A migration knowledge graph represents systems, capabilities, dependencies, candidate changes, and their evidence as connected assertions. To explain a proposed path, it must distinguish observed conditions from possible treatments and demonstrate which prerequisites and limits connect them.
A graph can make the reasoning inspectable without making it correct automatically. Its value depends on relationship meaning, evidence scope, and the treatment of missing information, not simply on the number of connected nodes.
Keep assessment dimensions separate
Application role, implementation platform, data format, deployment, criticality, support, and modernization objective describe different properties. A platform label should not silently determine a treatment, and age should not substitute for evidence of operational risk.
SKOS allows concepts to participate in multiple concept schemes. That supports flexible classification, but does not prove that every proposed business dimension is independent in practice.
The model should preserve aliases, disputed labels, and unknown values. A query can then select relevant candidates without converting an absent classification into an invented fact.
Give each relationship a precise meaning
An application can implement a capability, use a database, run on a host, or be a candidate for a transformation. These relations have different directions and evidentiary requirements.
Observed dependency and proposed applicability should remain distinct. A tool that offers a conversion capability has not thereby demonstrated that it converted this application successfully.
Technical modeling details can change inference. An RDF Schema domain declaration entails class membership for subjects using the property; it is not an ordinary input-validation restriction. Multiple domains imply membership in all of them, which can produce unintended classifications if modeled carelessly.
Approximation is not equivalence
SKOS closeMatch is not declared transitive, limiting the accumulation of approximate mapping errors. Two approximate matches cannot be treated as a verified identity merely because they form a traversable path.
The same discipline matters for migration claims. Similar platforms, formats, or application archetypes can suggest a candidate without establishing compatible data semantics, runtime behavior, or recovery needs.
A route query should return the specific evidence supporting applicability and the checks still required. Unsupported compatibility must remain visible rather than being inferred from lexical similarity.
Represent plans and executions separately
PROV-O defines a plan as an entity describing intended actions toward goals. A candidate migration path can connect source conditions, target state, strategy, mechanisms, and delivery responsibilities in that sense.
Actual execution needs another record with dates, observed outcomes, deviations, and failed or passed gates. A published plan should never count as a successful case simply because it names a source and target.
Prerequisites, contraindications, validation, and rollback belong on the candidate route. Evidence for a successful transition elsewhere should retain its environment and scope rather than certifying the new installation.
Queries must expose incomplete knowledge
SPARQL optional matching can retain a result when optional information is absent. An ASK query tests whether a pattern has a solution in the queried data; false does not establish absence in the world.
A decision query should therefore distinguish no recorded contraindication from verified suitability. It should return source dates, unresolved fields, conflicts, and superseded assertions where those affect the proposed route.
Test the model on a small curated graph containing missing and conflicting evidence as well as accepted examples. Acceptance means a reviewer can trace a candidate to its actual supporting claims and understand what would need to be verified before implementation. The graph explains a proposal while leaving feasibility and operational success as properties requiring their own evidence.