Intent as Infrastructure: Why Meaning Must Become a First-Class System Boundary
Layer Thirteen — Position Paper
Status: Public
Audience: CTOs, Platform Architects, AI Governance, Corp Dev
Contact: contact@layerthirteen.com
Abstract
As software development becomes increasingly automated and AI-assisted, organizations are encountering a failure mode that existing tools cannot address: systems continue to operate correctly while no longer reliably embodying the intent they were created to serve. This paper argues that the root cause is structural, not technical. Meaning is not captured as a first-class, authoritative artifact in modern software systems. Instead, it is inferred repeatedly through documentation, code, tests, and human judgment. AI removes the human interpretive layer that previously masked this flaw, making semantic drift unavoidable and increasingly costly. We argue that intent must become explicit, revision-controlled, and authoritative infrastructure. Without such a layer, correctness cannot be preserved in non-deterministic implementation environments.
1. The Problem People Are Already Hitting
Teams adopting AI-assisted development report a familiar and growing set of symptoms:
- AI-generated code passes tests but behaves incorrectly
- Code looks reasonable but violates business or policy intent
- Small prompt changes produce materially different implementations
- Tests validate behavior but not meaning
- Reviews confirm plausibility, not correctness of intent
- Systems drift from original requirements over time while remaining operational
These failures are often attributed to hallucination, nondeterminism, or immaturity of AI models.
They are not.
They are manifestations of a long-standing semantic gap between intent and execution — a gap that humans previously bridged informally. AI exposes this gap by executing exactly what is specified, without compensating for what was merely implied.
As automation increases, semantic drift moves from a slow, hidden problem to an immediate and visible one.
2. Why Existing Fixes Do Not Work
When teams encounter semantic drift, they instinctively reach for familiar remedies:
- better prompts
- more examples
- larger test suites
- stricter reviews
- tighter processes
These approaches operate downstream of meaning.
Tests validate behavior against an interpretation of intent, not intent itself. When tests are generated from the same assumptions as the implementation — increasingly common in AI-assisted workflows — they reinforce those assumptions rather than challenge them. Passing tests increases confidence without verifying semantic correctness.
Reviews face the same limitation. Reviewers must reconstruct intent from incomplete artifacts. When intent is implicit or distributed, reviews devolve into assessments of plausibility, style, or alignment with convention. They confirm that the code is reasonable, not that it preserves meaning.
Prompt refinement assumes that intent can be fully articulated and stably interpreted. In practice, prompts are incomplete, context-dependent, and subject to nondeterministic interpretation. Refinement reduces variance but does not establish authority.
Automation amplifies these failures. When generation, testing, and iteration are all automated, systems begin validating their own interpretations of intent. Errors no longer present as failures; they present as stable, passing behavior.
This is not a tooling failure. It is a failure of semantic authority.
3. The Structural Root Cause: Intent Inference
Modern software systems do not have a place where intent lives authoritatively.
Instead, intent is fragmented across:
- tickets and design documents
- code comments and conventions
- institutional memory
- individual expertise
- prompts and examples
At the point of execution, implementors — human or machine — must reconstruct intent from incomplete signals. Humans are good at improvising under ambiguity; AI is not. It resolves ambiguity consistently, but not necessarily correctly.
As systems scale, cross language boundaries, or rely on AI-assisted generation, this reconstruction becomes increasingly lossy. Semantic drift accumulates not because anyone is careless, but because inference replaces authority.
Once inference becomes the mechanism by which meaning is preserved, drift is inevitable.
4. A Concrete Proof That the Gap Is Real
Consider a small but representative case: a policy function responsible for determining whether an action is authorized based on subject identity, scope, issuance time, revocation, expiration, and conflict rules.
The intent appears straightforward. Yet when multiple competent implementors — including large language models — are asked to implement it, a consistent pattern emerges:
- Implementations are coherent and defensible
- All provided tests pass
- Edge cases involving temporal boundaries or mixed invalidity are handled differently
- Implementors confidently justify divergent choices as reasonable interpretations
The failures are not random. They arise precisely where intent was not made explicit and local. In the absence of authoritative semantics, implementors infer.
When the same intent is later elevated into an explicit, revision-controlled semantic artifact — with declared inputs, outputs, invariants, error surfaces, and normative examples — the pattern changes. The same implementors converge deterministically. The resulting implementations are simpler, not more complex. Ambiguity disappears, not because implementors improved, but because inference was no longer permitted.
This outcome is repeatable. It demonstrates that semantic drift is real, that it survives testing and review, and that it cannot be eliminated without an authoritative intent layer.
5. Intent Must Become Infrastructure
The implication is unavoidable: intent cannot remain informal in systems that are automated or AI-assisted.
Intent must be:
- explicit and local
- human-reviewable
- revision-controlled
- authoritative over implementation
- capable of producing deterministic acceptance or failure
This does not require encoding entire systems in formal logic. It requires capturing the semantic obligations of each unit of behavior in a way that does not rely on inference.
Once intent is treated this way, implementation becomes subordinate. Code, tests, and AI-generated artifacts must conform to declared meaning or fail. Semantic drift is detected at the boundary, not after the fact.
This layer sits between ideation and execution. It is not a runtime, a framework, or a development methodology. It is infrastructure.
6. What This Layer Is Not
To remain enforceable, an authoritative semantic layer must be deliberately constrained. It must not:
- infer or repair missing semantics
- guess intent heuristically
- act as an autonomous code generator
- define project-level composition semantics
- execute or optimize implementations
- integrate external DSLs or business logic
- dictate organizational workflow or tooling
- serve as a policy engine or decision service at runtime
These exclusions are intentional. By refusing inference, heuristics, and execution concerns, the layer preserves determinism and enforceability across languages, teams, and implementors.
Its role is narrow but foundational: to act as a deterministic semantic choke point where intent is either declared or rejected.
7. Why This Becomes Mandatory in an AI World
Vibe coding and AI-assisted development collapse software construction down to a single human contribution: intent. When the human cannot reliably verify code, meaning must be verifiable by the system itself.
Without an authoritative intent layer, AI will continue to produce systems that are operationally correct, increasingly automated, and semantically unmoored. With such a layer, AI becomes constrained rather than creative, and correctness becomes enforceable rather than assumed.
This is not about controlling AI behavior. It is about preserving meaning.
8. The Inevitable Conclusion
AI has not created a new class of semantic failures. It has removed the human mechanisms that previously concealed them.
As automation accelerates, organizations will be forced to answer a question they have long avoided:
Where does intent live, and who has authority over it?
Any system that cannot answer this explicitly will drift. Any organization that cannot enforce intent as infrastructure will lose control of its systems over time.
The only viable response is to make intent a first-class, revision-controlled, authoritative system boundary.
9. An Existence Proof
The intent-as-infrastructure boundary described in this paper is not hypothetical. A minimal, language-agnostic reference system implementing these constraints exists today. It demonstrates that intent can be captured as a deterministic, revision-controlled artifact, and that implementations — human or AI-generated — can be accepted or rejected without inference, heuristics, or runtime coupling.
The purpose of this system is not to automate development, but to make semantic authority explicit and enforceable.
Readers interested in examining a concrete implementation of this approach can explore a public reference at:
https://sliver.layerthirteen.com
10. Questions This Paper Addresses
- Why does AI-generated code pass tests but still feel wrong?
- Why don’t tests catch semantic errors?
- Is AI making software correctness worse?
- How can intent be preserved in AI-assisted development?
- Why do systems drift from their original meaning over time?
- What is the difference between operational correctness and semantic correctness?
- How can organizations prevent intent drift without relying on human inference?
Closing Note
This paper identifies a structural gap that is already affecting real systems. It does not propose incremental fixes or tooling improvements. It names a missing category of infrastructure that becomes unavoidable as AI adoption increases.
Organizations encountering these failures are encouraged to examine concrete approaches and reach out for discussion and clarification.
Contact: contact@layerthirteen.com