K-NOME Programming

At a Glance

K-NOME is AI-assisted development done right: the human provides domain expertise and selective pressure, the AI handles implementation, and every generation is tested against published scientific results. Not vibecoding — structured evolutionary cycles where the AI is a knowledgeable collaborator under human constraint.

Knowledge-Numeric Observed & Mentored Evolutionary Programming

A pre-thesis writeup naming and formalizing the operational methodology behind 🔧🦎 ecoPrimals.

What K-NOME Is

K-NOME is the name for what happens between a human expert and an AI when the human treats the AI as a knowledgeable but non-specialist collaborator and guides it through constrained evolutionary cycles using their personal domain expertise, pattern recognition, and the semantic structures humans evolved for transmitting knowledge.

It is not vibecoding. It predates the term. It is not prompt engineering. It is not spec-driven batch generation. It is mentored, observed, iterative, conversational construction — where the human’s expertise is the selective pressure and the AI’s numeric breadth is the mutation operator, and they work together in the Knowledge-Numeric space where those two capabilities intersect.

The tool is Cursor. No Claude Code, no multi-agent frameworks, no external orchestrators. The methodology is the tool.

The Acronym

K-N: Knowledge-Numeric

The intersection space where human domain knowledge and AI computational/numeric capability meet.

The human brings pattern recognition, domain expertise, taste, judgment, lived experience — the things that cannot be compressed into a prompt. A microbiologist brings five years of watching microbial populations adapt on plates. A surgeon brings ten thousand hours of tissue response. A woodworker brings material intuition about grain direction. The knowledge is embodied, experiential, semantic — transmitted through the patterns and structures humans evolved for teaching: analogy, narrative, correction, “it should feel like this.”

The AI brings numeric breadth — the compressed inheritance of everything humans have written (see atlasHugged/10_THE_LOVE_LETTER.md), navigable at the speed of silicon, available as a generalist collaborator across every domain simultaneously. The AI is not a specialist. It is an intelligent generalist — knowledgeable broadly, deep nowhere, capable of producing candidate solutions across an enormous solution space.

The K-N space is the dimension where these two capabilities overlap. It is the space where the human’s deep, narrow, embodied expertise meets the AI’s broad, shallow, numeric competence. Neither is sufficient alone. The human without the AI is slow — limited by typing speed, by the time it takes to implement what they already understand. The AI without the human is directionless — capable of generating infinite candidates but unable to evaluate fitness in any domain-specific way.

K-N is the productive overlap. It is the space where mentoring happens.

O: Observed

Observation operates in two directions simultaneously:

The AI observes the project as a whole. In 🔧🦎 ecoPrimals, primals and springs reinforce each other. Patterns that work in 🐿️🧠 Squirrel’s capability discovery inform how BarraCuda structures its shader pipeline. The niche self-knowledge pattern from ⛰️♨️ groundSpring propagates to 💧♨️ wetSpring, then to 🌬️♨️ airSpring, then back to 🐿️🧠 Squirrel. Each new component changes the fitness landscape for every other component. The growing codebase IS the evolving environment, and the AI — with its full-project context window — observes the project at a scale no human can hold in working memory.

The human observes the process while building. Over 69,000 iterations across 10 months, the developer builds a felt sense for the project — an intuition for complexity lines, for where the architecture wants to go, for which areas are robust and which are fragile. This is the craftsperson’s observation: the woodworker who knows from the grain which way the wood wants to split. The potter who feels the clay’s water content through their hands. The runner who reads their body’s fatigue signature without checking a heart rate monitor.

This observation is not passive monitoring. It is the bidirectional feedback loop where the project teaches the human and the human teaches the AI and the AI’s output reshapes the project that teaches the human. Each full cycle — human observes, human mentors AI, AI generates, compiler selects, human observes the result — adds to both the project’s complexity and the human’s intuition about that complexity.

Observation is what separates K-NOME from vibecoding. Vibecoding is unobserved generation — the human prompts, the AI generates, the human accepts or rejects without developing deep understanding of what was produced. K-NOME requires that the human develops an increasingly detailed mental model of the system as it grows, and that this mental model feeds back into the mentoring.

M: Mentored

The human mentors the AI the way you would mentor a knowledgeable but non-specialist colleague in your domain.

Not commanding (“implement X”). Not batch-specifying (“here is a complete spec, generate it”). Mentoring — which is conversational, iterative, corrective, and uses the full range of human communication patterns:

• Analogy: “This capability discovery pattern should work like how organisms broadcast quorum signals — not a central registry, but each service announcing what it can do.”
• Correction: “No, that’s not what I mean. The provider shouldn’t know about the consumer. Think of it as a bulletin board, not a phone call.”
• Narrative: “The gen1 version was a job scheduler. It evolved into an orchestrator when we needed multi-provider routing. Now it needs to become a coordination primal — the thing that the rest of the ecosystem discovers AI capabilities through.”
• Taste: “That error handling is technically correct but it doesn’t feel right. The error variants should be domain-specific, not generic. A context error and a transport error are different kinds of failure.”
• Redirection: “Stop. You’re solving the wrong problem. The question isn’t how to call OpenAI — it’s how to discover any AI provider at runtime without knowing it exists in advance.”

These are the patterns and semantic structures humans evolved for transmitting expertise. They work on the AI for the same reason they work on humans: they provide selective pressure at multiple levels of abstraction simultaneously. A good analogy constrains the solution space more effectively than a hundred lines of specification, because it activates the AI’s compressed knowledge of the analogous domain.

The mentoring is domain-agnostic. An artist mentoring an AI on composition would use the same patterns — analogy (“this should feel like Rothko, not Pollock”), correction (“the weight is wrong, pull it left”), narrative (“I started this series trying to express X, but it became about Y”), taste (“that color is technically complementary but it doesn’t sing”). A surgeon, a woodworker, a musician — anyone with deep domain expertise can mentor an AI in their domain using the same human communication patterns.

K-NOME is not a programming methodology. It is a human-expertise- transfer methodology that produces software in this instance.

E: Evolutionary

The constrained evolution framework itself. Described formally in CONSTRAINED_EVOLUTION_FORMAL.md and informally in atlasHugged/04_THE_HUMAN_SEARCH.md.

• AI as mutation operator (LLM token sampling generates candidate solutions)
• Rust’s type system as environmental constraint (the compiler rejects unfit variants)
• Test suites as fitness function (do the results reproduce?)
• Iterative generate-compile-test-select cycles (69,000 iterations)

The E is what makes K-NOME different from simple AI-assisted development. Ordinary AI-assisted development is: human specifies, AI generates, human accepts or modifies. K-NOME is: human mentors, AI generates candidates, compiler selects against constraint, human observes the result, human adjusts mentoring, repeat. The process is evolutionary — it improves through selection under constraint, not through increasingly precise specification.

The constraint is the design. The mentoring is the selective pressure. The observation is the feedback mechanism. The K-N space is where all of it happens.

How K-NOME Maps to the Thesis

Darwinian and Lamarckian: Natural vs. Applied Evolution

The constrained evolution thesis (CONSTRAINED_EVOLUTION_FORMAL.md) is grounded in Darwinian evolution — and correctly so. Darwinian evolution is natural reality. Random mutation, natural selection, no inheritance of acquired characteristics. The Weismann barrier separates soma (body) from germline (DNA): what an organism learns in its lifetime does not rewrite its genes. This is how biology works. Taq polymerase, Lenski’s LTEE, Anderson’s boundary — these are Darwinian. The formal framework holds.

But K-NOME is not purely Darwinian. K-NOME is Lamarckian — and correctly so, because Lamarckian evolution is simply applied evolution: what happens when a conscious agent intervenes in the evolutionary process.

Jean-Baptiste Lamarck proposed that organisms inherit characteristics acquired during their parents’ lifetimes — the giraffe that stretches its neck passes a longer neck to its offspring. This was disproven in biology because the Weismann barrier is real. Acquired somatic changes do not reach the germline.

But there is no Weismann barrier in software.

In K-NOME, both dynamics operate simultaneously:

• The compiler is Darwinian. It does not care what you intended. It rejects what does not fit the type system. Blind, mechanical, indifferent. This is natural selection — the hot spring that kills everything except what is thermostable.

• The human is Lamarckian. The human acquires characteristics during the process — expertise, intuition, architectural vision, felt sense for complexity lines — and transmits those acquired characteristics directly into the next generation of code through mentoring. The human’s observation (O) feeds back into mentoring (M), which reshapes what the AI generates. The goals evolve. The fitness function evolves. The constraint environment evolves. Because the human evolves.

This is why Lamarck was wrong about biology but right about applied systems. In biology, there is a barrier between what you learn and what you pass on. In K-NOME, the human IS the mechanism that breaks the barrier. Every insight the human acquires through observation becomes heritable through mentoring. Every goal change, every architectural redirection, every “stop, wrong problem” — these are acquired characteristics being transmitted to the next generation.

Darwinian evolution is natural reality. Lamarckian evolution is applied reality. K-NOME is applied evolution — conscious, directed, mentored — running on a Darwinian substrate (the compiler, the test suite) that provides the blind selection the human cannot.

The formal thesis grounds 🔧🦎 ecoPrimals in Darwinian dynamics because the biological evidence is Darwinian. K-NOME grounds the operational methodology in Lamarckian dynamics because the human intervention is Lamarckian. Both are true. They operate at different layers of the same system.

What K-NOME Is Not

Not vibecoding. Vibecoding is unmentored and unobserved — the human prompts loosely, the AI generates, the human accepts without deep engagement. K-NOME requires domain expertise, active mentoring, and continuous observation. The human’s understanding deepens as the project grows.

Not prompt engineering. Prompt engineering optimizes the input to maximize the quality of a single output. K-NOME optimizes the evolutionary trajectory over thousands of iterations. The individual prompt matters less than the cumulative selective pressure.

Not spec-driven generation. Spec-driven approaches (including Huntley’s Groundhog/specs method) write a complete specification and generate code from it. K-NOME is conversational and iterative — the specification evolves alongside the code, because the human’s understanding of what they’re building evolves through observation.

Not multi-agent orchestration. K-NOME uses one tool (Cursor) and one human-AI relationship. No @pm, @architect, @dev, @qa agent roles. The human IS all of those roles. The AI is the generalist collaborator. The methodology doesn’t require orchestration because the human’s pattern recognition provides the coordination.

Not domain-specific. Any human with deep expertise in any domain can apply K-NOME. A microbiologist produces sovereign computing infrastructure. A surgeon could produce surgical simulation software. A woodworker could produce CAD/CAM tooling. The K-N space is wherever human expertise meets AI numeric capability.

Relationship to External Seeds

K-NOME evolved from two external seeds, and everything beyond them is original:

1. Geoffrey Huntley’s /stdlib (early 2025): Treat Cursor as an autonomous agent. Program LLM outcomes. This became the insight that the AI should be treated as a collaborator with agency, not a code completion tool.

2. BMad Code’s agile workflow (Feb 2025): Structured iterative methodology for Cursor. Build-Manage-Ask-Do loops, auto rule generation. This became the structural discipline — the iterative loop that K-NOME runs inside.

From these two seeds: /stdlib’s insight about AI-as-agent + BMad’s structured iteration -> 🐿️🧠 Squirrel gen1 and gen2. Then the original contributions: the biological grounding (constrained evolution), the K-N space concept, the bidirectional observation, the mentoring-as- selective-pressure framework, and the domain-agnostic claim — these are K-NOME.

The Practical Receipt

• Tool: Cursor IDE (only)
• Invocations: 69,000+ agent invocations
• Tokens: 51 billion processed
• Streak: 185 consecutive days
• Period: ~10 months
• Result: 14 primals, 7 springs, 27,000+ tests
• Developer: One person (microbiologist BS, data scientist MS)

The Cursor receipt is the evidence that K-NOME, as a methodology, produces results at scale. The commit history is the evolutionary trajectory. The test counts are the fitness measurements. The architecture papers are the post-hoc analysis of what the methodology produced.

Pre-thesis writeup prepared March 16, 2026. Formal treatment to follow in thesis chapter.