Same 60 changes, two architectures: complexity amount holds steady while prompts and an AI reviewer each move placement

With the task, model and change list held fixed, the two architectures end with nearly the same total complexity, and the difference sits in where that complexity is placed. Tesler's conservation law and Brooks' separation of essential from accidental complexity are widely quoted but rarely measured, because a setting is needed in which the problem is genuinely fixed while the solution genuinely varies. Ten cells cover two architectures, five conditions and ten independent runs each, and four operationalisations resting on different theories give amount spreads of 7% to 20%, with the complexity the changes added varying by 21% to 29% within an architecture.

External instructions move the mutative architecture's placement a great deal and the additive architecture's hardly at all, while an AI reviewer is the only lever that moves the additive arm, and only on the concentration family. The four external controls, a bare change request, a plain-language cohesion prompt, a quality gate and the structural cost function disclosed in the prompt, move the additive arm's placement metrics by only 3% to 29%. Across eleven placement metrics the plasticity ratio between architectures is 2.7 to 10.0, with ten of eleven intervals entirely above 1; adding the AI reviewer roughly triples the additive arm's spread on five concentration metrics and drops the ratio to 1.2 to 2.1, while coupling, file count and per-node complexity barely move.

The structural gain comes with a correctness cost: strict pass rates fell in most architecture-by-intervention comparisons while the isolated pass rate barely moved. The strict rate requires every test of every rule so far to be green and therefore accumulates earlier breakage, whereas the isolated rate requires only that this checkpoint's own rule passes its own tests. The isolated rate is between 0.983 and 1.000 in all ten cells and all 6,000 checkpoints compiled; strict pass fell in most comparisons, the one exception being the mutative arm under the gate rising from 0.267 to 0.308, and the AI reviewer did not buy strict correctness back.

The findings apply to one REST endpoint, sixty accumulating change requests, the fixed model claude-opus-4-8, and two concrete codebases, a mutative Spring controller and an additive OfficeFloor pipeline. A planned replication rewrites the default OfficeFloor guide and adds a third arm built on a framework the author neither created nor maintains.

The statistic was defined after the runs existed, so it is exploratory and the intervals quantify sampling error only; the AI reviewer is one review design, one model and ten chains; the additive arm's YAML wiring is excluded from every measurement; there is one codebase per arm, so standard deviations describe model stochasticity rather than variation between architectures; and the five conditions were not run concurrently, confounding condition with date.