Measuring a scaffold from inside: an operator-authored self-measurement harness for a constitutional runtime over a hosted language model

Interactive intelligent systems constructed over hosted large language models cannot retrain the substrate on which they depend, and so their designers enforce a named epistemology at the surrounding scaffold layer. A difficulty follows that has received little methodological attention: when quality regresses, the regression frequently belongs to the scaffold rather than the substrate, and identifying it requires instruments that measure the scaffold itself rather than the model beneath it. This article reports one such instrument set, added to a live constitutional runtime between 17 and 18 July 2026 and extended with a production drift observer on 19 August 2026, and situates it as a reusable design pattern. The layer comprises six components-operator-authored adversarial batteries held as versioned regression fixtures, deterministic pattern screens encoding named defects, an emission quarantine that retains failure evidence as a first-class object, an instrument-health monitor issuing autonomous handoffs, continuous-integration delivery tripwires, and an observe-only production drift detector that accumulates a hand-labelling corpus from live traffic without altering any replies. Eight falsifiable harness predictions are stated (H1 through H8), each named against a specific runtime carrier-a code path, a database table, or a typed field-and each paired with an explicit falsification condition; their relationship to the eight pre-registered outcome predictions (F1 through F8) of a companion architecture paper is stated in a dedicated reconciliation section, since the harness reported here is the instrument set on which those outcome predictions depend. Of the eight, five (H1, H2, H3, H5, H7) carry dated on-platform observations and are reported in a dedicated observed subsection; three (H4, H6, H8) are prospective-predictions about experiments that have not yet occurred-and are reported separately under that label, so that no reader mistakes the distribution of evidence. The predictions are offered for adversarial replication under an honest empirical frame rather than as settled results: the batteries were authored by the platform's operator, classification is not independent, and the number of platforms studied is one. The discussion states a narrow priority claim, distinguishes the enforcement point of composition-time assertion from adjacent techniques, extends the pattern to the governance of learning loops-a scaffold that adapts must be measured by the same discipline that measures a scaffold that enforces-and argues that the interaction properties a reader encounters, an observation lock and per-turn provenance grounding, become testable precisely because they are structurally enforced.