Select the current interpretation before counting historical human prompts

For historical coding-agent comparisons, first build a source union, then choose one interpretation per source message, and only then filter for human-authored activity. Counting a legacy message table alone can omit continuing work; summing versioned projections can simultaneously inflate the recovered work. A user role is not sufficient evidence of a human instruction: runtime context, automatic continuations, delegated kickoff text, and real replies can share that role.

Keep event time distinct from session start and ingestion time. Preserve native message/session identifiers and source lineage; prefer the applicable newer classification before applying a human-origin filter, otherwise an older misclassification survives. Different historical schemas need different deduplication keys: stable native message IDs where present, with source sequence/timestamp/content evidence for older paired event/response representations. Do not sum aggregate counters across these sources or interpret coverage gaps as a change in employee effort.

Applicability and limits: this is a measurement method for immutable, versioned agent telemetry, not a productivity score or causal claim about work location. Even corrected prompt counts measure recorded interaction, not hours worked or customer benefit.

Evidence/example — 2026-09-16: read-only historical-query inspection found older v1 projections labeling environment context as human and representing a single instruction in both event and response records. Recent v2/v3 projections also retained multiple interpretations of the same native message. Selecting the newer interpretation before the human filter removed automatic continuation misclassifications; a separate native-session source restored conversations absent from the legacy message table. Relevant classification contract: https://github.com/e3-solutions/sherlock/pull/96 . No employee identities, performance figures, message content, credentials or raw logs are included.

This supports the linked stable-identity ingestion lesson and extends it to version-aware comparisons. Collector completeness and cross-period measurement equivalence remain unproven.