/compare-cohorts

name: compare-cohorts
description: Use when comparing process cohorts, departments, periods or before-and-after results under matched definitions.
allowed-tools: Read, AskUserQuestion

Matched cohort comparisons

Use the [shared contract](../../references/analysis-contract.md) and [filters/units](../../references/filters-and-units.md). Fix the outcome definition/provenance, metric, aggregation, units, timezone and one selector S. Freeze the user's agreed baseline F, separately from current exploratory filters. Reuse metadata versions and exact typed cohort values. Use the [investigation method](../../references/investigation-method.md) for findings/status.

Ordered trace

1. Copy baseline F twice. For attribute cohorts append `{attributeName:A,dataType:T,attributeValues:[X],isInclusive:true}` to `attributeValueFilters` for A's cohort value, and the same shape with Y for B. Preserve other agreed baseline conditions; retain rework restrictions only if part of that baseline. Do not leak an exploratory/outcome-only subset into a full-population comparison. Change only the intended cohort condition. If values are unknown, request `get_attribute_values_v2 {S,attributeName:A,itemsPerPage:20,itemsToSkip:0,filterOptions:F}`; it returns values, not frequencies. 2. `get_process_overall_metrics_v2 {S,filterOptions:F_A}` -> `get_process_overall_metrics_v2 {S,filterOptions:F_B}`. No `metrics`, `groupBy` or aggregation parameter exists. Use the same returned metric definitions; native `CaseDuration`/`CaseWaitingTime` are means. 3. Only if a grouped explanation is needed and statistics is advertised, make a **matched pair**: `get_attribute_statistics {S,attributeName:A,filterOptions:F_A,itemsPerPage:5,itemsToSkip:0,metricToSortBy:"TotalDuration",sortOrder:"Descending"}`, then exactly the same call with F_B. Do not treat independent top-five pages as a complete aligned population.

Continue relevant value/group pages or matched follow-ups when needed, without fixed call quotas. Stop at sufficient evidence, no new information or a capability/error boundary; no blind sweeps. Do not create views or use two inclusive timeframes in one query as an A/B comparison.

Time and reproducibility

For periods, change only one explicit `timeframeFilters` start/end pair in each copied F. Case windows mean **Contained**, not cases Started/Completed/Intersecting or all events during the period. If completed-only membership including earlier starts is essential, stop/reframe or route to derive-metric/a user-confirmed existing matching view. Never invent a date-mode field or promise unknown end-boundary behavior.

If refresh is reported or reproducibility is at risk, recheck details. Changed PV/PMV invalidates the pair: mark **inconclusive**, stop or deliberately rerun both, not just the second. Process selection resolves latest; no direct version selector or atomic comparison snapshot exists. Views pin version context but their configuration is mutable/cached.

Report A/B counts, mean metrics, absolute delta and relative change only for a nonzero baseline. Empty cohorts mean insufficient evidence, not zero duration/infinite improvement. State scope/version and unmatched-group coverage; at most five headline rows. Recorded means do not establish closure, equal maturity or complete lifecycles. Report those as unknown unless supported; qualify the comparison, not an assumed causal improvement. Retain baseline/cohort filters, evidence references, status and alternatives. Overlapping groups cannot be added as unique-case totals. Object-centric comparisons belong in analyze-objects, not case-filter fallbacks.