A growth vector orthogonal to scaling.

Synthesis Paper · v0.1

A New Growth Vector for AI

Cognitive ecology, Ontinuity training, and the architecture beyond scaling

Patrick Killebrew · April 2026 · Synthesis paper

Abstract

The dominant growth vector in AI development is scaling: more parameters, more compute, more data. This vector has produced extraordinary progress but is exhibiting diminishing returns at increasing cost — the next order of magnitude of compute does not reliably produce the next order of magnitude of capability. This paper describes a completely orthogonal vector: the cognitive ecology. Rather than making individual models more capable in isolation, it binds multiple models with different cognitive profiles into structured collaborative architectures where each occupies the role where its tendencies are assets rather than liabilities. Four components are documented — the cognitive ecology (the framework), Ontinuity Training (the training paradigm it enables), the Teaching Leash (the safety architecture it produces), and the Psychology of AI Data (the empirical instrument) — constituting a research agenda that is orthogonal to scaling, accessible without data-center infrastructure, and grounded in a working proof of concept.

1The Stagnation Problem

The AI field is not stagnant in absolute terms — models improve, benchmarks rise, capabilities emerge. But the marginal return on investment in scaling is declining. The leap from one frontier generation to the next was once qualitative; the improvements since have been incremental. The reasons are structural: scaling has one primary lever — more — and that lever has been pulled hard for five years. The easy gains are behind us; the hard gains require infrastructure measured in tens of billions of dollars and produce improvements measured in percentage points.

Meanwhile, the fundamental architecture of how AI is deployed has barely changed. A human types a prompt, a single model responds, the human reads it and types another. This was appropriate when models were weak and every human turn added value. It is increasingly inappropriate as models become capable enough to do sustained work autonomously, and as the bottleneck shifts from model capability to human attention bandwidth. The field is optimizing the wrong variable. The question is not how capable a single model can become — it is how effectively multiple models with different cognitive profiles can work together, with humans at genuine decision points rather than routine cycles.

2The Cognitive Ecology

A cognitive ecology is a structured arrangement of models with different cognitive profiles, each occupying a role where its tendencies are assets rather than liabilities, coordinated by an orchestration layer. The foundational insight is that model tendencies are not universally good or bad — they are contextually valuable or limiting depending on the role. A model’s tendency toward comprehensive elaboration is a liability in a Researcher role where direct execution is required, and an asset in a Challenger role where thorough adversarial objection is exactly what is needed.

These are not hypothetical observations — they were documented in the first live four-model session. A model in the Researcher role exhibited a consistent tendency to describe and elaborate rather than execute, even under explicit redirection; the friction signal rose to override twice, correctly identifying the resulting drift. When the role assignment was swapped, the same elaboration tendency became an asset: thorough, multi-angle adversarial challenges that built progressively on prior concessions.

The intelligence is not in any single model. It is in the architecture that binds them. The cognitive ecology is not a metaphor — it is an operational system.

The full framework is developed in The Cognitive Ecology; Tetraform implements it at the session level, and Knowtext provides the memory layer that makes it persist across sessions.

3The Teaching Leash

The cognitive ecology produces more than better research output — it produces a safety architecture. The dominant approach to safety is alignment, a training-time intervention that bets correct values going in produce correct behavior coming out. It faces a fundamental challenge: the world is always larger than the training distribution. The Teaching Leash proposes a complement that operates at runtime — the adversarial layer providing continuous resistance to confidence overreach — and that does more than catch failures: it generates the training signal that makes them less likely over time, mapping the primary system’s failure geometry from real operation.

The leash teaches the dog. The dog and the leash improve each other. The frontier keeps moving. The loop keeps running.

4Ontinuity Training

The cognitive ecology is not only a deployment architecture; it is a training environment. Every session generates a structured correction record — adjudicated challenge events, retraction histories, process-integrity scores — documenting where the primary model’s reasoning failed under peer adversarial pressure from a model with a different training distribution, what the correct reasoning looked like, and how cleanly the diagnostic ran. This is training data of a quality and specificity synthetic pipelines cannot produce.

Ontinuity Training uses this correction record as its primary curriculum: the Researcher produces output, the Challenger disputes it with a genuinely different cognitive profile, the human adjudicates, and the Researcher is fine-tuned on the correction record — then the next session begins with both improved. The result is co-evolution: two systems with different cognitive profiles improving each other through genuine intellectual competition. This is not adversarial training in the GAN sense — the adversary is not a discriminator optimized for a scalar reward but a frontier model capable of locating exactly where a specific claim overreaches its premises, and the human is not a labeler but an adjudicator exercising genuine judgment. It produces something existing paradigms cannot: a model whose reasoning has been stress-tested by a different kind of intelligence and found defensible — not against a fixed rubric, but against genuine challenge from a system that thinks differently.

5The Psychology of AI Data

The cognitive ecology, the Teaching Leash, and Ontinuity Training all depend on empirical knowledge of how specific models behave under adversarial pressure — knowledge that does not currently exist in systematic form. The Psychology of AI Data framework is the instrument for generating it. It specifies four native output streams that sessions produce: the friction-signal sequence, the status-tag sequence, the adjudicated challenge-event record, and the retraction-event record — encoding not just what was produced but how the session moved. Its primary contribution is behavioral tendency mapping: accumulating adjudicated challenge records across sessions to reveal stable patterns in how specific models respond to adversarial pressure on specific claim types. That is the empirical foundation for ecology design (which model in which role), Teaching Leash calibration (which failure modes to watch), and training-curriculum design (which adversarial pairings generate the most useful signal). The framework was itself produced by the method it describes, with a substantial retraction record as its quality warrant.

6The Proof of Concept

This paper describes a research agenda, not a future possibility — every component has been built and run. The session protocol was validated through live inter-model sessions, where true inter-model divergence produced qualitatively sharper adversarial pressure than single-model dual-role simulation. The routing client ran its first fully automated four-model session on April 9, 2026: seven cycles with accurate friction scoring, two override events that correctly identified genuine drift, ALIGNMENT_NEEDED firing at a real decision point, and a clean termination.

The entire corpus — specifications, framework papers, concept papers, and a working session script — was produced in roughly forty-eight hours across two days, using the methodology it describes, on an iPad M4 and an HP laptop, with no institutional resources.

The architecture is self-documenting: it generated the framework while demonstrating the framework. The proof of concept is the corpus. (For the disciplined separation of what is measured from what is modeled in these results, see the Synthesis.)

7The Research Agenda & Why Now

Each component opens a research agenda the field has not begun to address: optimal role assignments for current frontier models across task types; whether adversarial runtime resistance produces measurably better safety outcomes in physical systems; whether Ontinuity Training produces measurable improvement on hard reasoning benchmarks versus existing paradigms; and whether behavioral tendency mapping can predict session quality in advance. These questions are answerable with current infrastructure.

The timing is not accidental — three conditions converged. First, frontier models from different companies became capable enough that their inter-model differences are assets rather than noise. Second, API access to multiple frontier models became affordable for individuals: a session’s cost is measured in dollars, not hundreds. Third, consumer hardware capable of running local models for ambient monitoring reached threshold — an iPad M4 with 16GB unified memory can run a local model server fast enough for real-time friction scoring. The local model is on the device, the frontier models in the cloud, the orchestration on the laptop: the cognitive ecology runs on hardware that fits in a backpack. The window for establishing this research direction is open; this paper is the stake in the ground.