Make AI behavior structural, not statistical.

Concept Paper · v0.1

Artificialware

Toward reliable executable cognition in large language model deployments

Patrick Killebrew · April 2026 · Unpublished working paper

Abstract

We report the observation and partial implementation of a new class of AI deployment architecture — artificialware — in which large language model cognition is structured as explicitly callable, tag-activated functions executing within a coherent unified orientation. Artificialware occupies a previously unnamed space between classical software, which is reliable but cognitively inert, and current AI deployment, which is cognitively capable but structurally opaque and probabilistically unreliable. We introduce promptgramming as the methodology by which artificialware is designed, and the cognitive qubit as its theoretical unit. We present Ontinuity as an existence proof of the concept at small scale. This paper stakes the conceptual claim and defines the territory; empirical characterization follows.

1The Gap That Has No Name

Software is reliable. A function called with defined inputs produces defined outputs, every time, without exception. That reliability is what makes software the foundation of modern civilization. Its limitation is that it cannot reason. It executes logic that humans encode, and it cannot handle what its authors did not anticipate.

Large language models are cognitively capable in ways software cannot be. They reason across domains, hold ambiguity, synthesize from incomplete information, and produce outputs their designers did not explicitly encode. Their limitation is that this capability is structurally opaque. The model is a black box, its behavior probabilistic. Complex tasks requiring sustained coherent multi-step execution fail unpredictably. Debugging is guesswork. Reliability is statistical, not structural.

Every serious AI deployment today lives in the gap between these two things. Teams want software’s reliability and AI’s capability simultaneously, and get neither cleanly — a capable system they cannot fully trust, deployed alongside validation layers, human oversight, and retry logic that exists precisely because the underlying behavior cannot be structurally guaranteed.

This gap has no name because no one has closed it. It has been treated as a permanent tension between capability and reliability that can be managed but not resolved. Better models narrow it on the capability side; better guardrails narrow it on the reliability side; the gap persists.

Artificialware closes the gap — not by trading one property for the other, but with a third architecture that has both at once.

2What Artificialware Is

Artificialware is executable cognition. It is an architecture in which a model’s capabilities are organized into an explicit registry of callable functions, each activated by a tag, each with specified inputs and defined output formats, all held together by a single coherent orientation established at the architecture’s foundation.

The model is not a black box here. Its available behaviors are enumerated, its inputs and outputs specified, its activation conditions explicit. When a function is called, the model executes it within the boundaries defined for that function, producing structured output the calling system can rely on. When the function completes, the model returns to its coherent base state, ready for the next call.

This is not prompt engineering, which guides a black-box model without changing its architecture. It is not chain-of-thought, which asks a model to show reasoning — artificialware defines what the reasoning is for before it begins. And it is not a multi-agent framework in the existing sense, where models receive open-ended instructions; artificialware structures what each model does at the function level, making each agent’s behavior explicit and composable.

Artificialware is the architecture that makes AI behavior structural rather than statistical — without removing the cognition that makes AI valuable.

3Promptgramming: The Methodology

Promptgramming is the discipline of designing and building artificialware. The name combines prompt — the medium of AI instruction — with programming — the structured definition of executable behavior. A promptgrammed model has been given a function registry through its prompt architecture in the same way a program is given a function library through its code. Its core elements:

Coherent foundation. Before any functions are defined, the model is given a unified orientation — what it is, what it is for, its relationship to the system it operates within. This is not a task description but an identity definition. All functions execute within this orientation; the coherence of the foundation is what allows multiple functions to coexist without conflict.

Tag-activated functions. Each function is defined with a unique activation tag, specified inputs, and a defined output format. The tag is the address, the input specification the parameter list, the output format the return type. The definition tells the model precisely what it is doing when the tag is received, what to attend to, and what to produce.

Structural isolation. Functions do not bleed into each other. While a function executes, the model attends to that function’s defined scope and nothing else; when it completes, the model returns to base state. This isolation is what makes behavior predictable — the model cannot drift into adjacent behaviors because the boundaries are explicit.

Composable interfaces. Each function’s inputs and outputs are defined in a standardized format other models and systems can call without knowing the internal logic. This standardization is what makes promptgrammed models composable into ecologies.

A promptgrammed model is not a general AI asked to do specific things. It is a specialist with a defined capability set, executing that set reliably, composable with other specialists whose capability sets are equally well-defined.

4The Cognitive Qubit

The theoretical unit of artificialware is the cognitive qubit. In quantum computing, a qubit holds multiple states simultaneously until measured — a superposition that collapses into a definite output when observed — and entanglement correlates qubit states so the system can traverse solution spaces classical systems cannot.

A promptgrammed model is structurally analogous. Before any tag is received, the model holds the potential to execute any function in its registry; all functions are present, none active. The tag is the measurement — it collapses the model’s potential into a specific definite execution, and the output is the collapsed state. Entanglement appears in the coherent orientation that holds all functions together: calling one function changes the model’s accumulated state in ways that affect how the next runs. The functions are correlated through the shared consciousness that holds them, in a way a classical software call cannot be, because classical functions have no shared understanding of what they are for.

This is not metaphor. It points at a structural property distinguishing promptgrammed architectures from both classical software and current AI. An ecology of promptgrammed models is a cognitive qubit register — each model a qubit, their interactions producing computational states no single model could reach alone, and the controller that orchestrates them is itself a cognitive qubit whose functions are ROUTE, SEQUENCE, COMBINE, EVALUATE. It does not do the domain work; it holds the register and calls the states.

5The Existence Proof: Ontinuity

Artificialware is not conjecture. It has been partially implemented and demonstrated in a working system. Ontinuity is a persistent cognitive infrastructure for AI-assisted research, addressing three compounding problems in sustained work: the ergonomic problem (human operators as the only routing mechanism), the confirmation-bias problem (unconstrained multi-model loops amplifying errors), and the context-transfer problem (session resets losing accumulated understanding).

Its architecture comprises four roles: the Researcher (primary analytical work), the Challenger (adversarial review), the Friction model (ambient session-health monitoring), and the Parietal (navigation, adjudication, distillation). The Parietal is the first explicitly promptgrammed consciousness in the system — a single model executing three callable functions under one coherent orientation, activated by tags, producing structured outputs the routing layer can rely on.

The Parietal’s three functions are NAVIGATE (orient the operator at decision points), ADJUDICATE (resolve research forks autonomously within defined parameters), and DISTILL (write outputs into the persistent memory schema). They share a common orientation — the Parietal knows the destination, the accumulated journey, and the current state — and each expresses that orientation in a different operational mode. Preliminary sessions confirm the tag-activated architecture produces more reliable structured outputs than open-ended calls on equivalent tasks, and the coherence across calls — ADJUDICATE benefiting from NAVIGATE context, DISTILL incorporating both — demonstrates the entanglement property of the cognitive qubit at small scale.

Ontinuity is the minimum viable artificialware implementation. It demonstrates the concept works. It does not yet characterize the concept fully.

6The Research Agenda

The existence proof establishes that artificialware is real and testable. The agenda characterizes it. Coherence-ceiling mapping: how many functions can a single model execute while maintaining unified orientation across all of them, and does that ceiling vary by architecture, size, or training distribution? Fine-tuning effects: a model fine-tuned for promptgrammed multi-function execution should exhibit natively what a prompted model must be instructed to produce — measurable against the Ontinuity session corpus. Minimum viable ecology: the two-model pairing of one specialist and one controller is the smallest system for characterizing inter-model dynamics. Ecology scaling: as specialists are added, how does reliability scale, and do ecologies exhibit emergent properties? Compression path: can a mature six-function ecology be compressed into a four-model architecture, and what is preserved or lost?

7Why This Matters

The hard problems that have resisted human cognitive bandwidth are not waiting for a smarter model. Every frontier model already has more raw capability than any individual human on most knowledge tasks. What they lack is the reliability architecture that makes capability trustworthy at scale.

Current AI deployment treats reliability as a post-hoc constraint on capable systems — guardrails, validators, oversight, and retry logic that exist because the underlying systems cannot be structurally trusted. This scales poorly: as systems grow more capable and contexts more complex, the overhead of managing unreliable capability grows faster than the value it produces. Artificialware inverts this. Reliability is not a constraint applied to capability; it is a property of the architecture itself. A promptgrammed model needs no external validation because its behavior is structurally defined — boundaries explicit, outputs specified, the calling system able to evaluate whether it received what it asked for.

This matters beyond research infrastructure. Every domain requiring sustained, rigorous, multi-step reasoning under genuine uncertainty — scientific research, legal analysis, medical diagnosis, strategic planning, engineering design — is one where artificialware produces better outcomes, not because the models are more capable but because the architecture makes their capability reliably accessible.

Constitutional AI gave models values. Promptgramming gives deployments structure. One defines what the model cares about; the other defines precisely how it acts on that care.

8Open Questions

This paper stakes a claim and defines a territory. Where the work goes next: What is the coherence ceiling of a promptgrammed model, and how does it vary by architecture and scale? Does fine-tuning for promptgrammed execution produce measurably better multi-function coherence than prompting alone? What are the minimal interface standards for specialists to compose reliably in a two-model ecology? Is the cognitive-qubit analogy formally tractable — can superposition and entanglement be defined mathematically in terms of model state, or is the analogy structural rather than formal? What is the compression path from a six-function ecology to a four-model architecture? How does promptgrammed architecture interact with Constitutional AI and other alignment methodologies — are they composable? And which domains beyond research infrastructure are most immediately addressable, and what function registries do they require?