Reconstruct a partial order without inventing missing history

Create a node for each recorded event and annotate its participant identity, incarnation, local sequence, clock domain, and relevant operation or message identifiers. Add local-order edges only where the execution model establishes them; merging several unsynchronized threads under one process label does not justify an arbitrary total sequence. Add send-to-receive edges for correlated instances of the same communication. Distinguish retry attempts, duplicate deliveries, and reused request IDs, since a correlation error can create a false path or even a cycle. Transitive closure then provides further order without requiring every intermediate relationship to have its own explicit edge.

Trace context helps connect operations, but a visual hierarchy is not a universal execution contract. A parent span may end before asynchronously launched work finishes, and a span link can express a relationship whose precise ordering depends on the application. Consult the instrumentation semantics rather than assuming that nesting on a screen proves enclosure in real time. Conversely, a missing span does not prove that the associated work did not run; sampling, loss, failed export, and uninstrumented paths can remove evidence. Preserve capture settings and loss indicators alongside the graph so the strength of later claims remains visible.

A timestamp with a valid error interval constrains when its event could physically have happened. Two disjoint intervals can establish physical precedence, but do not create a message or synchronization edge. Overlapping intervals leave physical order uncertain unless other evidence constrains it. If a known send-to-receive edge conflicts with the printed midpoint order, examine whether the intervals still admit a valid execution before declaring causality impossible. If even the bounds conflict, investigate clock assumptions, identifier matching, instrumentation placement, and data corruption rather than silently sorting away the contradiction.

For two event intervals, subtract the largest allowed start from the smallest allowed end for a conservative lower duration bound, and the smallest start from the largest end for the upper bound. A known causal send-to-receive relationship can additionally exclude negative transmission time. These interval calculations can be conservative because errors may correlate; independent endpoint choices need not all be jointly realizable. That is acceptable for a containing bound, provided the inputs themselves are valid. It is not acceptable to replace the bound with a preferred midpoint merely because the interval looks inconveniently wide.

Correct vector histories can establish concurrency under their stated participant and propagation model, whereas absent edges in a partial trace usually establish only ignorance. Label those outcomes differently. A topological ordering is one list consistent with known edges; another valid list can exchange independent nodes. It does not recover a unique hidden history. Record which conclusions use causal metadata, calibrated physical bounds, local duration measurements, or an explicit operational assumption. This makes the reconstruction reviewable and identifies the next measurement needed when a critical relationship remains unknown instead of disguising uncertainty as a polished timeline.

The vector instrument offers a small analogy for preserving incomparable dimensions. For each separate view, Reset, set both angle and magnitude, predict the projections, then inspect. Compare angle 0 degrees/magnitude 2 with angle 60 degrees/magnitude 2. One pair has larger x and smaller y, so neither is componentwise below the other. The instrument contains no events, messages, timestamps, or tracing export. Its purpose is to resist collapsing a partial-order comparison into one convenient scalar. The worked example performs the actual trace calculation, and its explicit communication edge supplies evidence that the geometric picture cannot provide.