H-6736: Support campaign-based QA measurement, update low sample and outlier exclusion logic

apps/hash-frontend/src/pages/supply-chain/shared/docs/docs-modal/docs-content/settings.tsx

@@ -68,13 +68,12 @@ export const settingsSection: DocSectionDef = {
             </SettingRow>
           </Anchor>
           <Anchor id="outliers">
-            <SettingRow name="Exclude outliers">
-              Excludes values falling significantly outside the normal
-              distribution. This is done by computing the interquartile range
-              (IQR), i.e. the values falling between 25% and 75% of
-              observations, and then excluding any values which are greater than
-              1.5x IQR above the third quartile or less than 1.5x IQR below the
-              first quartile (aka Tukey fences).
+            <SettingRow name="Exclude outliers from mean">
+              Excludes outliers from every mean shown. Specifically, this
+              excludes values outside the Tukey fences (removing values which
+              fall greater than 1.5x the interquartile range below Q1 or greater
+              than 1.5x the interquartile range above Q3). Does not affect the
+              median, percentiles (P75, P95) or minimum and maximum.
             </SettingRow>
           </Anchor>
 

apps/hash-frontend/src/pages/supply-chain/shared/docs/docs-modal/docs-content/site-overview.tsx

@@ -137,7 +137,7 @@ export const siteOverviewSection: DocSectionDef = {
           <P>
             Dwell rows remain visible even when Exclude low samples is on,
             because a small number of high-value waits can still explain real
-            carrying cost. Low-sample dwell rows are labelled.
+            carrying cost. Low and limited-sample dwell rows are labelled.
           </P>
         </>
       ),
@@ -260,8 +260,8 @@ export const siteOverviewSection: DocSectionDef = {
             </LI>
             <LI>
               <Term>Exclude low samples</Term> &mdash; hide Planning and Trend
-              rows with fewer than 10 observations, so rankings are not
-              dominated by noisy single-event steps.
+              rows with fewer than 5 observations. Rows with 5&ndash;9
+              observations remain visible with a "limited" sample badge.
             </LI>
           </UL>
           <P>

apps/hash-frontend/src/pages/supply-chain/shared/docs/docs-modal/docs-content/step-detail.tsx

@@ -82,11 +82,11 @@ export const stepDetailDoc: DocEntry = {
           <Term>Time range</Term> buttons (3m, 6m, 12) filter the panel.
         </LI>
         <LI>
-          <Term>Outlier count</Term> appears in the header when the Exclude
-          outliers setting removes timing observations from the current step.
-          See{" "}
+          <Term>Outlier count</Term> appears in the header when values are
+          excluded from the mean. The raw observations and percentile statistics
+          remain unchanged. See{" "}
           <CrossRef to={{ section: "settings", sub: "outliers" }}>
-            Exclude outliers
+            Exclude outliers from mean
           </CrossRef>
           .
         </LI>

apps/hash-frontend/src/pages/supply-chain/shared/docs/docs-modal/docs-content/step-qa.tsx

@@ -8,19 +8,18 @@ export const qaDoc: DocEntry = {
   render: () => (
     <>
       <Lead>
-        QA hold measures the time a finished batch waits between production
+        QA hold measures the time a production campaign waits between production
         completion and QA release &mdash; the quality inspection and hold
         period.
       </Lead>
       <P>
-        <Term>What it measures:</Term> production receipt (the batch is complete
-        and received into inventory) to QA release (the batch passes inspection
-        and is cleared for use). One observation per finished-good batch.
+        <Term>What it measures:</Term> the time between a production campaign
+        finishing and the associated QA release. The full 'data' table also
+        shows the time between each batch's production finish and QA release.
       </P>
       <P>
-        <Term>Time filtering:</Term> observations are anchored to the
-        production-receipt date, so the selected window picks batches that
-        completed production inside that period.
+        <Term>Time filtering:</Term> observations are anchored to the campaign
+        end date.
       </P>
       <P>
         The wait that follows QA release &mdash; from release to dispatch

apps/hash-frontend/src/pages/supply-chain/shared/header-actions.test.tsx

@@ -0,0 +1,45 @@
+// @vitest-environment jsdom
+import { cleanup, fireEvent, render, screen } from "@testing-library/react";
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+import { OutlierContext } from "./cost";
+import { AnalysisSettingsPanel } from "./header-actions";
+import { MeasureContext } from "./measure-context";
+import { TimeRangeContext } from "./time-range-context";
+
+vi.mock("@hashintel/ds-components", () => ({
+  Icon: () => null,
+  NumberInput: () => <input readOnly />,
+}));
+
+describe("AnalysisSettingsPanel", () => {
+  afterEach(cleanup);
+
+  it("allows the mean outlier policy to change under non-mean headline measures", () => {
+    const setExcludeOutliers = vi.fn();
+
+    render(
+      <TimeRangeContext.Provider
+        value={{ timeRange: "12m", setTimeRange: vi.fn() }}
+      >
+        <MeasureContext.Provider
+          value={{ measure: "p95", setMeasure: vi.fn() }}
+        >
+          <OutlierContext.Provider
+            value={{ excludeOutliers: true, setExcludeOutliers }}
+          >
+            <AnalysisSettingsPanel />
+          </OutlierContext.Provider>
+        </MeasureContext.Provider>
+      </TimeRangeContext.Provider>,
+    );
+
+    const checkbox = screen.getByRole("checkbox", {
+      name: "Exclude outliers from mean",
+    });
+    expect(checkbox).toHaveProperty("checked", true);
+    expect(checkbox).toHaveProperty("disabled", false);
+    fireEvent.click(checkbox);
+    expect(setExcludeOutliers).toHaveBeenCalledWith(false);
+  });
+});

apps/hash-frontend/src/pages/supply-chain/shared/header-actions.tsx

@@ -293,7 +293,7 @@ export const AnalysisSettingsPanel = ({
               checked={excludeOutliers}
               onChange={(event) => setExcludeOutliers(event.target.checked)}
             />
-            Exclude outliers
+            Exclude outliers from mean
           </label>
           {children}
         </div>

apps/hash-frontend/src/pages/supply-chain/shared/node-badges.test.ts

@@ -55,20 +55,20 @@ describe("node R:/C: badge recompute from shipped series", () => {
     expect(windowed.yield_summary?.median).toBe(88);
   });
 
-  it("drops outliers from the yield series under the outlier rule", () => {
+  it("does not apply the timing mean rule to yield observations", () => {
     const withOutlier: NodeYieldSeries = {
       reference: 95,
       observations: [...yieldSeries.observations, obs("2026-05", 5)],
     };
     const node = makeNode({ type: "production", yield_series: withOutlier });
     const selected = applyOutlierSelectionToNode(node, true);
-    // The 5% point is far below the others' IQR fence and is dropped.
+    // Yield remains raw: the setting applies only to timing means.
     expect(
       selected.yield_series?.observations.some(
         (observation) => observation.value === 5,
       ),
-    ).toBe(false);
+    ).toBe(true);
     const windowed = windowGraphNodeToRange(selected, "12m");
-    expect(windowed.yield_summary?.n).toBe(4);
+    expect(windowed.yield_summary?.n).toBe(5);
   });
 });

apps/hash-frontend/src/pages/supply-chain/shared/normalize-contract.ts

@@ -104,10 +104,8 @@ function buildMonthlyFromObservations(obs: Observation[]): MonthlyBucket[] {
 }
 
 /**
- * Single-source records derive. When a step ships only the canonical
- * `detail_rows` (+ `value_col`/`ref_date_col`) and no precomputed timing series,
- * rehydrate observations/durations/monthly/stats from the rows. A strict no-op
- * whenever `observations` are already present.
+ * Single-source records derive. Campaign rows are the canonical timing source
+ * when present; detail rows remain batch evidence and the v1.2 fallback.
  */
 function ensureTimingSeriesStats<
   T extends {
@@ -223,13 +221,15 @@ function fillMonthlyTiming(
 export function deriveTimingFromRecords(step: StepDetail): StepDetail;
 export function deriveTimingFromRecords(step: StepDetailWire): StepDetailWire;
 export function deriveTimingFromRecords(step: StepDetailWire): StepDetailWire {
+  const campaignRows =
+    step.timing_grain === "campaign" ? step.campaign_rows?.rows : undefined;
   const existingObservations = step.observations ?? [];
-  if (existingObservations.length > 0) {
+  if (!campaignRows?.length && existingObservations.length > 0) {
     return step;
   }
   const valueCol = step.value_col;
   const dateCol = step.ref_date_col;
-  const rows = step.detail_rows?.rows;
+  const rows = campaignRows?.length ? campaignRows : step.detail_rows?.rows;
   if (!valueCol || !dateCol || !rows || rows.length === 0) {
     return step;
   }

apps/hash-frontend/src/pages/supply-chain/shared/observation-labels.test.ts

@@ -0,0 +1,33 @@
+import { describe, expect, it } from "vitest";
+
+import { countNoun, countTooltip, shortCountLabel } from "./observation-labels";
+
+describe("campaign observation labels", () => {
+  const qaCampaign = {
+    id: "prod_to_qa",
+    label: "Production to QA",
+    type: "qa_hold" as const,
+    timingGrain: "campaign" as const,
+  };
+
+  it("labels campaign-grain QA timing observations as campaigns", () => {
+    expect(countNoun(qaCampaign)).toBe("campaigns");
+    expect(shortCountLabel(7, qaCampaign)).toBe("7 campaigns");
+    expect(
+      countTooltip({ ...qaCampaign, count: 7, rangeLabel: "12m" }),
+    ).toContain("7 campaigns in the last 12 months");
+  });
+
+  it("keeps yield and consumption terminology ahead of timing grain", () => {
+    expect(countNoun({ ...qaCampaign, dimension: "yield" })).toBe(
+      "production orders",
+    );
+    expect(
+      countNoun({
+        ...qaCampaign,
+        dimension: "consumption",
+        selectedComponent: true,
+      }),
+    ).toBe("component events");
+  });
+});

apps/hash-frontend/src/pages/supply-chain/shared/observation-labels.ts

@@ -8,12 +8,14 @@ interface CountContext {
   type: StepType;
   dimension?: CountDimension;
   selectedComponent?: boolean;
+  timingGrain?: "campaign" | null;
 }
 
 interface CountTooltipContext extends CountContext {
   count: number;
   rangeLabel?: string | null;
   nBatches?: number | null;
+  nCampaigns?: number | null;
   nMovements?: number | null;
 }
 
@@ -34,6 +36,9 @@ export function countNoun(ctx: CountContext): string {
   if (ctx.dimension === "supplier") {
     return "schedule lines";
   }
+  if (ctx.timingGrain === "campaign") {
+    return "campaigns";
+  }
   return "events";
 }
 
@@ -80,7 +85,9 @@ export function dateAnchorLabel(ctx: CountContext): string {
     case "production":
       return "schedule start";
     case "qa_hold":
-      return "production receipt date";
+      return ctx.timingGrain === "campaign"
+        ? "campaign reference date"
+        : "production receipt date";
     case "post_qa_ship":
       return "QA release date";
     case "transit":
@@ -114,7 +121,9 @@ function eventMethodology(ctx: CountContext): string {
     case "production":
       return "each production schedule campaign contributes one duration event.";
     case "qa_hold":
-      return "each finished-good batch contributes one QA-hold event.";
+      return ctx.timingGrain === "campaign"
+        ? "each campaign contributes one timing observation; batch rows are supporting evidence."
+        : "each finished-good batch contributes one QA-hold event.";
     case "post_qa_ship":
       return "each dispatch (customer delivery or hub transfer) contributes one post-QA dwell event.";
     case "transit":
@@ -130,6 +139,9 @@ export function countTooltip(ctx: CountTooltipContext): string {
   const label = shortCountLabel(ctx.count, ctx);
   const period = rangeLabel(ctx.rangeLabel);
   const base = `${label} in ${period}. Filtered by ${dateAnchorLabel(ctx)}; ${eventMethodology(ctx)}`;
+  if (ctx.nCampaigns != null && ctx.nBatches != null) {
+    return `${base} All-time source coverage: ${ctx.nCampaigns} campaigns across ${ctx.nBatches} batches.`;
+  }
   if (ctx.nBatches != null && ctx.nMovements != null) {
     return `${base} All-time source coverage: ${ctx.nBatches} batches, ${ctx.nMovements} movements.`;
   }

apps/hash-frontend/src/pages/supply-chain/shared/outlier-selection.test.ts

@@ -10,10 +10,10 @@ import {
   applyOutlierSelectionToNode,
   applyOutlierSelectionToStep,
 } from "./outlier-selection";
+import { computeStats } from "./stats";
 
-// Characterization tests for the client-side Tukey 1.5x IQR outlier rule.
-// Behaviour pinned here MUST survive the contract refactor (the selected-view
-// semantics stay even as the duplicated outlier paths were collapsed).
+// Regression tests for mean-only client-side Tukey 1.5x IQR filtering.
+// Raw observations, counts, and percentile statistics remain unchanged.
 
 describe("applyOutlierSelectionToStep", () => {
   // One clear high outlier (100) far beyond the Tukey upper fence.
@@ -30,10 +30,14 @@ describe("applyOutlierSelectionToStep", () => {
       obs("2026-04", 100),
     ]);
 
-  it("drops out-of-fence points and recomputes stats when excluding outliers", () => {
+  it("filters only the mean while retaining raw observations and percentiles", () => {
     const selected = applyOutlierSelectionToStep(withOutlier(), true);
-    expect(selected.durations).toEqual([10, 11, 12, 13, 14, 15, 16, 17]);
-    expect(selected.stats.n).toBe(8);
+    expect(selected.durations).toEqual([10, 11, 12, 13, 14, 15, 16, 17, 100]);
+    expect(selected.observations).toHaveLength(9);
+    expect(selected.stats.n).toBe(9);
+    expect(selected.stats.mean).toBe(13.5);
+    expect(selected.stats.p95).toBe(withOutlier().stats.p95);
+    expect(selected.stats.max).toBe(100);
     expect(selected.excluded_count).toBe(1);
     expect(selected.excluded_pct).toBeCloseTo(100 / 9, 1);
   });
@@ -82,12 +86,13 @@ describe("applyOutlierSelectionToStep", () => {
     // Headline untouched, secondary loses its outlier.
     expect(selected.stats.n).toBe(4);
     expect(selected.excluded_count).toBe(0);
-    expect(selected.complete_timing?.stats.n).toBe(8);
+    expect(selected.complete_timing?.stats.n).toBe(9);
+    expect(selected.complete_timing?.stats.max).toBe(200);
     expect(
       selected.complete_timing?.observations.map(
         (observation) => observation.value,
       ),
-    ).not.toContain(200);
+    ).toContain(200);
   });
 
   it("leaves complete_timing untouched when including outliers", () => {
@@ -102,9 +107,6 @@ describe("applyOutlierSelectionToStep", () => {
 });
 
 describe("applyOutlierSelectionToNode", () => {
-  // Shipped v1 data carries a single base series (no raw_*/filtered_*), so today
-  // both toggle states return the base series untouched. The fixture uses a tight
-  // distribution so this invariant also holds once Tukey IQR filtering lands.
   it("returns the base series when including outliers", () => {
     const out = applyOutlierSelectionToNode(makeNode(), false);
     expect(out.observations?.map((observation) => observation.value)).toEqual([
@@ -118,4 +120,27 @@ describe("applyOutlierSelectionToNode", () => {
       10, 12, 11, 13,
     ]);
   });
+
+  it("keeps raw P95 while filtering an outlier from the mean", () => {
+    const observations = [
+      obs("2026-01", 10),
+      obs("2026-01", 11),
+      obs("2026-02", 12),
+      obs("2026-02", 13),
+      obs("2026-03", 14),
+      obs("2026-03", 15),
+      obs("2026-04", 16),
+      obs("2026-04", 17),
+      obs("2026-04", 100),
+    ];
+    const node = makeNode({
+      observations,
+      stats: computeStats(observations.map((observation) => observation.value)),
+    });
+    const selected = applyOutlierSelectionToNode(node, true);
+
+    expect(selected.stats.p95).toBe(node.stats.p95);
+    expect(selected.observations).toHaveLength(9);
+    expect(selected.mean_observations).toHaveLength(8);
+  });
 });