Research
My doctoral research treats the points where learned components hand results to planners, monitors, and policies as architectural seams: specified, checked at runtime, and able to evolve.
A learned component hands a result across the seam to a symbolic component. The seam contract sits on that line and has four parts:
- Typed boundary objects. What may cross: the fields and the metadata that travel with them, such as confidence, timestamp, and the model that produced them.
- Versioned constraint bundle. The rules, policies, and validity conditions that govern how a result may be used.
- Runtime monitor. Checks each result against the active constraints and triggers the configured fallback when one is violated.
- Decision trace. Records the contract version, the constraints checked, the evidence, and the outcome.
The problem
AI-enabled self-adaptive systems increasingly combine two kinds of component. Learned components, such as perception, prediction, or language models, handle uncertain, high-dimensional inputs. Symbolic components, such as planners, monitors, and policies, provide explicit structure and decision logic that can be checked.
Consider an autonomous mobile system working in a shared space. A learned perception component detects an ambiguous object and reports object probabilities, a location, a velocity, and a timestamp. A planner or runtime monitor then decides whether the system should slow down, stop, reroute, or ask a person for help. What matters here is not only how accurate the perception result is, but also what the planner may legitimately assume when it uses that result.
In many architectures this hand-off is underspecified. Scores, labels, embeddings, or fragments of text cross the boundary without explicit assumptions about calibration, freshness, the meaning of labels, or policy compliance, so the component downstream has to interpret them implicitly. Failures can then arise at the interface between components rather than inside them, and a change to a model, a policy, or a constraint can call for broad revalidation.
The problem is architectural rather than only algorithmic. A better model does not settle what other components may assume about its outputs, and a runtime monitor does not, by itself, define the contract that governs the hand-off.
Seam contracts
I treat such hand-offs as first-class parts of the architecture: seams with contracts. A seam contract is a specification at the level of the connector between components, not of either component, and it has the four parts shown in the drawing above: typed boundary objects, a versioned constraint bundle, a runtime monitor, and a decision trace.
These parts can build on established techniques: schema validation for boundary objects, policy and rule engines for constraint bundles, runtime verification and drift or calibration checks for monitors, and provenance records for traces.
The contract is meant to be used across a system’s life. During design, it helps to elicit and document the assumptions and responsibilities at these hand-offs. At runtime, monitors check results against the active constraints, trigger the configured fallback when a check fails, and leave a trace. When the system changes, contract versions and dependency links identify what has to be revalidated: the affected contracts, monitors, tests, and components.
Localised revalidation is justified only when the contract-relevant assumptions are unchanged, or when changes are versioned and propagated through dependency links. A replacement model can keep the same output format while changing its calibration or error behaviour; such a change may still need offline or statistical evidence and broader validation.
Research questions
- Specification
How can requirements, policies, uncertainty assumptions, and domain assumptions be represented as modular contracts at learned–symbolic boundaries?
- Runtime assurance
How can such contracts be monitored at runtime and linked to fallback or mitigation behaviour to support controlled and auditable adaptation?
- Evolution
How can self-adaptive systems add, replace, or modify learned and symbolic components while localising the required retesting, conformance checking, and model or policy revalidation?
Where the work stands
The research proposal (July 2026) sets out the problem, the seam-contract concept, the design objectives, and the evaluation strategy. The next steps it names are refining the contract model, building a prototype toolchain, and applying both to two scenarios: a simulated perception-to-decision pipeline and a policy-governed assistant based on a language model.
The evaluation asks three things: whether the contracts can express the assumptions found in those scenarios, whether monitors detect violations and leave complete and faithful traces, and whether a change can be mapped to a justified set of things to revalidate. It compares seam contracts with ordinary interface schemas, wrappers, validators, and application-specific checks, and it is not designed to establish statistical generality.
Four design questions are open: where a seam is worth making explicit, how precisely assumptions can be stated without false precision, when localised revalidation is justified, and how to evaluate the approach credibly against realistic alternatives.
Papers
- Research proposal: Architecting Self-Adaptive Systems with Learned and Symbolic Components. Accepted at the ECSA 2026 Doctoral Symposium. It defines the seam contract, the research questions, and the evaluation plan.
- Position paper: Beyond Monolithic Models: Symbolic Seams for Composable Neuro-Symbolic Architectures, with Vincenzo Scotti and Raffaela Mirandola. arXiv preprint. It argues for composable AI systems in which learned and symbolic parts meet at deliberate seams instead of forming one monolithic model, with an enterprise policy assistant as its example.
Related studies
Two studies look at reasoning in software: whether language models preserve how strongly a stakeholder meant a requirement, and how an image explanation can become logical facts that formal reasoning can check. Alongside this, I have worked on which crash evidence helps a language model repair a program and on evaluating explanations of federated models.