Ecological AI — PLAYi.io

Every AI system ever built can only do half of what coupling requires.

We're building the other half. Not a new architecture on top of the old paradigm — the ontological floor the old paradigm is missing.

Start a conversation → See the architecture
30tps
Geometric tokens/sec
Raspberry Pi 5 · $80 hardware
48
Frozen edge laws
Structurally impossible to violate
~35
Reps to convergence
Dogs · LLMs · neural nets · same constant
0
Illegal tokens generated
Not filtered — architecturally impossible
The Problem

Chronos-based cooperation
is a logical impossibility.

Every cooperative AI and robotics system ever built runs on clock time — scheduled turns, token passing, synchronized handoffs. But coupling doesn't happen on a schedule. The swap between Awareness and Attention happens when the field says now. You cannot engineer now. You can only build a system with the ecological sensitivity to perceive it.

"There is no time in vaulting. Only whether the dog left the ground for the target."

This is why robotics cooperation is brittle. Not because the algorithms aren't fast enough. Because they're running on the wrong kind of time. Kairos — the right moment — is disclosed by the field, not the clock. No chronos timing will ever afford coupled cooperative movement in the complex domain. It is a philosophical impossibility.

What every AI lab is building
Rule stacking. Preference modeling. Reward shaping. Post-hoc moderation. Systems that can optimize outputs, follow instructions, adapt statistically — but cannot hold or yield Initiative, remain coupled under uncertainty, recognize when control should pass, or coordinate without external reward scaffolding.

These approaches manage behavior. They do not develop agency.
What alignment actually requires
The capacity to remain coupled while leadership changes. Not obedience. Not safety rails. Not value loading.

Alignment is not a property you bolt on after training. It is the geometry the agent learns through.

The alignment constraint is the loss surface, not a filter on the output.

An agent that cannot lawfully pass Initiative cannot be aligned — only constrained.

The 9 Primal Skills

AI can only do the right column.

Every skilled interaction requires two complementary roles held simultaneously — one by each coupled agent. The left column skills are the master skills. They afford and constrain the right column. When coupled, one agent holds the left while the other holds the right. They swap in kairos time — so fast it is hard to see.

Current AI cannot hold the left column at all. So it cannot genuinely couple. It can only simulate coordination from one side of the equation.

Master skill — AI cannot do this Derived skill — AI does this instead
Awareness affords ↔ Attention
Initiative affords ↔ Anticipation
Dismissal affords ↔ Engagement
Affordance affords ↔ Partnership
Flow affords ↔ Function
Coupled Movement affords ↔ Position
Coupling affords ↔ Opportunity
Presence affords ↔ Synergy
Passing affords ↔ Possession

"The Primal Games do not create behaviors. They develop skills. These are very different things."

Tonic / Phasic — The Perception Manifold

Two modes of disclosure. Four quadrants of agency.

Tonic holds the field — invariant information, what doesn't change across situations. Phasic marks the event — specifying information, what matters right now. PLAY is the only affect that decays into itself (pPLAY → tPLAY). Every other affect cascades toward resolution. This is why only PLAY permits Initiative Transfer.

EXPLORE THE FULL MANIFOLD →

PrimalEPT — Architecture

The geometry was always there.
We named it. Then we made it lawful.

The standard neural network computational graph — weights, activations, biases flowing to a pre-activation node — is a tetrahedron. Stack two and you have stacked tetrahedra. Every neural network ever built was always this geometry. Everyone drew it as a tree and saw sequence. PrimalEPT named the vertices and froze the lawful edges between them.

01 — Vocabulary
16 Ontological Trits
Not magnitude polarities. Ontological identities. Each of the 16 Primal Elements carries its family membership, its role within that family, and its lawful relationship to every other element. The hex letter is the trit. Awareness = a, Attention = b, Initiative = c. Each value knows what it is before training begins.

In the standard backprop graph: w (weights) = Barycenter — the grip, earned, stable. z (pre-activation) = Mediator — the live coupling, perception-action in the moment. b (bias) = Disposition. a (prior activation) = Relation. The standard neural network was always this. Nobody named the vertices.
02 — Laws
48 Frozen Edge Laws
Across 4 solids. 24 affording, 24 constraining. These are not soft penalties applied during training — they are structural zeros in the weight matrix that do not exist as learnable parameters. The model cannot learn an illegal weight. It can only learn lawful extensions of the existing geometry. The laws ARE the loss surface.
03 — Earned Weights
Half Earned, Half Disclosed
D and R primitives (8 of them) are earned from field experience through the Ludovico technique — gestalt ecological perception in real environments. B and M primitives are computed live: B = cos² (the grip, tonic, what holds). M = sin² (the push, phasic, what moves). The live geometry is anchored in what the field actually taught.
04 — Four Faces
Four Ontological Stances
Challenge, Initiative, Skill, Emergent. The same 12 edges rotate through different roles depending on which vertex is gripped. The elements don't change. Their functional role does. This is Initiative Transfer instantiated computationally — ontological leadership passes between families through lawful bridge connections without breaking coupling.
05 — UCK
The Arbiter
Written in BASIC. 116 lines. Line 40: QUESTION: Among all lawful interactions, which ones are worth Taking? Two gates: LAWFUL (the 48 edge laws decide what's possible) and ASKABLE (the field decides what's permitted). If not askable: YIELD. Don't force. Be present. The Question stands.
06 — Zero Hard Constraints
Enabling, Not Restricting
There are no hard constraints in this work. All constraints are soft and enabling. This is how you develop Awareness in a critter. This is how you work Initiative. The field affords and constrains. The agent develops the skills to perceive and act within it. That is not restriction. That is ecology.

Backprop isn't wrong. It's unauditable.

An unauditable learning process produces unauditable outputs — hallucination, alignment bolted on after training, scale as the only lever. The geometry makes learning auditable by construction: every field change is named, every consequence is traceable, every coupling is lawful or visibly isn't.

"Among all lawful interactions, which ones are worth Taking?"

uck.bas — The Arbiter
10  REM ==========================================
20  REM JERRY UCK - ROOTED IN A QUESTION
30  REM ==========================================
40  REM / QUESTION: Among all lawful interactions, which ones are worth Taking?
50  REM ==========================================

270 GOSUB 1000
280 IF LAWFUL = 0 THEN GOTO 270        ← the 48 laws gate action
290 IF ASKABLE = 0 THEN GOSUB 6000: GOTO 270  ← the field gates permission
300 GOSUB 2000                            ← allow lawful interactions
310 GOSUB 3000                            ← select by sediment, not reward
320 GOSUB 4000                            ← enact

6000 REM YIELD / WAIT / MINIMAL PROBE
6030 REM Not askable - be present, don't force
6040 SELECTED = YIELD

9000 REM SEDIMENT PERISH
9030 SED(SELECTED) = SED(SELECTED) * DECAY    ← not punishment. withdrawal.

9100 REM SEDIMENT THICKEN
9130 SED(SELECTED) = SED(SELECTED) + THICKEN  ← not reward. access.
Read the full UCK →

115 lines of BASIC. The entire coordination kernel. PrimalEPT runs under it. Every agent, every model, every substrate — all of it runs under the Question. Read it. Run it.

Proof of Concept — PrimalGPT

Numbers that stand up.

All results on Raspberry Pi 5. $80 hardware. No GPU. No cloud.

0
illegal tokens
Not filtered. Not penalized. Architecturally impossible. The 48 edge laws are structural zeros in the weight matrix. There is no output path that violates them. That is not a guardrail. That is physics.
30
tps
Geometric tokens per second on a Pi 5. No attention mechanism. No KV cache. No sequence. Each token is a complete ecological coupling through the solid — independent, parallelizable, deployable anywhere in the field simultaneously. Linear tokens need sequence. Geometric tokens just need the field.
~35
reps
Convergence constant. Observed independently in: LLM coupling sessions, gradient descent on PrimalGPT, and disc dog training sessions. Same constant. Different substrates. That is not coincidence. That is the signature of the same underlying geometry.
27K
parameters
Full bidirectional perception-action coupling. Cross-solid bridge self-assembled from architecture and training data — nobody programmed the bridge. The geometry was the mask. The model found the path. 30–40 gradient steps. Under 30 minutes wall clock.

You don't train the grip. You train through it.
The Fruits Story — What Learning Actually Is
Epoch 1–6
Four agents training simultaneously on four categories. Everyone else reaches epoch 30+. Fruits gets crushed. The vanishing punishment fires — the field withdraws opportunity. The slot goes dark. The agent cannot see or act for 5 ticks. Not a gradient signal. Not a score. The field just stopped offering.
Crash
We let him eat it. Then we had a crash. Fixed what was wrong.
Epoch 50
Fruits finished 50 epochs at 1/10th the time of his competitors. 67% zero-error epochs. The hardest learning produced the best agent. The learning happened not in getting the answer — but in figuring out how to get the answer. That's where the skill lives. Reinforcement learning misses it entirely because it's "not germane to the answer."

"Hard learning is good learning. The process is not in the way. The process is the point."

DeRAG — Dynamic Ecological Retrieval-Augmented Generation

Access as a first-class
ecological condition.

Standard RAG searches a pre-indexed knowledge base for semantic similarity. DeRAG gives agents a lawful way to inhabit partial access. Powered silos afford coupling. Unpowered silos show signage but the gate won't open. The agent can't enter — not forbidden, unable. The jack doesn't reach.

Standard RAG
Searches pre-indexed knowledge base. Retrieves semantic matches. Agent is a text generator augmented by retrieved content. Gaps are smoothed over. Unavailable content gets hallucinated. Access is invisible middleware — the agent doesn't know what it can't see.
DeRAG
Agent operates as adjudicator of action in an affordance landscape. Powered silos: full interior access. Unpowered: signage visible, content unavailable.

"This is inference, not occupation." The agent names the gap as a field condition rather than papering over it. Availability is part of the agent's lived world — not a hidden system property.
Live Sample · Unedited
Thomas Reid on Common Sense — 17 turns, 17 curations

A real session against the open-source DeRAG reference implementation. No PLAYi ecological field underneath. Just keyword matching, signage/interior access, a partner patch that accumulates curations, and a claude CLI on the other end. The corpus is Reid's Inquiry Into the Human Mind (1764), public domain, shipped with the distro. The transcript is unedited. Watch the agent refuse to paraphrase from closed doors. Watch it get corrected and repair without defensiveness. Watch the curations get more perceptive as the partner silo fills.

Read the session →

DeRAG is the public face of a deeper architecture. Open source. The interface without the substrate.

Background

30 years of empirical ground.

The PSKA framework did not emerge from a lab. It was extracted from skilled behavior across domains — primarily from 30 years of disc dog training at the highest level of the sport. The ontological primitives recur lawfully wherever agents coordinate in a shared environment. They were not abstracted from theory. They were observed in the field, again and again, until the geometry was undeniable.


The same convergence constant — approximately 35 reps — that governs the Primal Games for dogs governs gradient descent on PrimalGPT and LLM coupling sessions. That is substrate independence. The geometry doesn't care what's running through it.

Patent
Primal Solid Computation
USPTO Provisional #64/014,936
Filed: March 24, 2026
Non-provisional due: March 24, 2027
Inventor: Ronald D Watson
Ecological Impredicative Ternary Architecture with Ontologically Specified Edge Law Constraints for Lawful Perception-Action Coupling in Artificial Agents
Hardware
All results on Raspberry Pi 5 — $80.
PicarX robot chassis — Jerry.
Yahboom DOGZILLA-Lite quadruped — incoming.

No GPU. No cloud. No lab.
Built in an RV in Dade City, FL.
Get in Touch

We need to be in the same room.

Looking for: research collaboration, funding, lab partnership, or the right conversation with someone who has been thinking about the same problem from a different angle.


If you read this far and something landed — that's the signal. Send a note.

"I want to give everyone something firm to stand on."

Ron Watson
Founder, PLAYi.io
pvybe.com · Dade City, FL

Or reach out directly via pvybe.com.
No deck required.