Arqua architectural interpretation of selected public documentation, reviewed 24 September 2026. Not a benchmark, certification or deployment recommendation.
A service-review question may need several different kinds of information: which assets support the service, what was observed, which procedure applies and what an authorised person actually decided. A convincing answer needs those accounts to fit together. More accessible information does not, by itself, establish that fit.
Selected Microsoft capabilities can contribute relationship representations, event histories, documentary retrieval, workplace context and evaluation evidence. This explanation connects those five responsibilities without proposing a mandatory five-product stack or assessing the entire Microsoft portfolio. Arqua supplies the architectural questions around meaning, participation, reliance and consequence; the implementation may combine Microsoft services with existing enterprise capabilities.
Keep definitions, bindings and relationship results distinct
Ontology in Microsoft Fabric, labelled Preview in the reviewed documentation, represents entity types, properties, relationships and their data bindings. Its graph can provide instantiated relationship information for a particular question. The enterprise's accepted definition, the binding to a source, the populated graph and the returned query result remain different things.
That distinction matters when a service changes. The documented graph refresh behaviour separates ontology-schema changes from upstream record changes, which need manual or scheduled refresh. A recent query can therefore describe an earlier source state. Record both the relationship's applicable period and the basis actually loaded into the representation.
The binding requirements also restrict eligible sources, including the documented exclusion of lakehouses with OneLake security enabled. Do not remove a required protection merely to make an integration work. Select a supported, authorised pattern or keep that use unresolved. These are conditions of the selected path, not a claim that all Fabric services have identical security limitations.
Preserve event meaning as well as time
Eventstreams can ingest, transform and route events, including to Eventhouse, and documents at-least-once delivery. Consumers need an explicit interpretation of duplicates, corrections and the subject each observation describes. Transport delivery does not establish that every relevant business occurrence was captured.
Keep source observation time, arrival, processing and the time a result was consumed distinguishable. The last message to arrive need not describe the latest business condition. Combining a fresh observation with an outdated service relationship can still produce an inapplicable conclusion.
Kusto's materialized-view documentation provides an important positive distinction: a full-view query combines stored results with the source-table delta. It need not wait for background materialisation to include newly ingested data. A materialized-only query is a different path, and neither path supplies observations that never reached the source table.
Failure rules also belong to the evidence contract. Microsoft directs consumers to ignore the entire result when a Kusto partial query failure occurs. Useful partial understanding can come from independently successful sources or a new successful narrower query—not from silently salvaging rows from that failed result.
Retrieve the applicable document, not just a relevant passage
Foundry IQ uses Azure AI Search for indexing and retrieval, with knowledge bases connecting selected sources. This is a substantive information-access contribution, not just a conversational interface.
Its migration documentation distinguishes a stable minimal, extractive contract from preview capabilities for richer planning and synthesis. Source and permission options vary with the interface and version. General availability of one subset is not proof that a different source-protection requirement is satisfied. This explanation does not prescribe a migration or weaken a control to fit a feature.
The actual retrieval response matters too. The documented REST response and MCP tool result expose different evidence structures. A successful call can still omit required material because of selection or output limits. Changing the interface must not silently discard evidence that the consumer needs.
A procedure's revision, approval, effective period and exceptions require their own basis. A newer draft is not necessarily the governing procedure. A citation supports tracing an answer to material; it does not independently establish that material's authority or complete coverage of the question.
Preserve the standing of workplace records
Work IQ describes internal context assembly and grounding across organisational information, alongside distinct chat, tool and persistent-workspace functions. Those can contribute more than raw document access, but a context summary and its source records remain distinguishable.
The selected API documentation specifies delegated user authentication, so available evidence is user-relative. The REST overview documents enterprise and web grounding by default, with web disablement applying to each message separately. Bind the actual source policy and interface rather than assuming one setting or file reference restricts every future interaction.
An email can carry a valid decision when the enterprise recognises it. Equally, a meeting summary can misstate an unapproved proposal. Preserve who decided, the scope and conditions, the effective period and any later suspension or replacement. Permission to read a record is not the author's decision authority, and absence from one user's response is not proof that no relevant record exists.
Persistent working storage can support continuity. It does not automatically turn every generated summary into accepted Institutional Memory.
Use evaluators as evidence, not a universal approval score
Microsoft's RAG evaluator reference distinguishes retrieval quality, grounding against supplied context and completeness against expected content. Response Completeness and Groundedness Pro are labelled Preview in the reviewed reference. These are useful checks when the actual input, evaluator and interpretation are appropriate.
A response may faithfully repeat an obsolete document. An expected answer can itself omit a mandatory exception. Establish the relevant source set and expected requirements independently of the candidate answer, and justify thresholds for the purpose. Do not turn several different scores into an unexplained universal confidence value.
Check the supported evaluation surface. The reviewed agent-evaluator documentation advises against specified evaluators on conversations containing certain tools, including Azure AI Search and Fabric Data Agent. A supported text-only check is not automatically evidence about the whole agent trace. Unsupported, failed or skipped checks must not be recorded as passing; corrected text requires assessment of that new rendition.
Compose evidence around a declared purpose
Arqua's Runtime Context Assembly and Context Qualification definitions distinguish forming understanding from evaluating its suitability for reliance. Both belong within Runtime Context Assembly. Qualified understanding does not itself authorise a consequential operation.
For each contributing product or service, identify the offered meaning, source and revision, relevant times, requesting and effective identities, permitted use, returned evidence and known gaps. Preserve native result and failure semantics rather than impose one convenient generic success flag. The required evidence depends on the claim being supported, not the number of connected systems.
An implementation can support a useful limited conclusion while withholding a broader one. Enterprise-approved definitions may be implemented within vendor artefacts; independent governance does not require another physical repository for every definition. Nor must all institutional information be centralised before selected evidence can be combined.
Begin with the smallest useful review
A proposed read-only service-review brief can separate four questions: represented dependencies, applicable observations, governing procedure and recorded decision standing. A fifth section records the evidence, gaps and limits on reliance. This is an illustrative design, not a report of an executed Microsoft composition.
Introduce a stale relationship, a delayed observation, a newer unapproved procedure, a restricted or missing decision record and a failed query. Assess each dependent claim without inventing inaccessible evidence or discarding independently valid information. A supported finding that further evidence is required can be useful without becoming permission to act.
Start with the sources and controls already adequate for the question. Add a graph, stream, retrieval service, workplace interface or evaluator only when its contribution is needed. Compare one implementation change at a time and preserve the evidence behind each result. No tenant, query, model or evaluator tests are reported here.
The architectural objective is to make different forms of evidence work together under explicit meaning, authority and reliance conditions—not to equate a collection of Microsoft capabilities with an implemented enterprise architecture.