Building is only useful when the boundaries and trade-offs are sound. These records make the reasoning behind architecture, integration, data, and implementation choices visible.
13 decisions documented
Historical reports needed to remain intuitive to find and use. Modeling reports and their lifecycle in the ontology created a stronger connection between current workflow actions and persistent historical records.
Reviewers needed to select relevant subordinate inputs and build the next report without dropping back into manual consolidation. Explicit report types and actions made that transition part of the application rather than an informal handoff.
Keeping source-specific mapping at the edge lets the bridge operate on a stable contract instead of accumulating vendor-specific branches. That makes additional sources easier to add and keeps canonical translation testable in isolation.
The transport needed to decouple producers and consumers while supporting a distributed integration model. Zenoh provided a way to move normalized records without making an upstream vendor platform the backbone of the architecture.
A canonical model gives downstream components a stable target even when upstream sources use different schemas and terminology. The proof of concept therefore translates normalized transport records into ISA constructs rather than propagating source-specific structures through the system.
A proof of concept is more useful when another technical team can reproduce the path and see where failures occur. Initialization, proof scripts, diagnostics, and repeatable packaging were treated as part of the implementation rather than after-the-fact documentation.
Late fusion made it possible to compare baseline-only and augmented predictions directly, isolate the effect of external knowledge, and run controlled ablations without retraining the full vision-language model.
ConceptNet includes large amounts of generic or weakly relevant information. Explicit limits on entities, hops, relations, neighbours, edge weights, and top-K facts made the retrieval process testable and reduced uncontrolled noise.
A control using unrelated knowledge was necessary to test whether any apparent improvement could be attributed to task-specific evidence rather than the existence of an auxiliary branch, trainable fusion parameters, or reranking effects.
The research question was whether the technique worked under matched conditions. Preserving configuration-linked results and failure analysis was more defensible than tuning toward an unsupported positive claim.
Interviews showed that registrar decisions are important but do not explain or own every blocker. All arrivals initially pass through a holding/in-processing function, while later hold populations can exist elsewhere. The system therefore needs to represent transitions and ownership across organizations.
An attention indicator is useful only if it leads users to the population they can actually work. The prototype therefore connects exceptions to the relevant roster/work view rather than creating separate summary screens with no resolution path.
The prototype can improve visibility, ownership, and coordination without pretending to replace the underlying systems of record or writing back before the workflow model is validated.