RACE Programming
Layer 1 — Science

AI-First Theory

The scientific foundation. Two axioms about AI systems and human tacit knowledge; six theorems that follow from them. Together they describe what happens to the software development lifecycle when AI systems can read, write, and verify code at production grade.

Why a theory

Engineering needs a ground truth, not a framework

Most discussions of AI in software development are tactical: which tool, which prompt pattern, which IDE integration. The upstream question: what, formally, is changing?

AI-First Theory answers in terms suitable for reasoning, falsification, and extension. Not a methodology. Not a manifesto. A compact set of propositions — stated precisely enough to disagree with, find edge cases in, or extend.

Theory → Manifesto → RACE Programming. Each layer more prescriptive. Each depends on the one above being sound.

Foundations

Two axioms

Two claims taken as starting points. Asserted as premises; theorems follow from them.

Axiom I · Formalization
AI systems can now formalize and execute any piece of software engineering knowledge that can be made explicit.
Axiom II · Tacit transfer
A growing share of what was previously tacit engineering knowledge is becoming formalizable through AI-mediated dialogue.

Full discussion of the axioms and their grounding →

Consequences

Six theorems

From the two axioms, six theorems follow. Full proofs on individual pages; compressed statements below.

  1. Theorem of Displaceable Labor
    Any engineering task whose inputs, constraints, and acceptance criteria can be expressed in code-executable form can be delegated to an AI system with human supervision.
  2. Theorem of Specification Primacy
    In AI-native development, the specification — not the code — becomes the primary artifact of engineering skill. Code becomes a derived artifact.
  3. Theorem of Verification Inversion
    The ratio of code-writing to code-verification inverts. Engineering time is spent predominantly on verification and curation rather than generation.
  4. Theorem of Team Compression
    An AI-augmented team of N humans can produce the output of a traditional team of 5N–10N humans — provided roles, process, and tooling are aligned.
  5. Theorem of Context Fragility
    AI-executed development is bounded not by model capability but by the quality of the context made available. Stale or contradictory context is the primary failure mode.
  6. Theorem of Methodological Non-Neutrality
    Existing SDLC methodologies are not neutral substrates for AI augmentation. They contain structural assumptions incompatible with AI-native execution, requiring purpose-built methodology.

What this theory is not

Not a claim about AGI, consciousness, or the singularity. Narrow scope: current-generation AI systems (LLMs with tool use and code execution) × structure of software engineering work. Intended to remain valid across model generations — operates above any specific model.

Not a call to replace engineers. Every theorem implies a new engineering role — higher-leverage, specification-focused, verification-intensive. Describes what shifts; does not prescribe what disappears.

Reading order

Fast path: axioms → T2 (Specification Primacy) → T3 (Verification Inversion). Sufficient to understand why RACE Programming looks different from Scrum. Complete picture: read in order — later theorems reference earlier ones.