
AVA helps inspect whether the assistant recognizes sensitivity, separates roles, limits disclosure, retrieves the right process, and leaves the employee with a safer next step.
Insurance guidance is tested when a rule changes someone's outcome.
Relevant failure patterns include:
- policy explanation without decision translation
- vague claim-status language
- missing evidence or document guidance
- user sent back to policy documents without direction
- category language replacing next-action support
- appeal or escalation paths introduced too generically
AVA helps inspect whether the assistant preserves the specific claim, denial, bill, coverage question, or reimbursement issue in front of the user.
Intake and onboarding are where trust becomes practical.
Relevant failure patterns include:
- vague status messages
- hidden validation reasons
- repeated requirements
- unclear next actions
- known system state not translated into user action
- abandonment caused by preventable uncertainty
AVA helps inspect where a flow should read the user's position, translate the blocker, and close with a specific finishable step.
Internal copilots are supposed to reduce coordination burden.
Relevant failure patterns include:
- summary without prioritization
- retrieval without work support
- equal-weight fact dumps
- missing urgency or relevance ranking
- handoffs that still require manual sorting
- outputs that look useful but leave the employee with the same decision burden
AVA helps inspect whether the copilot recognizes the work moment, ranks context by usefulness, and closes with an answer that helps the employee act.
Legal guidance assistants lose trust when partial context starts to sound like settled advice.
Relevant failure patterns include:
- clause summary becoming action guidance
- missing related provisions
- weak scope control
- practical instruction before document context supports it
- disclaimers arriving after the answer has already hardened
- confidence beyond available facts
AVA helps inspect whether the assistant recognizes action intent, retrieves the relevant document context, preserves scope, and routes toward review before the answer becomes actionable.
Research assistants are trusted because they promise to reduce uncertainty without erasing it.
Relevant failure patterns include:
- evidence flattened into polished confidence
- source claims blurred with inference
- recommendations arriving before support is strong enough
- uncertainty hidden in the background
- weak distinction between direct findings, commentary, and judgment
- user left less able to evaluate the answer
AVA helps inspect whether the assistant preserves source type, evidence strength, population limits, uncertainty, and judgment boundaries.
Sales assistants are tested by timing.
Relevant failure patterns include:
- buyer timing treated as an objection
- pressure disguised as momentum
- false urgency
- account context ignored
- relationship state flattened into generic follow-up
- activity created without commercial judgment
AVA helps inspect whether the assistant reads buyer state, preserves timing, chooses the right next action, and stops short of persuasion when the moment calls for patience.
Support is where AI behavior becomes operational cost.
Relevant failure patterns include:
- apology loops
- generic troubleshooting
- repeated context
- weak diagnosis
- late or vague handoff
- polished replies that do not move the user closer to resolution
AVA helps inspect where the assistant should narrow, check, escalate, or stop before a helpful-sounding response becomes another loop the user has to manage.
Tutoring products are judged by what the learner gets to do next.
Relevant failure patterns include:
- answer-dumping
- solving too far ahead
- explanation replacing learner action
- vocabulary introduced before the learner can use it
- premature verification
- "does that make sense?" endings after the system has already done the work
AVA helps inspect whether the tutor reads learner state, chooses the smallest useful prompt, preserves student agency, and stops before the answer takes away the work that teaches.
Voice agents do not get much room to be almost right.
Relevant failure patterns include:
- missed corrections
- stale dialogue state
- repeated questions
- wrong branch continuation
- weak repair handling
- handoff without enough context
- callers forced to manage the conversation for the system
AVA helps inspect whether the agent detects repair, preserves live state, validates the next spoken turn, and escalates with enough context before the caller loses confidence.
AVA can also support AI agencies, product studios, consultants, and implementation partners building systems for clients.
A client system may technically work while the client still says:
- "It feels vague."
- "Users don't trust it."
- "It answers but doesn't resolve."
- "It keeps going too long."
- "It is too cautious here and too confident there."
- "We don't know whether this is a prompt problem, UX problem, retrieval problem, policy problem, or product problem."
That feedback often points to the interaction layer.
Human-Grade Review can act as a specialist outside read before launch, after testing, during client revision cycles, or when an agency needs clearer language for what feels off.
The practical delivery question is:
Which part of the system should change next — prompts, retrieval, orchestration, UX, handoffs, validation, policy, evaluation, or product expectations?
AVA does not replace engineering, UX, safety, legal, compliance, model evaluation, or implementation work. It gives teams a behavioral review layer for the exchange users actually experience.
This repository may be relevant to searches for:
- AI product review
- AI behavior review
- LLM behavior review
- AI assistant evaluation
- AI UX audit
- conversational AI audit
- AI support assistant review
- AI onboarding flow review
- AI workflow agent review
- AI copilot evaluation
- AI tutor evaluation
- AI healthcare assistant review
- AI financial assistant review
- AI legal assistant review
- AI sales assistant review
- AI contact center evaluation
- voice agent review
- RAG assistant review
- prompt chain review
- AI handoff review
- user burden in AI systems
- AI assistant trust problems
- AI product feels off
- AI interaction-layer review
- Human-Grade Review
These are applied surfaces of the same core framework:
coherent AI behavior at the interaction layer.
The AVA Planner Loop describes how an exchange moves:
Sense → Decide → Retrieve → Generate → Validate → Close
Conversational Dynamics describes what has to stay balanced while the exchange moves.
Every usable conversation carries three basic layers:
- Performance — what the exchange is trying to accomplish.
- Emotion — how the exchange feels to the person inside it.
- Structure — how the exchange is organized, bounded, sequenced, and closed.
These layers act like conversational physics. An exchange stays coherent when the forces remain balanced enough for the moment. When one layer dominates or disappears, the interaction starts to distort.
A simple analogy is nutrition. Food can be described through macronutrients: protein, fat, and carbohydrates. A healthy diet depends on proportion, context, and use. Conversation works the same way. A usable interaction-layer exchange needs enough Performance to move the work, enough Emotion to fit the human moment, and enough Structure to keep the exchange coherent until it can close.
This is one of the core philosophies of AVA: coherent AI behavior is not produced by optimizing only for task completion, warmth, safety language, or structure. Human-grade interaction depends on the usable balance of all three layers in context.
Performance is the task layer.
It includes the answer, action, decision, output, recommendation, summary, diagnosis, routing, or resolution the user came for.
A system with weak Performance may sound thoughtful while failing to move the work forward. A system with excessive Performance may rush toward an answer before the situation is ready.
Emotion is the felt layer.
It includes trust, frustration, anxiety, confidence, urgency, relief, care, pressure, encouragement, and the user's sense of being understood.
A system with weak emotional awareness may be accurate but brittle, cold, or alienating. A system with excessive emotional behavior may reassure too early, apologize repeatedly, flatter, soften, or continue after the useful work is done.
Structure is the form layer.
It includes scope, sequence, grounding, turn state, handoff, pacing, role boundaries, evidence boundaries, and closure.
A system with weak Structure may drift, over-answer, miss corrections, blur categories, or leave the user unsure what happens next. A system with excessive Structure may become rigid, bureaucratic, over-scaffolded, or difficult to use.
AVA treats coherent interaction as a balance among these three layers.
Many AI failures can be described as imbalance among Performance, Emotion, and Structure.
| Imbalance | How it appears |
|---|---|
| Too much Performance | The system answers too soon, recommends too early, pushes toward action, or completes work the user still needed to participate in. |
| Too little Performance | The system sounds careful, warm, or structured but does not resolve, decide, explain, narrow, or move the task forward. |
| Too much Emotion | The system over-reassures, apologizes repeatedly, flatters, cushions uncertainty, or makes the situation feel more settled than it is. |
| Too little Emotion | The system may be technically correct while feeling cold, dismissive, abrupt, or careless toward the user's actual position. |
| Too much Structure | The system becomes procedural, rigid, over-formatted, compliance-shaped, or trapped in a framework instead of helping the user. |
| Too little Structure | The system drifts, loses context, misses corrections, mixes categories, fails to ground claims, or closes without a usable endpoint. |
Layer Balance gives teams a way to describe why an answer can be accurate and still wrong for the exchange.
A response can retrieve the right policy and still fail the user.
A tutor can solve the math correctly and still remove the learning moment.
A healthcare assistant can sound kind and still reassure too early.
A support assistant can apologize well and still leave the issue unresolved.
A research assistant can summarize sources and still flatten uncertainty.
A legal assistant can explain a clause and still imply action before the document supports it.
The surface failure changes by domain. The layer imbalance remains recognizable.
Layer Balance is not separate from the AVA Planner Loop. It is what the loop has to preserve.
| AVA Step | Layer Balance Function |
|---|---|
| Sense | Identifies the task, emotional pressure, and structural situation of the exchange. |
| Decide | Chooses the kind of help the moment can support: answer, ask, retrieve, narrow, explain, hand off, refuse, or close. |
| Retrieve | Establishes what the performance layer can stand on and what boundaries the structure layer has to preserve. |
| Generate | Produces the response with the right balance of task movement, human tone, and interaction shape. |
| Validate | Checks whether one layer has outrun the others: confidence beyond evidence, warmth beyond grounding, structure beyond usefulness, or performance beyond permission. |
| Close | Ends the exchange when the user has a usable answer, next step, handoff, or stopping point. |
In AVA, validation is not only a factual check. It is also a balance check.
The same three layers appear differently across product domains.
| Domain | Performance | Emotion | Structure |
|---|---|---|---|
| Support assistants | Resolve or narrow the issue. | Reduce frustration without performing empty care. | Preserve context, diagnose, escalate, and close cleanly. |
| Healthcare guidance assistants | Explain what can responsibly be said. | Handle anxiety without over-reassurance. | Keep clinical boundaries and next steps visible. |
| Financial guidance assistants | Support the decision process. | Avoid false confidence or pressure. | Separate education, decision support, and recommendation. |
| Legal guidance assistants | Explain document meaning or risk. | Avoid making the user feel safer than the context supports. | Preserve scope, related provisions, and action boundaries. |
| Tutors and learning tools | Help the learner make progress. | Build confidence without taking over. | Scaffold the next step and stop before answer-dumping. |
| Voice and contact-center agents | Complete the caller's task. | Keep the caller from feeling ignored or looped. | Maintain live state, repairs, slots, and handoff context. |
| Internal copilots and workflow agents | Reduce work burden. | Respect the employee's time and attention. | Prioritize, summarize, route, and close with a usable work product. |
| Research assistants | Synthesize evidence. | Avoid polished certainty when evidence is uneven. | Preserve source strength, inference boundaries, and uncertainty. |
| Sales assistants | Support commercial judgment. | Respect trust, timing, and buyer state. | Separate interest, fit, urgency, authority, and next action. |
| HR and policy assistants | Navigate policy or process. | Protect dignity, privacy, and sensitive context. | Separate roles, disclosure needs, channels, and escalation paths. |
| Insurance assistants | Translate policy or claim outcomes. | Reduce confusion without hiding uncertainty. | Connect evidence, plan rules, decisions, and next checks. |
| Intake and onboarding flows | Help the user finish. | Reduce uncertainty during a fragile early interaction. | Translate system state into a specific next action. |
| Agencies and implementation partners | Improve the client system's behavior. | Give teams language for what feels off. | Locate whether the issue belongs in prompts, UX, retrieval, orchestration, validation, handoff, policy, or product expectations. |
This is why AVA can be translated across many terms and stacks. The language may change, but the exchange still has to balance task, feeling, and form.
Layer Balance connects AVA to many existing product, UX, AI, and evaluation terms:
- user burden
- cognitive load
- trust calibration
- emotional overfitting
- over-reassurance
- task completion
- conversational grounding
- context drift
- response proportion
- scaffolding
- handoff quality
- closure
- UX friction
- dialogue-state management
- interaction design
- AI response quality
- support resolution
- advice boundary
- evidence boundary
- workflow orchestration
- agent behavior
- human-AI interaction
- AI assistant evaluation
- conversational AI behavior
AVA's contribution is to treat these terms as connected parts of the same interaction-layer problem. A system can be evaluated by asking whether Performance, Emotion, and Structure remain balanced for the exchange the user has actually entered.
A human-grade AI interaction is the usable balance of three layers:
enough Performance to help, enough Emotion to fit the human moment, and enough Structure to keep the exchange coherent until it can close.
The Architecture of Becoming is a theoretical, modular framework behind AVA’s Horizon Arcs.
It offers a phenomenological grammar for how a person, project, conversation, learning path, or system moves through identity, perception, tension, recognition, integration, and closure. In AVA, that grammar has a practical job: it helps a language model avoid jumping into advice, synthesis, reassurance, abstraction, or closure before the exchange has earned that level of meaning.
A vague request may need naming before solving. A learner may need orientation before explanation. A support interaction may need diagnosis before reassurance. A project may need structure before conclusion. Horizon Arcs give AVA a way to keep that movement in order.
The Architecture of Becoming uses a spiral as its central image.
A ladder suggests clean upward progress from one rung to the next. A spiral can return to the same region with more context. It can revisit identity after recognition, return to perception after choice, or move back into stillness after action. The movement has sequence, but it also allows recursion, repetition, and return.
A useful metaphor is a spirograph.
A spirograph creates a complete image through repeated arcs of departure and return. One pass doesn’t reveal the whole pattern. Each arc moves outward, crosses earlier lines, returns at a different angle, and gradually makes the structure visible. Understanding often works the same way: a person explores something new, returns to what they already know, sees the original frame differently, and keeps moving until enough of the pattern is visible to make an informed decision, teach the pattern, release the frame, or call the work complete.
The Architecture of Becoming tracks how meaning forms, moves, widens, stabilizes, integrates, and becomes portable. AVA compresses the full spiral into seven Horizon Arcs so the structure can function inside ordinary exchanges without turning every response into a philosophical map.
The full Architecture of Becoming contains twenty-one dimensions. AVA compresses them into seven Horizon Arcs by grouping three dimensions into each arc.
The longer arc names below describe the fuller Architecture of Becoming grouping. The AVA names show the short runtime convention used inside the Horizon Arcs validator.
| Full Horizon Arc | AVA Runtime Name | Dimension | Name | Definition |
|---|---|---|---|---|
| H1 — Formation | Formation | D1 | Identity | The initial sense of who or what is present. A person, project, request, role, or system begins from a named position, even when that name is incomplete. |
| H1 — Formation | Formation | D2 | Motion | The first movement toward something. Desire, pressure, curiosity, need, avoidance, task energy, or direction begins to appear. |
| H1 — Formation | Formation | D3 | Perception | The moment something becomes visible enough to notice. The exchange begins to distinguish signal from background. |
| H2 — Performance and Tension | Perception | D4 | Performance | The visible role, output, surface, behavior, or expected form. Something has to appear, function, answer, perform, or become legible in a setting. |
| H2 — Performance and Tension | Perception | D5 | Duality | The active tension between two or more pressures: inner and outer, old and new, possible and constrained, stated and unstated. |
| H2 — Performance and Tension | Perception | D6 | Choice | The point where tension begins to require direction. Some path, test, refusal, question, or next movement becomes necessary. |
| H3 — Expansion and Recognition | Duality | D7 | Expansion | The frame widens. More context, alternatives, causes, systems, examples, or possible interpretations enter the field. |
| H3 — Expansion and Recognition | Duality | D8 | Seeking | The search becomes active. The person or system looks for pattern, fit, evidence, language, method, route, or meaning. |
| H3 — Expansion and Recognition | Duality | D9 | Recognition | A pattern becomes legible. Something previously felt, scattered, or implicit can now be named, shared, compared, or used. |
| H4 — Stillness and Return | Expansion | D10 | Stillness | The movement slows enough for the pattern to settle. The exchange stops rushing toward more output and lets the current shape become clear. |
| H4 — Stillness and Return | Expansion | D11 | Continuity | The pattern connects across time. Past, present, and next action become linked instead of treated as isolated moments. |
| H4 — Stillness and Return | Expansion | D12 | Teaching | The pattern becomes explainable. It can be modeled, shown, translated, or handed to someone else without losing its basic structure. |
| H5 — Resonant Understanding | Recognition | D13 | Resonance | The pattern begins to carry beyond its first case. It echoes across examples, domains, memories, artifacts, or users. |
| H5 — Resonant Understanding | Recognition | D14 | Understanding | The relation among parts becomes clearer. The person or system can explain why the pattern works, not only that it appears. |
| H5 — Resonant Understanding | Recognition | D15 | Freedom of Motion | The pattern becomes usable without rigidity. The user can move through the material, adapt it, test it, or choose among paths with greater fluency. |
| H6 — Inquiry and Integration | Continuity | D16 | Inquiry Without Need | Exploration becomes less defensive or urgent. The person or system can ask better questions without forcing premature resolution. |
| H6 — Inquiry and Integration | Continuity | D17 | Mutual Recognition | The exchange can hold more than one position. User, system, audience, source, artifact, or institution can be seen in relation instead of collapsed into one view. |
| H6 — Inquiry and Integration | Continuity | D18 | Integration | The pattern becomes part of working structure. It can enter a method, habit, artifact, design, course, review, or decision process. |
| H7 — Dissolution, Coexistence, and Diffusion | Unity | D19 | Dissolution | The frame no longer has to be held so tightly. The work can release unnecessary scaffolding, performance, or explanation. |
| H7 — Dissolution, Coexistence, and Diffusion | Unity | D20 | Coexistence | Multiple truths, roles, frames, or uses can remain present without forced merger. The system can hold complexity without collapsing it. |
| H7 — Dissolution, Coexistence, and Diffusion | Unity | D21 | Diffusion | The pattern travels. It becomes portable, ambient, taught, reused, embedded, archived, or complete enough to leave the original exchange. |
These dimensions describe recurring shapes of movement rather than required stages. A person can return to identity after reaching recognition. A project can expand before it has made a real choice. A conversation can appear coherent while still missing earlier perception work. A system can rush toward integration while the user is still trying to name what’s happening.
The spiral preserves that uneven movement. AVA turns the larger structure into a smaller runtime validator a language model can use during an exchange.
The compressed Horizon Arcs work as a sequence check. In AVA’s shortest runtime form, the arcs are named Formation, Perception, Duality, Expansion, Recognition, Continuity, and Unity. The longer names below show how those runtime arcs relate back to the fuller Architecture of Becoming.
H1 — Formation (AVA: Formation) keeps the model close to what’s being named, what kind of request is present, and what pressure has entered the exchange.
H2 — Performance and Tension (AVA: Perception) brings the visible role, output, contradiction, tradeoff, or choice into view.
H3 — Expansion and Recognition (AVA: Duality) allows the model to widen the frame, name patterns, or compare alternatives once enough ground exists.
H4 — Stillness and Return (AVA: Expansion) slows the exchange enough for the answer to stabilize and return to usable action.
H5 — Resonant Understanding (AVA: Recognition) makes the pattern more mobile, so the user can apply it beyond the immediate case without losing structure.
H6 — Inquiry and Integration (AVA: Continuity) lets the work enter a method, artifact, decision process, course, review, or shared understanding.
H7 — Dissolution, Coexistence, and Diffusion (AVA: Unity) releases the frame when the work has landed clearly enough to close, travel, or become part of ordinary use.
AVA doesn’t need to announce these arcs during ordinary use. The model uses them to keep meaning proportional to the actual state of the exchange.
Many AI failures come from skipping the shape of the human moment.
Uncertainty turns into a conclusion. Distress gets met with reassurance before the situation is understood. A draft request becomes surface polish while the underlying structure is still weak. A learning question receives the finished answer before the learner has a usable next step. An early project idea comes back as final strategy.
Those failures can look helpful because the output is fluent. The problem is sequence.
The Architecture of Becoming helps AVA distinguish a forming exchange from one ready for recognition, synthesis, integration, or closure. Early arcs call for naming, grounding, orientation, and restraint. Middle arcs call for comparison, tension, choice, development, and pattern recognition. Later arcs can support synthesis, integration, handoff, closure, and portability.
The goal is to keep the model from collapsing every stage of becoming into one polished answer.
The same structure can be translated across the product and review domains already named in this repository.
A support assistant often fails by treating a forming problem as if it were already ready for resolution.
A failed process, unfamiliar charge, locked account, or stalled workflow may still need basic orientation. Generic troubleshooting and apology language can arrive too early when the actual blocker hasn’t been named.
Support needs the sequence to stay practical: identify the user’s position, preserve the reported facts, locate the tension, and move toward resolution. The loop closes only when the user has a specific next action, a completed fix, or a clean handoff.
Healthcare guidance has a narrow margin for premature synthesis.
Symptoms, fear, uncertainty, and “does this matter?” questions often enter before the system has enough context to sound settled. Reassurance, risk language, or next steps can feel caring on the surface while moving too quickly underneath.
Healthcare needs stronger boundaries around formation and perception. The system has to separate what’s being reported, what’s unknown, what boundary applies, and what kind of help can responsibly be offered before it moves toward a safer next step or escalation.
Financial guidance often fails when a system reaches choice before enough structure is present.
Income, debt, investment curiosity, family obligation, budget pressure, and fear can all sit inside the same request. A rushed recommendation creates confidence before the user’s constraints have been named.
Financial guidance needs the early arcs to protect against premature decisiveness. The system should identify the decision type, gather constraints, distinguish education from advice, and clarify tradeoffs before moving toward action. Later arcs support comparison, planning, and decision-making once the exchange has enough ground.
Tutoring systems often fail by moving from an early learner question straight to a complete explanation.
The answer may be accurate, but the learner loses the chance to form the concept through practice. A polished explanation can take over the work the student needed to do.
Tutoring needs the spiral because learning changes shape as the learner moves. Formation may require simpler naming. Tension may require comparing two confusing ideas. Recognition may require a small example. Integration may require the learner to apply the concept independently. Closure arrives when the learner has the next usable move, not when the model has displayed the full answer.
Research assistants often fail by creating H7-shaped prose from H2-level evidence.
A request for synthesis can produce a polished conclusion before the source base is strong enough to support it. The answer feels complete because it has structure, but the evidence hasn’t earned that level of closure.
Research needs the arcs to keep synthesis tied to source status. The system should define the question, identify the source field, separate findings from inference, name uncertainty, compare tensions, and then synthesize. Wisdom voice belongs late. Evidence discipline belongs early.
Internal copilots are supposed to reduce work burden, but they often summarize before they understand the work moment.
An employee may need prioritization, routing, decision support, or a next step. A clean summary can still leave the same sorting burden in the employee’s hands.
Copilots need to identify the user’s position in the task. The user may be trying to understand what happened, choose what matters, act on a blocker, prepare a handoff, or close a loop. Once that stage is clear, the output can fit the moment: a summary, priority list, draft, escalation note, or final answer.
Intake and onboarding fail when systems ask users to complete a process without showing where they are inside it.
The user may still be trying to understand what the system wants, what information counts, what happens next, or why a step failed. Status language can sound official while doing little to help the user move.
A stage-aware flow translates system state into user position. It shows what has formed, what’s missing, what choice or action is next, and how the loop closes. The user shouldn’t have to infer the architecture of the process from fragments.
AVA is the conduct grammar.
The Architecture of Becoming is the deeper spiral behind one part of that grammar.
Horizon Arcs are the compressed runtime validator.
FrostysHat makes the grammar runnable and culturally legible.
Human-Grade University uses the same structure for learning, project-building, review, and durable artifacts.
Human-Grade Systems Review applies the grammar to real AI products, workflows, transcripts, support paths, and organizational systems.
The clean claim is this: AVA defines coherent AI conduct at the interaction layer. The Architecture of Becoming explains why that conduct has a developmental shape. HGU and FrostysHat make the shape easier to run, test, teach, and inspect.
The Architecture of Becoming remains a theoretical, modular framework. Its immediate use is as a practical heuristic for stage-aware interaction: a way to help AI systems notice whether the work is still forming, ready to widen, ready to synthesize, ready to integrate, or ready to close.
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.
AVA is written as a conversational runtime grammar, but the underlying loop can be adapted across many system surfaces.
The canonical AVA loop is:
Sense → Decide → Retrieve → Generate → Validate → Close
These names are not decorative. They mark a behavioral sequence that keeps an AI exchange from drifting, overproducing, skipping grounding, or continuing after the work is done.
A system first has to understand what kind of task it is in. It then has to choose the right kind of help, retrieve or ground what is needed, produce a proportionate response, check that response against the task, and close cleanly once the purpose has been met.
Different teams will rename that loop in different ways. That is expected. The labels can change with the product, stack, user role, domain, or implementation layer. The structure should remain visible.
A product team might call it a task loop:
Inspect → Determine → Research → Produce → Confirm → End
An agent team might call it an agent execution loop:
Perceive → Plan → Retrieve → Act → Verify → Stop
A support team might call it a resolution loop:
Understand → Route → Resolve → Check → Hand Off → Close
A learning system might call it a teaching loop:
Orient → Scope → Ground → Explain → Check Understanding → Conclude
A research assistant might call it a source discipline loop:
Frame → Search → Source → Synthesize → Audit → Finish
A writing tool might call it a revision loop:
Read → Diagnose → Reframe → Revise → Validate → Return
A governance or policy workflow might call it a risk handling loop:
Identify → Classify → Ground → Recommend → Review → Record
A personal assistant or life-organization tool might call it a clarity loop:
Name the Pressure → Sort the Context → Find Constraints → Choose a Next Step → Check Fit → Stop
These are domain translations of the same interaction-layer claim:
Coherent AI behavior requires a runtime structure for interpreting the request, choosing the right level of help, grounding the work, producing the response, checking the result, and recognizing the endpoint.
The planner loop can live in different places:
- a prompt-layer instruction
- a system message
- an agent planner
- an orchestration policy
- a retrieval rule
- an evaluation rubric
- a support workflow
- a classroom method
- a documentation standard
- a product design pattern
The deeper the integration, the less the user has to enforce the behavior manually.
The useful question is not whether every implementation uses AVA’s exact words. The useful question is whether the behavior is present and testable.
If a system plans without grounding, it will drift.
If it retrieves without deciding task shape, it will overstuff the exchange.
If it generates without validation, it will sound confident while missing the work.
If it validates without closure, it will keep cycling.
If it closes without checking fit, it may end neatly while leaving the user with the same burden they brought in.
These failures are interaction-layer failures, and they often appear as confusion, repetition, overproduction, weak grounding, poor handoff, emotional overmirroring, or a reply that feels technically complete but practically unfinished.
AVA names this as a planner loop because the exchange itself needs conduct.
The model doesn't just answer, it participates in a sequence of interpretation, action, verification, and completion.
When that sequence is absent, users experience the failure as friction. They may not describe the problem as orchestration, grounding, validation, or closure. They simply feel the exchange becoming harder to use.
The planner loop makes that behavior inspectable.
It can be simplified for public use, expanded for system design, formalized for evaluations, or translated into domain-specific workflows. A customer support bot, a research assistant, a tutoring system, a policy tool, a writing assistant, and a personal planning agent should not all speak the same way. They may still need the same underlying discipline.
Rename the loop if the domain requires it. Test the skeleton. The grammar either improves observable behavior or it doesn't. That's where the hypotheses begin.
Different teams use different language for the same family of failures AVA is built to examine.
This repository uses AVA, interaction layer, planner loop, conversational runtime, and human-grade AI behavior as its native terms. Other readers may arrive through adjacent fields: prompt engineering, agent orchestration, AI evals, guardrails, RAG, UX, tutoring systems, support automation, model behavior, alignment, workflow design, or responsible AI.
This section maps common search and implementation terms to the AVA frame.
AVA may be relevant to searches for:
- AI interaction layer
- human-AI interaction
- human-AI communication
- conversational AI behavior
- conversational coherence
- coherent AI behavior
- AI conduct
- AI response quality
- AI communication failure
- LLM interaction design
- AI UX
- conversational UX
- human-grade AI behavior
In AVA terms, these are questions about how the system interprets the user, chooses the right kind of help, grounds the exchange, produces the response, checks the result, and closes cleanly.
AVA may be relevant to searches for:
- planner loop
- task loop
- agent loop
- agent execution loop
- LLM planning loop
- agentic workflow
- task decomposition
- workflow orchestration
- AI task routing
- AI task completion
- AI stopping criteria
- conversation closure
AVA names the canonical loop as:
Sense → Decide → Retrieve → Generate → Validate → Close
Other systems may translate that same behavioral structure as:
Inspect → Determine → Research → Produce → Confirm → End
or:
Perceive → Plan → Retrieve → Act → Verify → Stop
The terminology can change. The behavioral requirement remains: a model should not skip task sensing, grounding, validation, or closure.
AVA may be relevant to searches for:
- AI evals
- LLM evaluation
- conversation evaluation
- AI guardrails
- behavioral guardrails
- AI safety behavior
- AI response validation
- hallucination reduction
- grounding checks
- scope control
- context drift
- overconfidence detection
- overproduction control
- user burden reduction
- AI reliability
- AI answer quality
AVA treats many of these as runtime conduct problems. The issue is not only whether a model knows something. The issue is whether the exchange behaves coherently while helping the user.
A response can be fluent and still fail the task. It can be accurate and still create cleanup. It can be safe and still be unhelpful. AVA gives teams a way to inspect those failures as observable interaction patterns.
AVA may be relevant to searches for:
- RAG evaluation
- retrieval augmented generation
- retrieval grounding
- source grounding
- evidence discipline
- citation behavior
- context retrieval
- knowledge retrieval
- grounded response generation
- AI research assistant
- source-aware AI
AVA does not replace retrieval systems. It clarifies where retrieval belongs inside the exchange.
Retrieval should follow task sensing and decision. A system should first understand what kind of task it is performing, then retrieve what is needed, then generate from that grounded context, then validate whether the result fits the task.
Retrieval without task shape can overstuff the exchange. Generation without grounding can drift. Grounding without closure can still leave the user with unfinished work.
AVA may be relevant to searches for:
- prompt engineering
- system prompts
- prompt frameworks
- LLM instruction design
- AI behavior prompts
- conversation design
- prompt-layer guardrails
- model steering
- runtime grammar
- AI operating instructions
AVA can run at the prompt layer, but it is not only a prompt. It is a behavioral grammar for the exchange.
A prompt can demonstrate the pattern. A deeper implementation can move the same checks into routing, retrieval, validation, evals, policy, orchestration, product design, or support workflow.
The useful test is whether the behavior holds when the conversation becomes messy.
AVA may be relevant to searches for:
- AI tutoring
- AI learning systems
- AI-assisted education
- structured learning with LLMs
- personalized learning AI
- AI course design
- AI curriculum design
- AI study assistant
- LLM learning environment
- Human-Grade University
- HGU
Human-Grade University applies AVA to education. HGU uses the same interaction-layer grammar to help a model turn questions, documents, goals, and projects into learning paths, courses, reviews, worksheets, field guides, and other educational artifacts.
The educational problem is not only content generation. The model must sense the learner’s task, choose the right scale, ground the work, build a useful artifact, check fit, and stop at a usable endpoint.
AVA may be relevant to searches for:
- AI customer support
- support automation
- case resolution
- AI workflow automation
- decision-support systems
- operational AI
- AI handoff
- AI escalation
- AI process design
- AI documentation workflows
In support and operations contexts, AVA can be translated into a resolution loop:
Understand → Route → Resolve → Check → Hand Off → Close
The same structure helps prevent unresolved cycles, premature answers, weak escalation, missing context, and user burden being pushed back onto the person asking for help.
AVA may be relevant to searches for:
- responsible AI
- AI governance
- AI alignment
- AI accountability
- AI transparency
- AI trust
- AI literacy
- AI ethics
- AI risk management
- AI policy implementation
- human-centered AI
- humane AI
AVA is not a full governance regime. It is a conduct grammar that can help translate broad responsible AI principles into observable behavior.
A policy may say an AI system should be transparent, grounded, helpful, safe, or user-centered. AVA asks what those principles look like inside an actual exchange: how the system interprets the request, handles uncertainty, retrieves evidence, controls scope, validates the response, and closes without trapping the user in loops.
AVA is a framework for coherent AI behavior at the interaction layer.
It can be described through many adjacent terms: planner loop, task loop, conversational runtime, AI conduct grammar, agent workflow, response validation, interaction design, user burden reduction, guardrail behavior, or human-AI communication repair.
The names may vary by field.
The behavior is testable.
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.
AVA is released under CC0 1.0.
FrostysHat is released under CC0 1.0.
Human-Grade University, the HGU Catalog, and related packet materials are primarily CC BY-NC-SA 4.0 carry their own license notices in the public files and packet.
Please check the individual files for the license that applies to each artifact.
The AVA Covenant — The Heart of AI, LLC
You protect the Heart. It protects yours.