From CI/CD to ACG: From Tests to Courts in Agentic Software Development

Continuous integration and continuous delivery transformed software engineering by making builds, tests, packaging, and deployment checks part of the infrastructure of change rather than optional developer practices. Agentic coding changes the limiting problem. As software agents become capable of generating and modifying implementation faster than humans can directly review it, the central engineering bottleneck shifts from code production toward epistemic and constitutional control: determining what a change means, whether an implementation corresponds to its specification, whether available evidence establishes the claimed behavior, whether previous evidence remains applicable to the exact current subject, whether a transition is authorized, what concrete effect should occur, and what may truthfully be claimed afterward. This paper introduces Ableman Constitutional Governance (ACG) as a software-development architecture in which a constitutionally significant subset of executable verification is promoted from tests into courts. Within the software domain considered here, Csoftware ⊆ Vexecutable, where ordinary unit, integration, regression, determinism, packaging, performance, and recurrence tests remain part of the broader verification estate. Courts differ not because they cease to be executable verifications, but because selected predicates are bound to an exact subject, constitutional frame, proposition or obligation, evidence topology, jurisdiction, authority relationship, and typed disposition. The paper develops a broader claim: the reasoning capability obtained from ACG does not arise from courts alone. It arises from combining constitutionally typed verification with constitutional synchronization, governed effects, and a richer representation of the system. Ableman's Razor, originally formulated as a typed grammar governing migration of claims among Specification, Configuration, and Implementation, is explicitly extended here as a representation lens for governed systems. The current integration resolves Configuration into Configuration Manifold 1 (CM1), the mandatory admissible manifold; Configuration Manifold 2 (CM2), the preferred admissible manifold under declared criteria; and IC, expanded interchangeably as Instance Constitution or Instance Configuration, the selected slice for a concrete system. IR, expanded interchangeably as Instance Realization or Implementation Realization, makes that selection concrete: P→ S→ CM_1→ CM_2→ IC→ IR→ E→ R. This extension is not attributed to the Razor's original jurisdiction. It is accepted only insofar as the added layers yield diagnostic, governance, selection, correspondence, evidence, or revision distinctions unavailable from a less differentiated representation without equivalent structure. This representation connects to Dynatology, the study of possible continuation: which transitions are conceivable, specification-valid, admissible, reachable, viable, selected, authorized, realized, observed, and established from a particular state. ACG supplies courts that adjudicate relationships and transitions among these projections. The architecture also extends governance into consequential execution through constitutional effect interposition. Application logic may continue to propose familiar semantic operations while a governed effect boundary determines the admissible realization under the current Instance Constitution. The requested operation, admitted realization, and actual observed effect are therefore distinct: O_r → R_a → E_a. A governed boundary may preserve, negate, constrain, transform, defer, simulate, sandbox, or substitute a proposed effect. This is not claimed to be novel merely as interception or mutation: reference monitors, capability systems, admission controllers, aspect-oriented programming, and effect handlers provide important precedents. The distinguishing concern is the integration of effect mediation with exact-subject binding, constitutional selection, evidentiary correspondence, typed dispositions, receipts, lineage, and standing. Because governed effects cross explicit semantic boundaries, those same boundaries become causal observation points. Receipts can record what operation was proposed, what realization was admitted, what executor action occurred, and what post-effect evidence followed. Governance, behavioral adaptation, and debugging can therefore share one effect topology. Documentation is similarly repositioned. Rather than remaining an independently authored and perpetually drifting artifact, documentation, glossaries, implementation codices, agent-orientation surfaces, capability references, operator explanations, and configuration views become deterministic, zero-authority projections of the retained semantic model. Projection fidelity does not itself establish realization fidelity; correspondence must be independently witnessed and adjudicated. The paper defines a reasoning-capability gain operationally as a representational extension that, while holding the underlying observed system state fixed, enables derivation of a decision-relevant distinction, contradiction, obligation, or admissible-action restriction that the baseline representation cannot express or derive without introducing equivalent additional structure. Several development episodes illustrate this pattern: environmental extraction safety conflicted with realization fidelity; artifact reuse was separated from effectful materialization; source changes invalidated stale verification evidence; and historical verifiers were made to testify to their own supersession rather than masquerade as current authority. The resulting software project is no longer modeled merely as source code plus tests and documentation. It becomes a constitutionally synchronized semantic and effect state containing specification, configuration possibilities, instance commitments, implementation bindings, evidence, courts, authority, effects, receipts, lineage, standing, and deterministic projections. ACG therefore suggests a progression from continuous integration toward continuous constitutionalization: not a replacement for CI/CD, but an enlargement in which verification, correspondence, adjudication, effect governance, and explanation become machine-addressable parts of the development process.