AI SDLC - Model-Adaptive

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

The problem: the model does not stay still

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.

What "model-adaptive" means in IntDEx

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:

  • Assess more than one model. During generative creation, developers test multiple models to find the best behavioural fit for each intent, and prompts may be optimised for specific model families rather than written once for all.
  • Route deliberately. The delivery engine architecture includes model routing, the mechanism that decides which model handles which task, adapting to performance, cost and latency. Routing is versioned with the rest of the engine. Research on learned routers suggests the cost case is real: Ong et al. (2024) report cost reductions of more than half on public benchmarks without a loss in response quality.
  • Treat a model change as a behaviour change. When a new model version arrives, prompts, constraints and architecture may need updating. Release packages record model versions alongside prompt, constraint and context versions, and rollback returns to the previous bundle as a whole, not just the previous code.
  • Watch for drift. Monitoring tracks degradation across model versions and over time. Testers identify drift that needs prompt or model updates; the Risk Manager owns the resulting risk log.
  • Prove portability, do not just declare it. The minimum adoption criteria require a tested ability to route a delivery task to more than one model family or provider, including a locally hosted or self-managed model where policy and infrastructure allow it. The exercise is repeated after material model or provider changes, and what is recorded is the exercised result, not the stated possibility. The Architect also records processing location, transfer basis and provider dependency for every model in use.

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).

Research-style validation

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.

References

IntDEx Framework

Other references

  • Chen, L., Zaharia, M., & Zou, J. (2023). How is ChatGPT's behavior changing over time? arXiv:2307.09009. https://arxiv.org/abs/2307.09009
  • European Parliament and Council (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Article 15. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  • Ong, I., Almahairi, A., Wu, V., Chiang, W.-L., Wu, T., Gonzalez, J. E., Kadous, M. W., & Stoica, I. (2024). RouteLLM: Learning to route LLMs with preference data. arXiv:2406.18665. https://arxiv.org/abs/2406.18665
  • Sclar, M., Choi, Y., Tsvetkov, Y., & Suhr, A. (2024). Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. International Conference on Learning Representations (ICLR 2024). https://arxiv.org/abs/2310.11324

AI was used for readability improvements.