This preprint sets Vibe Coding and Spec-driven development against each other as two archetypes of AI-assisted practice and asks which of them can carry enterprise-grade software. Its method is design science: a designed artefact — the Specification Governance Reference Model — is constructed from a systematically verified evidence corpus and then evaluated by informed argument against an external quality standard, with no new experimental data collected. The corpus was assembled from systematic reviews of LLM-based software engineering, the citation graphs of foundational treatments of the two practices, and targeted searches for empirical studies of AI-assisted development quality; every candidate reference was verified against a primary source before inclusion, and claims without a verifiable source were excluded rather than attributed. The article reports the resulting corpus as 44 verified sources across controlled experiments, user studies, grey-literature reviews, benchmark studies, enterprise case studies, methodological treatments, standards and foundational works.
It is organised around three research questions: which failure modes of ungoverned conversational generation the verifiable evidence identifies with respect to enterprise quality attributes; through which formal mechanisms specification governance addresses those failure modes, and how those mechanisms can be consolidated into a tool-independent reference model; and to what extent the published evidence supports specification-driven development as the foundation of AI-native enterprise software engineering, with what boundary conditions. A failure mode was admitted only if supported by at least two independent verified sources, and counter-evidence was recorded rather than discarded.
Its stated position is deliberately conditional rather than absolutist. Vibe coding is affirmed as a legitimate and valuable instrument for ideation, prototyping, learning and accessibility, and specification governance is said to carry authoring and maintenance costs that simpler contexts do not repay. The claim defended is narrower: that where enterprise-grade quality attributes — testability, scalability, security, maintainability, traceability and regulatory compliance — are required, specification governance is the mechanism by which they are obtained from stochastic generators.
Key Points
- Four recurring failure modes of ungoverned generation are identified and each traced to a causal mechanism: a productivity–reliability paradox, architectural erosion from localized model context, security exposure, and technical-debt accumulation. “Ungoverned” is used precisely, attaching to practices where acceptance is decided by observational sampling of behaviour rather than verification against an explicit contract, not to AI assistance as such.
- The four failure modes are argued to share one causal structure: a stochastic generator optimizing localized context, coupled to an acceptance criterion — observed behaviour — that is strictly weaker than the quality attributes an enterprise requires.
- A specification is formalized as a machine-readable tuple of four components: functional obligations (interface signatures with preconditions, postconditions and invariants, acceptance scenarios and behavioural examples), quality constraints (measurable thresholds over ISO/IEC 25010 characteristics), constitutional constraints (non-negotiable security, privacy and regulatory rules), and architectural structure (module boundaries, bounded contexts, interface contracts and dependency rules).
- A generator is defined as a stochastic function from specification and context to an implementation artefact, and a validator as a deterministic, decidable procedure decomposed into static and constitutional analysis, specification-derived tests, contract and proof obligations, and architectural conformance. A process is specification-governed only if every integrated artefact lies in the acceptance set of the current specification and every intended behaviour change is initiated by a change to the specification.
- The article argues this vocabulary makes the contrast with vibe coding precise: there the specification is latent, existing only as implicit intent distributed across working memory and conversation history, so the acceptance set is undefined and acceptance degenerates to observational sampling — a strictly weaker criterion than membership in that set.
- Three rigor tiers taken from the practitioner literature — spec-first, spec-anchored and spec-as-source — are restated as progressively stronger process invariants over the specification–implementation pair. The article presents them as a costed design space rather than a ranking of virtue, with each step buying stronger guarantees at higher authoring and tooling cost.
- The reference model is organised into four layers — specification, generation, verification and governance — crossed by a traceability spine, with an operational core stated as a rejection-sampling loop in which validator diagnostics feed back as corrective context and budget exhaustion escalates to human governance rather than silently lowering the bar.
- Six falsifiable design propositions state the model’s empirical commitments, covering the determinism boundary, reproducibility across sessions and model versions, verification by construction, architectural integrity, maintenance economics, and traceability as a process byproduct.
- Every ISO/IEC 25010 product quality characteristic is mapped to a documented risk under ungoverned generation, the mechanism intended to address it, and the evidence bearing on that mechanism. The article observes that because the specification has a component for each family of obligations, each characteristic of the standard has a designated place to be demanded and a designated validator component to check it.
- The evidence is explicitly weighted by maturity into three tiers: replicated experimental findings (the productivity effect of AI assistance, the security risk of unreviewed generation), single rigorous studies awaiting replication (the correctness gap, practitioner-perceived unreliability, a null maintainability result), and unreplicated case evidence — under which the article places its most striking pro-specification quantifications, a reported 73% security-defect reduction under constitutional constraints and enterprise delivery multipliers from a single financial-institution case study. The author states that the argument does not rest its weight on these.
- A counter-result is recorded rather than discarded: one preregistered controlled experiment did not detect systematic code-level maintainability disadvantages when new developers later evolved code co-developed with AI assistants. The article’s reconciliation — that the moderating variable is governance rather than the presence of AI — is labelled a hypothesis consistent with both datasets rather than a directly tested finding.
- Boundary conditions are tabulated: conversational generation for ideation, prototyping, learning and MVP assembly; spec-first for standalone utilities and short-lived tools; spec-anchored for multi-developer products with sustained evolution; and spec-anchored to spec-as-source with constitutional constraints for enterprise, regulated or safety-critical systems.
Notes
The article positions specification-driven development as an adaptation of four mature traditions rather than an invention of the LLM era: program synthesis, model-driven engineering, design by contract, and requirements economics. It reads the practice as a synthesis of two of them in particular — the fluency of language models joined to the rigor of specification, on the argument that specification-driven techniques historically struggled with exactly the scalability and expressiveness that models supply.
Its threats-to-validity section names four limits: construct validity, internal validity, external validity and reliability of the corpus. The central constructs are young and contested, so the article anchors each to a published definition and formalizes specification governance so the evaluation cannot equivocate. The evaluation is analytical, and no causal claim in the article was tested by an experiment conducted for it; the sharpest internal threat named is the moderating-variable inference. The corpus over-represents certain settings — controlled experiments on bounded tasks, security benchmarks on scenario suites, one regulated financial enterprise and one banking-microservices case — so generalization is described as plausible but unestablished. On corpus reliability the article notes that several load-bearing sources are preprints under review, and names two disciplines that limit the exposure: verification of every reference against a primary source before inclusion, and the evidence-maturity weighting that keeps unreplicated results from carrying the argument. The research agenda names the isolation of the moderating variable as the single most valuable next experiment, alongside replication of the case results, the design of specification languages that are simultaneously human-authorable, machine-verifiable and model-legible, and measurement instruments for specification effort comparable to those the discipline has for code.