Georgia Kali · 16 November 2026 · AI SDLC series, IntDEx framework
Why the model underneath an AI delivery engine should be treated as a replaceable, tested component, and how the IntDEx framework keeps public-service behaviour stable when models change
In conventional software, a dependency changes only when the team upgrades it. Hosted language models do not work that way. Providers release new versions, retire old ones and adjust behaviour behind the same name, often on their own schedule. A prompt that worked well last quarter may quietly produce different results today.
This is documented, not anecdotal. Chen, Zaharia and Zou (2023) compared the March and June 2023 versions of two widely used model services on the same tasks. One model's accuracy at identifying prime numbers fell from 84% to 51%, largely because it stopped following the chain-of-thought instruction it had followed months earlier. Their conclusion was that the behaviour of the "same" service can change substantially in a short time, and that continuous monitoring is needed. Models also differ from each other in ways that prompts expose. Sclar et al. (2024) found accuracy differences of up to 76 points from meaning-preserving changes in prompt formatting, and showed that a format that suits one model only weakly predicts what suits another.
For public organisations, the consequences go beyond quality. A service that depends on a single model from a single provider inherits that provider's release calendar, pricing and jurisdiction.
The IntDEx framework (https://intdex.org) lists model-adaptive delivery among its foundational characteristics. Models "evolve, drift, and differ in capability", so delivery must adapt continuously, and models are treated as interchangeable components whose fitness is evaluated rather than assumed.
In practice this rests on five commitments:
This aligns with the EU AI Act's expectation that high-risk systems perform consistently in accuracy and robustness throughout their lifecycle (Regulation (EU) 2024/1689, Article 15).
Problem. How can a public organisation benefit from rapid model improvement without letting unannounced model changes alter approved service behaviour, or locking itself into one provider?
How IntDEx addresses it. By evaluating models for fitness per intent, versioning routing and model choices as engine artefacts, re-running behaviour regression on every model change, monitoring drift, and periodically exercising a real switch to another model family.
Worked example (illustrative only). A hypothetical municipal service uses an AI delivery engine to draft plain-language summaries of planning decisions. Its engine manifest includes:
INTENT-12 v2 Summarise planning decisions at reading age 12
ROUTE-R1 v3 Default: hosted model family A; fallback: self-hosted family B
REGRESSION-S1 40 reference decisions with approved summaries
SOV-K1 v1 Processing location EU; provider dependency recorded
The provider announces that the current version of family A will be retired. Before switching, the Tester runs REGRESSION-S1 against the new version and against family B. The new version fails on legal terms that must be kept verbatim; the team adds a constraint, re-runs the suite, and releases a new behavioural bundle with a named human approving it. The family B run doubles as this year's portability exercise. All figures here are illustrative, not measured results.
Proposed evaluation criteria (not results). Regression pass rate before and after each model change; time from a provider's announcement to a re-tested, approved bundle; the number of intents that pass on at least two model families; and drift alerts per model version.
Limitations. Regression suites sample behaviour; they cannot prove equivalence between models. Testing several models costs time and money, and self-hosted alternatives may not match frontier capability for every task. Providers may change behaviour without announcing a new version, so drift monitoring can only detect change after the fact. The approach still needs independent case studies.
Conclusion. In AI-native delivery, the model is a supplier, not a constant. IntDEx treats it as a replaceable component under test: chosen per intent, versioned in the release bundle, monitored for drift, and proven portable through real exercises. A practical first step is to write down which model and version each service uses, build a small regression set for its most important intent, and run it once against a second model family.
AI was used for readability improvements.