Measurement Levels, Units, and Evidence

Correspondence Measurement requires explicit level and unit specification because the same evidence can support different claims at different levels. A note in a decision record may be an evidence unit. The decision itself may be an observation unit. The relationship between a decision and work artifact may be the measurement object. A capability may be the reporting unit. An organization-level assessment may later use the result, but the organization is not automatically the unit of analysis.

The framework distinguishes four level systems. Organizational object levels include representations, representation pairs, relationships, correspondence relations, paths, loops, subnetworks, representation networks, capabilities, processes, organizational units, organizations, ecosystems, states, transitions, and trajectories. Unit-role levels include evidence unit, observation unit, measurement object, measurement unit, unit of analysis, reporting unit, and assessment unit. Result-representation levels include categorical, ordinal, interval-like, ratio-like, probabilistic, interval-valued, relational, graph, configuration, profile, state, transition, and trajectory results. Epistemic levels distinguish source evidence, extracted observation, coded observation, verified observation, inferred observation, interpreted result, and assessment judgment.

Figure 2 summarizes this logic through the Correspondence Measurement Instance. A measurement instance is a bounded record of one measurement execution. It includes identity, construct, property, object, level, context, time, evidence, observations, inferences, procedure, result, warrant qualifiers, provenance, review status, limitations, and permitted use.

FIG-0017 depicts Measurement Instance and Traceability Chain. It represents Identity, Construct and property, Object and level, Context and time, Source evidence, Observation, AI-assisted inference, Procedure and rules, Result package, Coverage, Confidence, Uncertainty, Provenance, Review status, Limitations, Permitted interpretation, Bounded use. The intended relationships are: A measurement result is a traceable package; AI-assisted inference remains distinct from source evidence and reviewed input. The figure should be read with this boundary: Result does not equal score; AI inference does not equal evidence. — Figure 2. Measurement Instance and Traceability Chain. Main Figure 2 shows that a Correspondence Measurement Instance contains identity, construct, object, level, context, time, evidence, observation, inference, procedure, result, warrant qualifiers, provenance, review status, limitations, and permitted use. Dashed arrows indicate inference or downstream interpretation.

Evidence is central but bounded. Evidence may include representations, relationships, traces, artifacts, records, events, conversations, logs, outcome claims, governance records, human judgments, AI outputs, or external artifacts. But evidence is not self-interpreting. It must be admissible for the measurement purpose, temporally relevant, sufficiently complete, and traceable. Evidence also has states: available, missing, inaccessible, stale, conflicting, contested, inferred, non-applicable, or superseded. These states must be visible because they affect warrant differently.

Coverage is not value. If evidence is missing, the correct result may be unknown or underdetermined, not low correspondence. If evidence is extensive, the result may still be weak. If evidence conflicts, conflict must be represented rather than averaged away. If AI infers a relationship, the relationship remains inferred unless independently verified. If an observation is coded, the coding method and review state must be preserved.

This evidence model has consequences for research design. Empirical studies of Organizational Correspondence should record the source and state of evidence, the method of extraction or coding, the role of human and AI actors, the temporal interval, and the limitations of inference. Without this structure, researchers cannot know whether they are comparing correspondence, evidence availability, coding decisions, or measurement procedures.

For example, a decision-to-work measurement instance might identify a product decision as the source representation, an implemented work artifact as the target representation, and dependency preservation as the measured property. The evidence package may include the decision record, linked work item, change history, acceptance evidence, and reviewer notes. The result may state that the work artifact preserves the decision's required dependency relation with moderate warrant, while also recording that outcome evidence is not yet available. Such a result is not an assessment of success. It is a bounded measurement of a specified relationship under a specified evidence state.