diff --git a/.trivyignore b/.trivyignore
index ccf17286e..34a98f1b7 100644
--- a/.trivyignore
+++ b/.trivyignore
@@ -31,3 +31,11 @@ CVE-2026-48815
 # entry, the matching `auditConfig.ignoreGhsas` entry in pnpm-workspace.yaml,
 # and raise the floor once node-forge 1.4.1 is out.
 CVE-2026-85393
+
+# braces <= 3.0.3 stack overflow on deeply nested patterns (CVE-2026-93687).
+# braces arrives only through micromatch -> fast-glob in development tools
+# (@next/eslint-plugin-next, @ts-morph/common, the shadcn CLI); `pnpm why
+# braces --prod` is empty and the production image does not contain it, and
+# those tools only expand patterns written in this repository. No fixed
+# release is published (latest is 3.0.3). Remove this entry once one is.
+CVE-2026-93687
diff --git a/docs/api/openapi.yaml b/docs/api/openapi.yaml
index 611091141..d270f2e96 100644
--- a/docs/api/openapi.yaml
+++ b/docs/api/openapi.yaml
@@ -14064,7 +14064,16 @@ paths:
         "401": *a1
         "403": *a10
         "409": *a7
-        "422": *a4
+        "422":
+          description: Nothing was written. `meta.errorCode` = `labs.create.invalid` when the body failed validation (every issue
+            under `details.issues`), or `labs.unit.mismatch` when a numeric reading states a `unit` other than the one
+            its marker is tracked in (`meta.markerUnit`, `meta.readingUnit`). A unit is compared only when one is sent;
+            with a `biomarkerId` it may be omitted and the marker's unit is then the stated one. Spelling is forgiven
+            (`mmol/l`, `ug/L` for `µg/L`); a different unit is refused, never relabelled.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ErrorEnvelope"
         "429": *a11
   /api/labs/{id}:
     get:
@@ -14265,9 +14274,9 @@ paths:
         - Labs
       summary: Commit confirmed Lab-OCR rows
       description: 'Writes ONLY the rows the human confirmed on the review screen. Each row resolves-or-mints a user-scoped
-        biomarker and creates a `LabResult` with `source: "OCR"`; a row that duplicates a live reading is skipped.
-        Idempotent (Idempotency-Key). Audits as `labs.ocr.commit`. `userId` is narrowed from the session, never a body
-        field.'
+        biomarker and creates a `LabResult` with `source: "OCR"`; a row that duplicates a live reading, or states a unit
+        other than that of its marker, is skipped (`skipped[].reason`). Idempotent (Idempotency-Key). Audits as
+        `labs.ocr.commit`. `userId` is narrowed from the session, never a body field.'
       parameters:
         - *a6
       requestBody:
@@ -31762,15 +31771,16 @@ components:
                 labs.biomarker.notFound, labs.biomarker.update.invalid, labs.biomarker.update.referenceRangeInvalid,
                 labs.create.invalid, labs.ocr.budgetExceeded, labs.ocr.extractFailed, labs.ocr.fileTooLarge,
                 labs.ocr.fileType, labs.ocr.localOcrDisabled, labs.ocr.pdfNeedsAnthropic, labs.ocr.providerUnsupported,
-                labs.ocr.rateLimited, labs.restore.invalid, labs.result.notFound, labs.update.invalid,
-                labs.update.linkedFieldsImmutable, labs.update.numericExpected, labs.update.qualitativeExpected,
-                labs.update.referenceRangeInvalid, labs.update.sourceReferenceRangeInvalid; managed_profile —
-                managed_profile.guardian.duplicate, managed_profile.guardian.managed_invitee,
-                managed_profile.guardian.required, managed_profile.not_found; measurement — measurement.batch.invalid,
-                measurement.batch.source_not_permitted, measurement.batch.too_large, measurement.bulk-delete.invalid,
-                measurement.create.invalid, measurement.create.source_not_permitted, measurement.delete.too_large,
-                measurement.duplicate_timestamp, measurement.restore.invalid, measurement.update.server_owned_source;
-                medication — medication.intake.import.invalid_format; medications —
+                labs.ocr.rateLimited, labs.restore.invalid, labs.result.notFound, labs.unit.mismatch,
+                labs.update.invalid, labs.update.linkedFieldsImmutable, labs.update.numericExpected,
+                labs.update.qualitativeExpected, labs.update.referenceRangeInvalid,
+                labs.update.sourceReferenceRangeInvalid; managed_profile — managed_profile.guardian.duplicate,
+                managed_profile.guardian.managed_invitee, managed_profile.guardian.required, managed_profile.not_found;
+                measurement — measurement.batch.invalid, measurement.batch.source_not_permitted,
+                measurement.batch.too_large, measurement.bulk-delete.invalid, measurement.create.invalid,
+                measurement.create.source_not_permitted, measurement.delete.too_large, measurement.duplicate_timestamp,
+                measurement.restore.invalid, measurement.update.server_owned_source; medication —
+                medication.intake.import.invalid_format; medications —
                 medications.intake.bulk.apple_health_not_mirrored, medications.intake.bulk.invalid,
                 medications.intake.bulk.too_large, medications.intake.force_slot.invalid,
                 medications.intake.force_slot.occupied, medications.intake.injection_site.disallowed,
@@ -63380,6 +63390,13 @@ components:
           enum:
             - new
             - existing
+        markerUnit:
+          anyOf:
+            - type: string
+            - type: "null"
+          description: "The unit the matched marker is tracked in, or null when the marker does not exist yet (it will adopt this
+            row's unit). A row whose `unit` differs from it cannot be committed: the commit skips it with `reason:
+            unit_mismatch`, and the review screen marks it before Save."
         duplicateOf:
           anyOf:
             - type: string
@@ -63395,6 +63412,7 @@ components:
         - takenAt
         - confidence
         - biomarkerMatch
+        - markerUnit
         - duplicateOf
       additionalProperties: false
       description: "One proposed reading transcribed from the upload. UNTRUSTED model output annotated server-side:
@@ -63435,7 +63453,11 @@ components:
                 type: string
               reason:
                 type: string
-                const: duplicate
+                enum:
+                  - duplicate
+                  - unit_mismatch
+                description: "`duplicate`: the row matched a live reading at commit time. `unit_mismatch`: the row states a unit other
+                  than the one its marker is tracked in, so it was not written."
             required:
               - analyte
               - reason
@@ -63455,7 +63477,8 @@ components:
         - skipped
         - outcome
       additionalProperties: false
-      description: 'The rows written (`source: "OCR"`), those skipped as commit-time duplicates, and the resolved outcome.'
+      description: 'The rows written (`source: "OCR"`), those skipped (commit-time duplicates, or a unit that differs from
+        that of the marker), and the resolved outcome.'
     OcrCommittedLabResult:
       type: object
       properties:
diff --git a/e2e/dialog-horizontal-overflow.spec.ts b/e2e/dialog-horizontal-overflow.spec.ts
index 72d49564f..54267f858 100644
--- a/e2e/dialog-horizontal-overflow.spec.ts
+++ b/e2e/dialog-horizontal-overflow.spec.ts
@@ -363,6 +363,7 @@ const REVIEW_ROWS = ["LDL", "HDL", "Glucose", "Creatinine", "ALT", "CRP"].map(
     takenAt: index === 0 ? null : "2026-06-10",
     confidence: { analyte: 0.9, value: 0.9, unit: 0.9, range: 0.9 },
     biomarkerMatch: "existing",
+    markerUnit: "mmol/L",
     duplicateOf: null,
   }),
 );
diff --git a/messages/de.json b/messages/de.json
index d319c56a1..64d0b6e38 100644
--- a/messages/de.json
+++ b/messages/de.json
@@ -9075,6 +9075,10 @@
       "errorValue": "Bitte einen Wert angeben.",
       "errorResult": "Bitte ein Ergebnis angeben.",
       "errorUnit": "Bitte eine Einheit angeben.",
+      "errorUnitMismatch": "Dieser Marker wird in {markerUnit} geführt. Ändere die Einheit, falls der Befund falsch gelesen wurde, oder wähle den Wert ab.",
+      "unitDiffersBadge": "Geführt in {markerUnit}; dieser Befund nennt {unit}",
+      "savedPartialUnitToast": "{count} Laborwerte gespeichert. Nicht gespeichert, weil die Einheit von der des Markers abweicht: {analytes}",
+      "savedNothingUnitToast": "Nichts gespeichert: Die Einheit weicht bei {analytes} von der des Markers ab",
       "saveFailed": "Werte konnten nicht gespeichert werden. Bitte erneut versuchen.",
       "providerUnsupported": "Dein KI-Anbieter kann keine Bilder lesen. Konfiguriere ein Modell mit Bildverständnis.",
       "consentRequired": "Für das Scannen eines Befunds ist die Einwilligung zur KI-Verarbeitung erforderlich.",
diff --git a/messages/en.json b/messages/en.json
index d077dfde3..1e8f3ae7c 100644
--- a/messages/en.json
+++ b/messages/en.json
@@ -9075,6 +9075,10 @@
       "errorValue": "Add a value.",
       "errorResult": "Add a result.",
       "errorUnit": "Add a unit.",
+      "errorUnitMismatch": "This marker is tracked in {markerUnit}. Change the unit if the report was misread, or deselect this reading.",
+      "unitDiffersBadge": "Tracked in {markerUnit}; this report says {unit}",
+      "savedPartialUnitToast": "{count} lab results saved. Not saved, because the unit differs from the marker's: {analytes}",
+      "savedNothingUnitToast": "Nothing saved: the unit differs from the marker's for {analytes}",
       "saveFailed": "Couldn't save the readings. Please try again.",
       "providerUnsupported": "Your AI provider can't read images. Configure a vision-capable model.",
       "consentRequired": "AI processing consent is required to scan a report.",
diff --git a/messages/es.json b/messages/es.json
index b1038b4e0..87c229f67 100644
--- a/messages/es.json
+++ b/messages/es.json
@@ -9075,6 +9075,10 @@
       "errorValue": "Añade un valor.",
       "errorResult": "Añade un resultado.",
       "errorUnit": "Añade una unidad.",
+      "errorUnitMismatch": "Este marcador se registra en {markerUnit}. Cambia la unidad si el informe se leyó mal o deselecciona este valor.",
+      "unitDiffersBadge": "Registrado en {markerUnit}; este informe indica {unit}",
+      "savedPartialUnitToast": "{count} resultados guardados. No guardados porque la unidad difiere de la del marcador: {analytes}",
+      "savedNothingUnitToast": "No se guardó nada: la unidad difiere de la del marcador en {analytes}",
       "saveFailed": "No se pudieron guardar los valores. Inténtalo de nuevo.",
       "providerUnsupported": "Tu proveedor de IA no puede leer imágenes. Configura un modelo con visión.",
       "consentRequired": "Se requiere consentimiento de procesamiento de IA para escanear un informe.",
diff --git a/messages/fr.json b/messages/fr.json
index 9a8f0222e..97364f463 100644
--- a/messages/fr.json
+++ b/messages/fr.json
@@ -9075,6 +9075,10 @@
       "errorValue": "Indiquez une valeur.",
       "errorResult": "Indiquez un résultat.",
       "errorUnit": "Indiquez une unité.",
+      "errorUnitMismatch": "Ce marqueur est suivi en {markerUnit}. Modifiez l’unité si le compte rendu a été mal lu, ou désélectionnez cette valeur.",
+      "unitDiffersBadge": "Suivi en {markerUnit} ; ce compte rendu indique {unit}",
+      "savedPartialUnitToast": "{count} résultats enregistrés. Non enregistrés, car l’unité diffère de celle du marqueur : {analytes}",
+      "savedNothingUnitToast": "Rien n’a été enregistré : l’unité diffère de celle du marqueur pour {analytes}",
       "saveFailed": "Impossible d'enregistrer les valeurs. Veuillez réessayer.",
       "providerUnsupported": "Votre fournisseur d'IA ne peut pas lire les images. Configurez un modèle capable de vision.",
       "consentRequired": "Le consentement au traitement par IA est requis pour scanner un compte rendu.",
diff --git a/messages/it.json b/messages/it.json
index 13192eaee..3f40e06b7 100644
--- a/messages/it.json
+++ b/messages/it.json
@@ -9075,6 +9075,10 @@
       "errorValue": "Inserisci un valore.",
       "errorResult": "Inserisci un risultato.",
       "errorUnit": "Inserisci un'unità.",
+      "errorUnitMismatch": "Questo marcatore è registrato in {markerUnit}. Cambia l’unità se il referto è stato letto male, oppure deseleziona questo valore.",
+      "unitDiffersBadge": "Registrato in {markerUnit}; questo referto indica {unit}",
+      "savedPartialUnitToast": "{count} risultati salvati. Non salvati perché l’unità è diversa da quella del marcatore: {analytes}",
+      "savedNothingUnitToast": "Nulla salvato: l’unità è diversa da quella del marcatore per {analytes}",
       "saveFailed": "Impossibile salvare i valori. Riprova.",
       "providerUnsupported": "Il tuo provider IA non può leggere le immagini. Configura un modello con visione.",
       "consentRequired": "Per scansionare un referto è richiesto il consenso al trattamento tramite IA.",
diff --git a/messages/ko.json b/messages/ko.json
index 271f03069..6ed108d52 100644
--- a/messages/ko.json
+++ b/messages/ko.json
@@ -9075,6 +9075,10 @@
       "errorValue": "값을 입력하세요.",
       "errorResult": "결과를 입력하세요.",
       "errorUnit": "단위를 입력하세요.",
+      "errorUnitMismatch": "이 지표는 {markerUnit} 단위로 기록됩니다. 결과지를 잘못 읽었다면 단위를 바꾸거나 이 값의 선택을 해제하세요.",
+      "unitDiffersBadge": "{markerUnit} 단위로 기록 중; 이 결과지는 {unit}",
+      "savedPartialUnitToast": "검사 결과 {count}개를 저장했습니다. 단위가 지표와 달라 저장하지 않은 항목: {analytes}",
+      "savedNothingUnitToast": "저장된 항목이 없습니다: {analytes}의 단위가 지표와 다릅니다",
       "saveFailed": "측정값을 저장하지 못했어요. 다시 시도해 주세요.",
       "providerUnsupported": "AI 제공자가 이미지를 읽지 못해요. 이미지를 읽는 모델로 설정해 주세요.",
       "consentRequired": "결과지를 스캔하려면 AI 처리 동의가 필요해요.",
diff --git a/messages/pl.json b/messages/pl.json
index 388671013..84bb9d1dc 100644
--- a/messages/pl.json
+++ b/messages/pl.json
@@ -9075,6 +9075,10 @@
       "errorValue": "Podaj wartość.",
       "errorResult": "Podaj wynik.",
       "errorUnit": "Podaj jednostkę.",
+      "errorUnitMismatch": "Ten wskaźnik jest prowadzony w {markerUnit}. Zmień jednostkę, jeśli wynik został źle odczytany, albo odznacz tę wartość.",
+      "unitDiffersBadge": "Prowadzony w {markerUnit}; ten wynik podaje {unit}",
+      "savedPartialUnitToast": "Zapisano {count} wyników. Nie zapisano, bo jednostka różni się od jednostki wskaźnika: {analytes}",
+      "savedNothingUnitToast": "Nic nie zapisano: jednostka różni się od jednostki wskaźnika dla {analytes}",
       "saveFailed": "Nie udało się zapisać wyników. Spróbuj ponownie.",
       "providerUnsupported": "Twój dostawca AI nie może odczytywać obrazów. Skonfiguruj model z obsługą obrazu.",
       "consentRequired": "Do zeskanowania wyniku wymagana jest zgoda na przetwarzanie przez AI.",
diff --git a/src/app/api/labs/ocr/commit/route.ts b/src/app/api/labs/ocr/commit/route.ts
index 6f193ef36..8a8647108 100644
--- a/src/app/api/labs/ocr/commit/route.ts
+++ b/src/app/api/labs/ocr/commit/route.ts
@@ -33,6 +33,7 @@ import { withIdempotency } from "@/lib/idempotency";
 import { enqueueReminderSatisfy } from "@/lib/jobs/reminder-satisfy";
 import { emitDataArrival } from "@/lib/arrivals/emit-shared";
 import { resolveOrMintBiomarker } from "@/lib/labs/biomarker-store";
+import { sameLabUnit } from "@/lib/labs/unit-normalise";
 import { parseReferenceRange } from "@/lib/labs/parse-reference-range";
 import { serialiseLabResult } from "@/lib/labs/serialise";
 import {
@@ -178,6 +179,21 @@ async function commitOcrRows(request: NextRequest) {
       panel: row.panel ?? null,
     });
 
+    // The marker's unit is the one the reading is stored under, so a row that
+    // states another unit cannot be written: the number would be kept under a
+    // unit it is not in. The review screen marks such a row before Save; this
+    // is the backstop for a client that did not, and for a marker whose unit
+    // changed between the scan and the commit. A freshly minted marker adopted
+    // this row's unit, so it cannot disagree.
+    if (
+      !isQualitative &&
+      row.unit !== undefined &&
+      !sameLabUnit(row.unit, biomarker.unit)
+    ) {
+      skipped.push({ analyte: row.analyte.trim(), reason: "unit_mismatch" });
+      continue;
+    }
+
     // Field-by-field — never spread the parsed row. The row stamps the resolved
     // catalog name/unit/range as historical truth and keeps the FK.
     const created = await prisma.labResult.create({
diff --git a/src/app/api/labs/route.ts b/src/app/api/labs/route.ts
index d4fbf47d9..bd463890a 100644
--- a/src/app/api/labs/route.ts
+++ b/src/app/api/labs/route.ts
@@ -18,6 +18,7 @@ import { withIdempotency } from "@/lib/idempotency";
 import { enqueueReminderSatisfy } from "@/lib/jobs/reminder-satisfy";
 import { emitDataArrival } from "@/lib/arrivals/emit-shared";
 import { resolveOrMintBiomarker } from "@/lib/labs/biomarker-store";
+import { sameLabUnit } from "@/lib/labs/unit-normalise";
 import {
   type ResolvedBiomarker,
   serialiseLabResult,
@@ -252,6 +253,32 @@ async function postLabResult(request: NextRequest) {
     });
   }
 
+  // The marker's unit is the one every reading is stored under, so a reading
+  // that states another unit cannot be written: the number would be kept and
+  // the label swapped, and nothing downstream could tell. Compared only when a
+  // unit was actually sent: on the `biomarkerId` path it may be omitted, and
+  // then the marker's unit is the stated one. A freshly minted marker adopted
+  // the reading's own unit, so it cannot disagree with it.
+  if (
+    !isQualitative &&
+    unit !== undefined &&
+    !sameLabUnit(unit, biomarker.unit)
+  ) {
+    annotate({
+      action: { name: "labs.create.unit-mismatch" },
+      meta: { biomarkerId: biomarker.id },
+    });
+    return apiError(
+      `The reading is in ${unit}, but ${biomarker.name} is tracked in ${biomarker.unit}.`,
+      422,
+      {
+        errorCode: "labs.unit.mismatch",
+        markerUnit: biomarker.unit,
+        readingUnit: unit,
+      },
+    );
+  }
+
   // Field-by-field assignment — never spread `parsed.data`. The row stamps the
   // resolved name/unit/range as historical truth (so a later catalog edit does
   // not silently rewrite a past reading) AND keeps the FK; reads resolve the
diff --git a/src/components/labs/__tests__/ocr-review-dialog.test.tsx b/src/components/labs/__tests__/ocr-review-dialog.test.tsx
index 852746337..88d61a6e3 100644
--- a/src/components/labs/__tests__/ocr-review-dialog.test.tsx
+++ b/src/components/labs/__tests__/ocr-review-dialog.test.tsx
@@ -12,6 +12,7 @@ import {
   handleFilePickerChange,
   OcrPageCoverageNote,
   OcrReviewDialog,
+  unitSkippedAnalytes,
 } from "../ocr-review-dialog";
 
 const hookState = vi.hoisted(() => ({ extractPending: false }));
@@ -223,3 +224,40 @@ describe("<OcrPageCoverageNote> — a long PDF read from its first pages", () =>
     expect(html).toBe("");
   });
 });
+
+describe("unitSkippedAnalytes", () => {
+  const unit = (analyte: string) => ({
+    analyte,
+    reason: "unit_mismatch" as const,
+  });
+
+  it("is null when nothing was skipped for a unit", () => {
+    expect(unitSkippedAnalytes([])).toBeNull();
+    expect(
+      unitSkippedAnalytes([{ analyte: "LDL", reason: "duplicate" }]),
+    ).toBeNull();
+  });
+
+  it("names the analytes skipped for a unit, ignoring duplicates", () => {
+    expect(
+      unitSkippedAnalytes([
+        { analyte: "LDL", reason: "duplicate" },
+        unit("Glucose"),
+        unit("Creatinine"),
+      ]),
+    ).toBe("Glucose, Creatinine");
+  });
+
+  it("lists three and counts the rest, once each", () => {
+    expect(
+      unitSkippedAnalytes([
+        unit("A"),
+        unit("A"),
+        unit("B"),
+        unit("C"),
+        unit("D"),
+        unit("E"),
+      ]),
+    ).toBe("A, B, C +2");
+  });
+});
diff --git a/src/components/labs/__tests__/ocr-review-validation.test.ts b/src/components/labs/__tests__/ocr-review-validation.test.ts
index 1052269b2..ac47ea7a9 100644
--- a/src/components/labs/__tests__/ocr-review-validation.test.ts
+++ b/src/components/labs/__tests__/ocr-review-validation.test.ts
@@ -1,9 +1,11 @@
 import { describe, expect, it } from "vitest";
 
 import type { OcrReviewRow } from "../ocr-review-types";
+import { seedReviewRows } from "../ocr-review-types";
 import {
   collectRowErrors,
   planSave,
+  readingUnitDiffers,
   toCommitRow,
   validateReviewRow,
 } from "../ocr-review-validation";
@@ -21,6 +23,7 @@ function row(overrides: Partial<OcrReviewRow> = {}): OcrReviewRow {
     takenAt: "2026-06-10",
     confidence: { analyte: 1, value: 1, unit: 1, range: 1 },
     biomarkerMatch: "existing",
+    markerUnit: "mmol/L",
     duplicateOf: null,
     confirmed: true,
     ...overrides,
@@ -140,3 +143,72 @@ describe("planSave", () => {
     }
   });
 });
+
+describe("a reading in another unit than its marker's", () => {
+  it.each([
+    ["the same unit", { unit: "mmol/L" }, false],
+    ["the same unit, spelled differently", { unit: "mmol/l" }, false],
+    ["a different unit", { unit: "mg/dL" }, true],
+    ["a different magnitude of the same substance", { unit: "µmol/L" }, true],
+    ["a new marker (no unit yet)", { unit: "mg/dL", markerUnit: null }, false],
+    ["no unit stated", { unit: "", markerUnit: "mmol/L" }, false],
+    [
+      "a qualitative result",
+      { value: null, valueText: "negative", unit: "mg/dL" },
+      false,
+    ],
+  ] as const)("%s → differs: %s", (_name, overrides, differs) => {
+    expect(readingUnitDiffers(row(overrides))).toBe(differs);
+  });
+
+  it("is a blocking error on the unit field, distinct from a missing unit", () => {
+    expect(validateReviewRow(row({ unit: "mg/dL" }))).toEqual({
+      unitMismatch: true,
+    });
+    expect(validateReviewRow(row({ unit: "" }))).toEqual({ unit: true });
+  });
+
+  it("cannot be turned into a payload, whatever its other fields say", () => {
+    expect(toCommitRow(row({ unit: "mg/dL" }))).toBeNull();
+    expect(toCommitRow(row({ unit: "mmol/l" }))?.unit).toBe("mmol/l");
+  });
+
+  it("blocks the whole save when it is selected, and not when it is not", () => {
+    expect(
+      planSave([row({ key: "a" }), row({ key: "b", unit: "mg/dL" })]),
+    ).toEqual({ kind: "blocked" });
+    const plan = planSave([
+      row({ key: "a" }),
+      row({ key: "b", unit: "mg/dL", confirmed: false }),
+    ]);
+    expect(plan.kind).toBe("ready");
+    expect(
+      collectRowErrors([row({ unit: "mg/dL", confirmed: false })]).size,
+    ).toBe(0);
+  });
+
+  it("starts unselected on the review screen, like a duplicate, so it is a choice", () => {
+    const dto = (unit: string, markerUnit: string | null) => ({
+      analyte: "Glucose",
+      value: 5.6,
+      valueText: null,
+      unit,
+      referenceLow: null,
+      referenceHigh: null,
+      referenceText: null,
+      takenAt: "2026-06-10",
+      confidence: { analyte: 1, value: 1, unit: 1, range: 1 },
+      biomarkerMatch: "existing" as const,
+      markerUnit,
+      duplicateOf: null,
+    });
+    const [same, other, fresh] = seedReviewRows(
+      [dto("mmol/L", "mmol/L"), dto("mmol/L", "mg/dL"), dto("mmol/L", null)],
+      null,
+    );
+    expect(same.confirmed).toBe(true);
+    expect(other.confirmed).toBe(false);
+    expect(other.markerUnit).toBe("mg/dL");
+    expect(fresh.confirmed).toBe(true);
+  });
+});
diff --git a/src/components/labs/__tests__/ocr-row-editor.test.tsx b/src/components/labs/__tests__/ocr-row-editor.test.tsx
index 88e28812a..0079dfafd 100644
--- a/src/components/labs/__tests__/ocr-row-editor.test.tsx
+++ b/src/components/labs/__tests__/ocr-row-editor.test.tsx
@@ -18,6 +18,7 @@ const baseRow: OcrReviewRow = {
   takenAt: null,
   confidence: { analyte: 1, value: 1, unit: 1, range: 1 },
   biomarkerMatch: "existing",
+  markerUnit: "mmol/L",
   duplicateOf: null,
   confirmed: true,
 };
@@ -115,3 +116,42 @@ describe("<OcrRowEditor> layout", () => {
     }
   });
 });
+
+describe("<OcrRowEditor> unit differs from the marker's", () => {
+  const mismatched: OcrReviewRow = {
+    ...baseRow,
+    takenAt: "2026-06-10",
+    unit: "mg/dL",
+    markerUnit: "mmol/L",
+  };
+
+  it("says so before any save attempt, in calm wording", () => {
+    const html = render(undefined, "en", mismatched);
+    expect(html).toContain("Tracked in mmol/L; this report says mg/dL");
+    expect(html).not.toContain('aria-invalid="true"');
+    expect(html).not.toContain('role="alert"');
+  });
+
+  it("marks the unit field and names the marker's unit once a save was refused", () => {
+    const html = render({ unitMismatch: true }, "en", mismatched);
+    const invalid = html.match(/<input[^>]*aria-invalid="true"[^>]*>/g) ?? [];
+    expect(invalid).toHaveLength(1);
+    expect(invalid[0]).toMatch(/id="[^"]*-unit"/);
+    expect(html).toContain("This marker is tracked in mmol/L.");
+  });
+
+  it("shows nothing when the units agree, however they are spelled", () => {
+    const html = render(undefined, "en", { ...mismatched, unit: "mmol/l" });
+    expect(html).not.toContain("Tracked in");
+  });
+
+  it("shows nothing for a new marker, which adopts the row's unit", () => {
+    const html = render(undefined, "en", { ...mismatched, markerUnit: null });
+    expect(html).not.toContain("Tracked in");
+  });
+
+  it("speaks German too", () => {
+    const html = render(undefined, "de", mismatched);
+    expect(html).toContain("Geführt in mmol/L; dieser Befund nennt mg/dL");
+  });
+});
diff --git a/src/components/labs/ocr-review-dialog.tsx b/src/components/labs/ocr-review-dialog.tsx
index f734509a6..a11205fa7 100644
--- a/src/components/labs/ocr-review-dialog.tsx
+++ b/src/components/labs/ocr-review-dialog.tsx
@@ -19,6 +19,7 @@ import { toastWrittenOutcome } from "@/components/outcome/outcome-toast";
 import { Button } from "@/components/ui/button";
 import { ResponsiveSheet } from "@/components/ui/responsive-sheet";
 import { ApiError } from "@/lib/api/api-fetch";
+import type { OcrSkippedRowDto } from "@/lib/validations/labs-ocr";
 import { EncounterSuggestionField } from "@/components/encounters/encounter-suggestion-field";
 import { useTranslations } from "@/lib/i18n/context";
 
@@ -53,6 +54,25 @@ export function handleFilePickerChange(
   if (file) onFilePicked(file);
 }
 
+/**
+ * The analytes the commit skipped for a unit other than their marker's, as one
+ * short phrase for a toast (the first three, then a count), or null when none.
+ */
+export function unitSkippedAnalytes(
+  skipped: readonly OcrSkippedRowDto[],
+): string | null {
+  const names = [
+    ...new Set(
+      skipped
+        .filter((row) => row.reason === "unit_mismatch")
+        .map((row) => row.analyte),
+    ),
+  ];
+  if (names.length === 0) return null;
+  const shown = names.slice(0, 3).join(", ");
+  return names.length > 3 ? `${shown} +${names.length - 3}` : shown;
+}
+
 /**
  * Translate an extract failure into a friendly, error-code-aware message.
  * Refusals are read by `meta.errorCode`: a 403 is a missing consent only when
@@ -241,19 +261,32 @@ export function OcrReviewDialog({
           // used to be green on any 200 and print the count regardless; the
           // server resolves the outcome from the two lists and the dialog
           // renders it.
+          // A row skipped for its unit is the one the person has to act on, so
+          // when there is one the message names it; the plain "already
+          // recorded" wording would be wrong for it.
+          const unitSkipped = unitSkippedAnalytes(result.skipped);
           toastWrittenOutcome(
             result.outcome,
             result.outcome === "success"
               ? t("labs.ocr.savedToast", { count: result.inserted.length })
               : result.outcome === "partial"
-                ? t("labs.ocr.savedPartialToast", {
-                    count: result.inserted.length,
-                    skipped: result.skipped.length,
-                  })
-                : result.outcome === "failed"
-                  ? t("labs.ocr.savedNothingToast", {
+                ? unitSkipped
+                  ? t("labs.ocr.savedPartialUnitToast", {
+                      count: result.inserted.length,
+                      analytes: unitSkipped,
+                    })
+                  : t("labs.ocr.savedPartialToast", {
+                      count: result.inserted.length,
                       skipped: result.skipped.length,
                     })
+                : result.outcome === "failed"
+                  ? unitSkipped
+                    ? t("labs.ocr.savedNothingUnitToast", {
+                        analytes: unitSkipped,
+                      })
+                    : t("labs.ocr.savedNothingToast", {
+                        skipped: result.skipped.length,
+                      })
                   : t("labs.ocr.savedEmptyToast"),
           );
           onCommitted();
diff --git a/src/components/labs/ocr-review-types.ts b/src/components/labs/ocr-review-types.ts
index 825aa77ad..7def77782 100644
--- a/src/components/labs/ocr-review-types.ts
+++ b/src/components/labs/ocr-review-types.ts
@@ -7,6 +7,8 @@
  */
 import type { OcrExtractedRowDto } from "@/lib/validations/labs-ocr";
 
+import { readingUnitDiffers } from "./ocr-review-validation";
+
 export interface OcrReviewRow {
   /** Stable client key for the list (the analyte+index at seed time). */
   key: string;
@@ -22,6 +24,11 @@ export interface OcrReviewRow {
   takenAt: string | null;
   confidence: OcrExtractedRowDto["confidence"];
   biomarkerMatch: "new" | "existing";
+  /**
+   * The unit the matched marker is tracked in; null for a marker that does not
+   * exist yet. A numeric row whose `unit` differs from it cannot be saved.
+   */
+  markerUnit: string | null;
   /** Non-null when the server flagged a likely duplicate of a live reading. */
   duplicateOf: string | null;
   /** Whether the row is selected for saving. Defaults false for duplicates. */
@@ -45,8 +52,18 @@ export function seedReviewRows(
     takenAt: row.takenAt ?? fallbackDate,
     confidence: row.confidence,
     biomarkerMatch: row.biomarkerMatch,
+    markerUnit: row.markerUnit,
     duplicateOf: row.duplicateOf,
-    // A flagged duplicate starts unchecked; everything else starts confirmed.
-    confirmed: row.duplicateOf === null,
+    // A flagged duplicate starts unchecked, and so does a reading in another
+    // unit than its marker's (it carries a badge saying why, and ticking it
+    // marks the unit field until it is fixed); everything else starts
+    // confirmed.
+    confirmed:
+      row.duplicateOf === null &&
+      !readingUnitDiffers({
+        valueText: row.valueText,
+        unit: row.unit,
+        markerUnit: row.markerUnit,
+      }),
   }));
 }
diff --git a/src/components/labs/ocr-review-validation.ts b/src/components/labs/ocr-review-validation.ts
index 8e48e69bc..a2ab7fea2 100644
--- a/src/components/labs/ocr-review-validation.ts
+++ b/src/components/labs/ocr-review-validation.ts
@@ -7,10 +7,13 @@
  * build the payload: it returns null exactly when `validateReviewRow` reports
  * at least one field.
  */
+import { sameLabUnit } from "@/lib/labs/unit-normalise";
+
 import type { OcrReviewRow } from "./ocr-review-types";
 import type { OcrCommitRowInput } from "./use-ocr-extract";
 
-export type OcrRowField = "analyte" | "value" | "valueText" | "unit" | "date";
+export type OcrRowField =
+  "analyte" | "value" | "valueText" | "unit" | "unitMismatch" | "date";
 
 /** The fields of one row that block saving; empty when the row is writable. */
 export type OcrRowErrors = Partial<Record<OcrRowField, true>>;
@@ -26,6 +29,23 @@ function isQualitativeRow(row: OcrReviewRow): boolean {
   return row.valueText !== null && row.valueText !== undefined;
 }
 
+/**
+ * Whether a numeric reading states a unit other than the one its marker is
+ * tracked in. A qualitative reading has no unit, a new marker adopts the row's
+ * own, and a row with no unit yet is a different problem (`unit`), so none of
+ * those count. The same rule the commit applies when it skips such a row.
+ */
+export function readingUnitDiffers(row: {
+  valueText: string | null | undefined;
+  unit: string | null | undefined;
+  markerUnit: string | null | undefined;
+}): boolean {
+  if (row.valueText !== null && row.valueText !== undefined) return false;
+  const stated = (row.unit ?? "").trim();
+  if (!row.markerUnit || !stated) return false;
+  return !sameLabUnit(stated, row.markerUnit);
+}
+
 export function validateReviewRow(row: OcrReviewRow): OcrRowErrors {
   const errors: OcrRowErrors = {};
   if (!row.analyte.trim()) errors.analyte = true;
@@ -36,6 +56,7 @@ export function validateReviewRow(row: OcrReviewRow): OcrRowErrors {
   } else {
     if (row.value === null || !Number.isFinite(row.value)) errors.value = true;
     if (!(row.unit ?? "").trim()) errors.unit = true;
+    if (readingUnitDiffers(row)) errors.unitMismatch = true;
   }
   return errors;
 }
@@ -59,6 +80,7 @@ export function toCommitRow(row: OcrReviewRow): OcrCommitRowInput | null {
   if (row.value === null || !Number.isFinite(row.value)) return null;
   const unit = (row.unit ?? "").trim();
   if (!unit) return null;
+  if (readingUnitDiffers(row)) return null;
   return {
     analyte,
     value: row.value,
diff --git a/src/components/labs/ocr-row-editor.tsx b/src/components/labs/ocr-row-editor.tsx
index 306e5eeb9..b103f09f7 100644
--- a/src/components/labs/ocr-row-editor.tsx
+++ b/src/components/labs/ocr-row-editor.tsx
@@ -32,7 +32,7 @@ import { Label } from "@/components/ui/label";
 import { useTranslations } from "@/lib/i18n/context";
 
 import type { OcrReviewRow } from "./ocr-review-types";
-import type { OcrRowErrors } from "./ocr-review-validation";
+import { readingUnitDiffers, type OcrRowErrors } from "./ocr-review-validation";
 
 /** Below this per-field confidence the field is flagged for a second look. */
 const CONFIDENCE_THRESHOLD = 0.6;
@@ -65,10 +65,21 @@ export function OcrRowEditor({
           "aria-describedby": `${fieldId}-${field}-error`,
         }
       : {};
-  const fieldError = (field: keyof OcrRowErrors, messageKey: string) =>
+  const fieldError = (
+    field: keyof OcrRowErrors,
+    messageKey: string,
+    params?: Record<string, string>,
+  ) =>
     errors?.[field] ? (
-      <FieldError id={`${fieldId}-${field}-error`} message={t(messageKey)} />
+      <FieldError
+        id={`${fieldId}-${field}-error`}
+        message={t(messageKey, params)}
+      />
     ) : null;
+  // Stated before Save, not only after a failed attempt: a reading in another
+  // unit than its marker's cannot be written, and the person should see why
+  // the row is unticked (or will not save) while they are still reading it.
+  const unitDiffers = readingUnitDiffers(row);
 
   const isQualitative = row.valueText !== null && row.valueText !== undefined;
   const lowValueConfidence =
@@ -158,6 +169,17 @@ export function OcrRowEditor({
                   {t("labs.ocr.duplicateWarning")}
                 </Badge>
               ) : null}
+              {unitDiffers ? (
+                <Badge variant="secondary" className="max-w-full">
+                  <Ruler aria-hidden />
+                  <span className="truncate">
+                    {t("labs.ocr.unitDiffersBadge", {
+                      markerUnit: row.markerUnit ?? "",
+                      unit: (row.unit ?? "").trim(),
+                    })}
+                  </span>
+                </Badge>
+              ) : null}
               {lowRowConfidence ? (
                 <Badge variant="secondary" className="text-muted-foreground">
                   <AlertCircle aria-hidden />
@@ -224,9 +246,14 @@ export function OcrRowEditor({
                     id={`${fieldId}-unit`}
                     value={row.unit ?? ""}
                     onChange={(e) => onChange({ ...row, unit: e.target.value })}
-                    {...invalidProps("unit")}
+                    {...invalidProps(
+                      errors?.unitMismatch ? "unitMismatch" : "unit",
+                    )}
                   />
                   {fieldError("unit", "labs.ocr.errorUnit")}
+                  {fieldError("unitMismatch", "labs.ocr.errorUnitMismatch", {
+                    markerUnit: row.markerUnit ?? "",
+                  })}
                 </div>
               </>
             )}
diff --git a/src/lib/labs/__tests__/unit-normalise.test.ts b/src/lib/labs/__tests__/unit-normalise.test.ts
new file mode 100644
index 000000000..5fb4e9d12
--- /dev/null
+++ b/src/lib/labs/__tests__/unit-normalise.test.ts
@@ -0,0 +1,90 @@
+import { describe, expect, it } from "vitest";
+
+import { normaliseLabUnit, sameLabUnit } from "../unit-normalise";
+
+describe("sameLabUnit — spellings of one unit", () => {
+  it.each([
+    ["mmol/L", "mmol/l"],
+    ["mg/dL", "mg/dl"],
+    ["mg/dL", "mg/DL"],
+    ["mg/dL", " mg / dL "],
+    ["mg/dL", "mg/ dL"],
+    ["µg/L", "ug/L"],
+    ["µg/L", "μg/L"],
+    ["µg/L", "mcg/L"],
+    ["µmol/L", "umol/l"],
+    ["µIU/mL", "uIU/mL"],
+    ["x10^3/µL", "x10^3/uL"],
+    ["x10^3/µL", "x10^3/µl"],
+    ["mg/dL", "MG/DL"],
+    ["mmol/L", "MMOL/L"],
+    ["µmol/L", "UMOL/L"],
+    ["ng/mL", "NG/ML"],
+    ["IU/L", "IU/l"],
+    ["IU/L", "IU/L"],
+  ])("%s = %s", (a, b) => {
+    expect(sameLabUnit(a, b)).toBe(true);
+    expect(sameLabUnit(b, a)).toBe(true);
+  });
+});
+
+describe("sameLabUnit — never merges units of different magnitude or kind", () => {
+  it.each([
+    // The point of the module: nothing that changes the number.
+    ["mg/dL", "mmol/L"],
+    ["mg/dL", "mg/L"],
+    ["mg/L", "µg/L"],
+    ["g/dL", "g/L"],
+    ["ng/mL", "ng/dL"],
+    ["µg/L", "mg/L"],
+    ["%", "mmol/mol"],
+    ["mmol/L", "µmol/L"],
+    ["mmol/L", "mol/L"],
+    // Case that carries meaning stays significant.
+    ["mIU/L", "MIU/L"],
+    ["g/L", "G/L"],
+    ["u/L", "U/L"],
+    ["mmol/L", "Mmol/L"],
+    ["pmol/L", "Pmol/L"],
+    // A unit the module does not know compares exactly, so it cannot be guessed.
+    ["mg/dL", "mg/dl."],
+    ["mg/dL", "mg per dL"],
+    ["mg/dL", ""],
+  ])("%s ≠ %s", (a, b) => {
+    expect(sameLabUnit(a, b)).toBe(false);
+    expect(sameLabUnit(b, a)).toBe(false);
+  });
+
+  it("does not turn a unit printed in capitals into one that is not on the list", () => {
+    // `G/L` shouted is the blood-count unit, not grams per litre.
+    expect(normaliseLabUnit("G/L")).toBe("G/L");
+    expect(normaliseLabUnit("MIU/L")).toBe("MIU/L");
+  });
+
+  it("keeps enzyme units apart from the micro prefix", () => {
+    expect(normaliseLabUnit("u/L")).toBe("u/L");
+    expect(normaliseLabUnit("U/L")).toBe("U/L");
+  });
+});
+
+describe("normaliseLabUnit", () => {
+  it("is idempotent", () => {
+    for (const unit of [
+      "ug/l",
+      "MG/DL",
+      "mmol/l",
+      " x10^3 / uL ",
+      "mIU/L",
+      "G/L",
+    ]) {
+      const once = normaliseLabUnit(unit);
+      expect(normaliseLabUnit(once)).toBe(once);
+    }
+  });
+
+  it("names the canonical spelling", () => {
+    expect(normaliseLabUnit("mmol/l")).toBe("mmol/L");
+    expect(normaliseLabUnit("ug/l")).toBe("µg/L");
+    expect(normaliseLabUnit("MG/DL")).toBe("mg/dL");
+  });
+});
diff --git a/src/lib/labs/ocr-extract.ts b/src/lib/labs/ocr-extract.ts
index af2926a09..f217f2657 100644
--- a/src/lib/labs/ocr-extract.ts
+++ b/src/lib/labs/ocr-extract.ts
@@ -131,7 +131,7 @@ export class OcrExtractError extends Error {
  * server-computed biomarker match + duplicate id. `reportDate` fills a row's
  * date when the row itself carries none.
  */
-async function annotateRow(
+export async function annotateRow(
   userId: string,
   row: ExtractedRow,
   reportDate: string | null,
@@ -144,7 +144,7 @@ async function annotateRow(
   // Biomarker match — case-insensitive on the catalog identity, presence-only.
   const existingBiomarker = await prisma.biomarker.findFirst({
     where: { userId, name: { equals: analyte, mode: "insensitive" } },
-    select: { id: true },
+    select: { id: true, unit: true },
   });
 
   // Duplicate check: a live lab_results row with the same analyte (fuzzy,
@@ -198,6 +198,7 @@ async function annotateRow(
     takenAt,
     confidence: row.confidence,
     biomarkerMatch: existingBiomarker ? "existing" : "new",
+    markerUnit: existingBiomarker?.unit ?? null,
     duplicateOf,
   };
 }
diff --git a/src/lib/labs/unit-normalise.ts b/src/lib/labs/unit-normalise.ts
new file mode 100644
index 000000000..b67cc3346
--- /dev/null
+++ b/src/lib/labs/unit-normalise.ts
@@ -0,0 +1,98 @@
+/**
+ * Spelling-level comparison of lab units.
+ *
+ * A reading may only be stored under a marker whose unit it actually is in, so
+ * the write paths compare the unit a reading arrives with against the unit the
+ * marker is tracked in. This module answers one narrow question: are these two
+ * strings the same unit written two ways?
+ *
+ * It resolves spelling, never magnitude. Two spellings are treated as equal
+ * only where no reading of either could mean a different quantity:
+ *
+ *   - whitespace around the slash (`mg / dL`),
+ *   - the micro prefix (`µg/L`, `μg/L` with the Greek mu, `ug/L`, `mcg/L`),
+ *   - the case of the litre in the denominator (`mmol/l`, `mg/dl`, `mg/DL`),
+ *   - a unit printed entirely in capitals (`MG/DL`, `MMOL/L`), but only for the
+ *     short list of units where capitals cannot denote another real unit.
+ *
+ * Case is otherwise significant, on purpose. `mIU/L` and `MIU/L` differ by a
+ * factor of a billion (a lone capital M is mega, a lowercase m is milli), and
+ * `g/L` (grams) is not `G/L` (the 10^9/L of a blood count). A unit this module
+ * does not recognise is compared exactly, so an unfamiliar spelling surfaces as
+ * a mismatch rather than being guessed at.
+ *
+ * Pure and dependency-free.
+ */
+
+/** Units that may be recognised when printed entirely in capitals. */
+const SHOUTED_UNITS: ReadonlyMap<string, string> = new Map(
+  [
+    "mg/dL",
+    "g/dL",
+    "ng/dL",
+    "µg/dL",
+    "mmol/L",
+    "µmol/L",
+    "nmol/L",
+    "pmol/L",
+    "ng/mL",
+    "pg/mL",
+    "µg/L",
+    "mEq/L",
+    "IU/L",
+    "U/L",
+    "fL",
+  ].map((unit) => [unit.toLowerCase(), unit]),
+);
+
+const MICRO_SIGNS = /[µμ]/g;
+/** What may follow a spelled-out micro prefix: g, mol, IU, U, L, kat, Eq. */
+const MICRO_BEFORE = "(?=g|mol|IU|U|L|l|kat|Eq|eq)";
+const MICRO_BEFORE_LOWER = "(?=g|mol|iu|u|l|kat|eq)";
+
+function foldSpacing(raw: string): string {
+  return raw
+    .trim()
+    .replace(/\s+/g, " ")
+    .replace(/\s*\/\s*/g, "/");
+}
+
+function foldMicro(text: string, lookahead: string): string {
+  const unified = text.replace(MICRO_SIGNS, "µ");
+  // A spelled-out prefix counts only at the start of the unit or straight after
+  // a slash, and only in front of a unit that takes it. `u/L` (enzyme units)
+  // and `U/L` therefore stay as they are.
+  const prefix = new RegExp(`(^|/)(?:u|mc)${lookahead}`, "g");
+  return unified.replace(prefix, "$1µ");
+}
+
+/** `/l`, `/dl`, `/ml`, `/µl` in any case → `/L`, `/dL`, `/mL`, `/µL`. */
+function foldLitre(text: string): string {
+  return text.replace(
+    /\/([dmµ]?)l$/i,
+    (_, prefix: string) => `/${prefix.toLowerCase()}L`,
+  );
+}
+
+/**
+ * The comparison form of a lab unit: the same string for every spelling that
+ * means the same thing, and the (trimmed) string itself otherwise.
+ */
+export function normaliseLabUnit(raw: string): string {
+  const spaced = foldSpacing(raw);
+
+  const hasLowercase = /[a-zµμ]/.test(spaced);
+  const hasUppercase = /[A-Z]/.test(spaced);
+  if (hasUppercase && !hasLowercase) {
+    const lowered = foldMicro(spaced.toLowerCase(), MICRO_BEFORE_LOWER);
+    const known = SHOUTED_UNITS.get(lowered);
+    if (known) return known;
+  }
+
+  return foldLitre(foldMicro(spaced, MICRO_BEFORE));
+}
+
+/** Whether two strings name the same lab unit (see the module comment). */
+export function sameLabUnit(a: string, b: string): boolean {
+  return normaliseLabUnit(a) === normaliseLabUnit(b);
+}
diff --git a/src/lib/openapi/error-codes.ts b/src/lib/openapi/error-codes.ts
index 8932c74da..79376a1e6 100644
--- a/src/lib/openapi/error-codes.ts
+++ b/src/lib/openapi/error-codes.ts
@@ -213,6 +213,7 @@ export const ERROR_CODE_CATALOGUE: Readonly<Record<string, readonly string[]>> =
       "labs.ocr.rateLimited",
       "labs.restore.invalid",
       "labs.result.notFound",
+      "labs.unit.mismatch",
       "labs.update.invalid",
       "labs.update.linkedFieldsImmutable",
       "labs.update.numericExpected",
diff --git a/src/lib/openapi/routes/labs.ts b/src/lib/openapi/routes/labs.ts
index 81dd942af..ca0b92046 100644
--- a/src/lib/openapi/routes/labs.ts
+++ b/src/lib/openapi/routes/labs.ts
@@ -162,6 +162,11 @@ export const labsPaths: NonNullable<ZodOpenApiObject["paths"]> = {
           },
         },
         ...stdResponses,
+        "422": {
+          description:
+            "Nothing was written. `meta.errorCode` = `labs.create.invalid` when the body failed validation (every issue under `details.issues`), or `labs.unit.mismatch` when a numeric reading states a `unit` other than the one its marker is tracked in (`meta.markerUnit`, `meta.readingUnit`). A unit is compared only when one is sent; with a `biomarkerId` it may be omitted and the marker's unit is then the stated one. Spelling is forgiven (`mmol/l`, `ug/L` for `µg/L`); a different unit is refused, never relabelled.",
+          content: { "application/json": { schema: errorEnvelope } },
+        },
         ...recordWriteRateLimitResponse,
       },
     },
diff --git a/src/lib/openapi/routes/ocr.ts b/src/lib/openapi/routes/ocr.ts
index f894cf8aa..9a816fcd0 100644
--- a/src/lib/openapi/routes/ocr.ts
+++ b/src/lib/openapi/routes/ocr.ts
@@ -72,6 +72,10 @@ const extractedRow = z
     takenAt: z.string().nullable(),
     confidence: extractConfidence,
     biomarkerMatch: z.enum(["new", "existing"]),
+    markerUnit: z.string().nullable().meta({
+      description:
+        "The unit the matched marker is tracked in, or null when the marker does not exist yet (it will adopt this row's unit). A row whose `unit` differs from it cannot be committed: the commit skips it with `reason: unit_mismatch`, and the review screen marks it before Save.",
+    }),
     duplicateOf: z.string().nullable(),
   })
   .meta({
@@ -138,7 +142,10 @@ const commitResponse = z
     skipped: z.array(
       z.object({
         analyte: z.string(),
-        reason: z.literal("duplicate"),
+        reason: z.enum(["duplicate", "unit_mismatch"]).meta({
+          description:
+            "`duplicate`: the row matched a live reading at commit time. `unit_mismatch`: the row states a unit other than the one its marker is tracked in, so it was not written.",
+        }),
       }),
     ),
     outcome: z.enum(["empty", "failed", "partial", "success"]).meta({
@@ -149,7 +156,7 @@ const commitResponse = z
   .meta({
     id: "OcrCommitResponse",
     description:
-      'The rows written (`source: "OCR"`), those skipped as commit-time duplicates, and the resolved outcome.',
+      'The rows written (`source: "OCR"`), those skipped (commit-time duplicates, or a unit that differs from that of the marker), and the resolved outcome.',
   });
 
 export const ocrPaths: NonNullable<ZodOpenApiObject["paths"]> = {
@@ -223,7 +230,7 @@ export const ocrPaths: NonNullable<ZodOpenApiObject["paths"]> = {
       tags: ["Labs"],
       summary: "Commit confirmed Lab-OCR rows",
       description:
-        'Writes ONLY the rows the human confirmed on the review screen. Each row resolves-or-mints a user-scoped biomarker and creates a `LabResult` with `source: "OCR"`; a row that duplicates a live reading is skipped. Idempotent (Idempotency-Key). Audits as `labs.ocr.commit`. `userId` is narrowed from the session, never a body field.',
+        'Writes ONLY the rows the human confirmed on the review screen. Each row resolves-or-mints a user-scoped biomarker and creates a `LabResult` with `source: "OCR"`; a row that duplicates a live reading, or states a unit other than that of its marker, is skipped (`skipped[].reason`). Idempotent (Idempotency-Key). Audits as `labs.ocr.commit`. `userId` is narrowed from the session, never a body field.',
       requestBody: {
         required: true,
         content: { "application/json": { schema: ocrCommitRequest } },
diff --git a/src/lib/validations/labs-ocr.ts b/src/lib/validations/labs-ocr.ts
index 000a78ab5..59c076f61 100644
--- a/src/lib/validations/labs-ocr.ts
+++ b/src/lib/validations/labs-ocr.ts
@@ -125,6 +125,12 @@ export interface OcrExtractedRowDto {
     range: number;
   };
   biomarkerMatch: "new" | "existing";
+  /**
+   * The unit the matched marker is tracked in, or null for a marker that does
+   * not exist yet (it will adopt this row's unit). Lets the review screen say
+   * so before Save when the scanned unit differs.
+   */
+  markerUnit: string | null;
   duplicateOf: string | null;
 }
 
@@ -231,10 +237,14 @@ export const ocrCommitSchema = z.object({
 
 export type OcrCommitInput = z.infer<typeof ocrCommitSchema>;
 
-/** A row the commit route skipped (a re-checked duplicate at commit time). */
+/**
+ * A row the commit route did not write. `duplicate`: it matched a live reading
+ * at commit time. `unit_mismatch`: it states a unit other than the one its
+ * marker is tracked in.
+ */
 export interface OcrSkippedRowDto {
   analyte: string;
-  reason: "duplicate";
+  reason: "duplicate" | "unit_mismatch";
 }
 
 export interface OcrCommitResponseDto {
diff --git a/tests/integration/backup-round-trip.test.ts b/tests/integration/backup-round-trip.test.ts
index 3b0e4510e..2ae7458d8 100644
--- a/tests/integration/backup-round-trip.test.ts
+++ b/tests/integration/backup-round-trip.test.ts
@@ -126,6 +126,12 @@ vi.mock("@/lib/cache/invalidate", () => ({
 
 const OWNER_ID = "round-trip-owner";
 const AT = (iso: string) => new Date(iso);
+// The restore drops a measurement tombstone older than
+// `TOMBSTONE_RETENTION_DAYS` on purpose, so the fixture's deletion has to stay
+// recent on whatever day the suite runs. A fixed date aged out on 2026-10-03.
+const TOMBSTONE_DELETED_AT = new Date(
+  Math.floor((Date.now() - 10 * 86_400_000) / 1000) * 1000,
+).toISOString();
 const PHQ9_RESPONSES = JSON.stringify({
   items: [2, 1, 2, 2, 1, 1, 2, 1, 2],
   schema: 1,
@@ -351,7 +357,7 @@ async function seedEveryTwoEndedModel(prisma: PrismaClient): Promise<void> {
       source: "APPLE_HEALTH",
       externalId: "round-trip-deleted-sample",
       syncVersion: 4,
-      deletedAt: AT("2026-07-20T09:15:00.000Z"),
+      deletedAt: AT(TOMBSTONE_DELETED_AT),
     },
   });
   // The EVENT row an ECG strip is filed against, and the strip itself. The
@@ -1425,7 +1431,7 @@ describe("every model the plan claims two-ended survives a real restore", () =>
       ),
       "a soft-deleted reading is missing from the disaster-recovery payload",
     ).toMatchObject({
-      deletedAt: "2026-07-20T09:15:00.000Z",
+      deletedAt: TOMBSTONE_DELETED_AT,
       syncVersion: 4,
     });
 
@@ -2342,7 +2348,7 @@ describe("every model the plan claims two-ended survives a real restore", () =>
       },
       "the deleted reading came back, but not as a deletion — a tombstone " +
         "restored without its `deletedAt` is a resurrected measurement",
-    ).toEqual({ deletedAt: "2026-07-20T09:15:00.000Z", syncVersion: 4 });
+    ).toEqual({ deletedAt: TOMBSTONE_DELETED_AT, syncVersion: 4 });
 
     const restoredMeasurement = await prisma.measurement.findFirstOrThrow({
       where: { userId: OWNER_ID, type: "WEIGHT" },
diff --git a/tests/integration/lab-unit-mismatch.test.ts b/tests/integration/lab-unit-mismatch.test.ts
new file mode 100644
index 000000000..0596330b7
--- /dev/null
+++ b/tests/integration/lab-unit-mismatch.test.ts
@@ -0,0 +1,283 @@
+/**
+ * A reading may only be stored under the unit it is actually in.
+ *
+ * Before this, a reading that stated another unit than its marker's (5.6
+ * mmol/L against a marker kept in mg/dL) was written with its value unchanged
+ * and the marker's unit stamped on it, and the range check then judged 5.6
+ * against a mg/dL band. These tests follow that reading through the real
+ * routes and the real database, and assert both what is refused and what is
+ * not: a different spelling of the same unit, a qualitative reading, a
+ * reading with no unit on the structured path and a brand-new marker must all
+ * still write.
+ */
+import { NextRequest } from "next/server";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { cookieJar } from "./mock-next-headers";
+import { getPrismaClient, truncateAllTables } from "./setup";
+
+const USER_ID = "user-lab-unit-mismatch";
+
+vi.mock("next/headers", async () => {
+  const { cookieJar, headerJar } = await import("./mock-next-headers");
+  return {
+    headers: vi.fn(async () => ({
+      get: (name: string) => headerJar.get(name.toLowerCase()) ?? null,
+    })),
+    cookies: vi.fn(async () => ({
+      get: (name: string) => {
+        const value = cookieJar.get(name);
+        return value ? { name, value } : undefined;
+      },
+      set: (name: string, value: string) => {
+        cookieJar.set(name, value);
+      },
+      delete: (name: string) => {
+        cookieJar.delete(name);
+      },
+    })),
+  };
+});
+
+vi.mock("@/lib/db-compat", () => ({
+  ensureDbCompatibility: vi.fn().mockResolvedValue(undefined),
+}));
+
+beforeEach(async () => {
+  await truncateAllTables(getPrismaClient());
+  cookieJar.clear();
+  const prisma = getPrismaClient();
+  await prisma.user.create({
+    data: {
+      id: USER_ID,
+      username: "lab-unit-mismatch",
+      email: "lab-unit-mismatch@example.test",
+    },
+  });
+  const session = await prisma.session.create({
+    data: { userId: USER_ID, expiresAt: new Date(Date.now() + 3_600_000) },
+  });
+  cookieJar.set("healthlog_session", session.id);
+});
+
+const TAKEN_AT = "2026-07-10T12:00:00.000Z";
+
+async function seedGlucose() {
+  return getPrismaClient().biomarker.create({
+    data: {
+      userId: USER_ID,
+      name: "Glucose",
+      unit: "mg/dL",
+      lowerBound: 70,
+      upperBound: 99,
+    },
+  });
+}
+
+function jsonReq(path: string, body: unknown): NextRequest {
+  return new NextRequest(`http://localhost${path}`, {
+    method: "POST",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify(body),
+  });
+}
+
+async function postLab(body: Record<string, unknown>) {
+  const { POST } = await import("@/app/api/labs/route");
+  const res = await POST(jsonReq("/api/labs", { takenAt: TAKEN_AT, ...body }));
+  return { status: res.status, body: await res.json() };
+}
+
+async function commitRows(rows: Record<string, unknown>[]) {
+  const { POST } = await import("@/app/api/labs/ocr/commit/route");
+  const res = await POST(
+    jsonReq("/api/labs/ocr/commit", {
+      rows: rows.map((row) => ({ takenAt: TAKEN_AT, ...row })),
+    }),
+  );
+  return { status: res.status, body: await res.json() };
+}
+
+const labCount = () =>
+  getPrismaClient().labResult.count({ where: { userId: USER_ID } });
+
+describe("POST /api/labs", () => {
+  it("refuses a reading in another unit than its marker's and writes nothing", async () => {
+    await seedGlucose();
+    const { status, body } = await postLab({
+      analyte: "Glucose",
+      value: 5.6,
+      unit: "mmol/L",
+    });
+    expect(status).toBe(422);
+    expect(body.meta).toMatchObject({
+      errorCode: "labs.unit.mismatch",
+      markerUnit: "mg/dL",
+      readingUnit: "mmol/L",
+    });
+    expect(await labCount()).toBe(0);
+  });
+
+  it("refuses it on the structured path too, when a unit is sent", async () => {
+    const marker = await seedGlucose();
+    const { status, body } = await postLab({
+      biomarkerId: marker.id,
+      value: 5.6,
+      unit: "mmol/L",
+    });
+    expect(status).toBe(422);
+    expect(body.meta.errorCode).toBe("labs.unit.mismatch");
+    expect(await labCount()).toBe(0);
+  });
+
+  it("writes a reading on the structured path that sends no unit, under the marker's", async () => {
+    const marker = await seedGlucose();
+    const { status, body } = await postLab({
+      biomarkerId: marker.id,
+      value: 95,
+    });
+    expect(status).toBe(201);
+    expect(body.data).toMatchObject({ value: 95, unit: "mg/dL" });
+  });
+
+  it.each(["mg/dl", "mg/DL", "MG/DL", " mg / dL "])(
+    "writes the same unit spelled %j",
+    async (spelling) => {
+      await seedGlucose();
+      const { status, body } = await postLab({
+        analyte: "Glucose",
+        value: 95,
+        unit: spelling,
+      });
+      expect(status).toBe(201);
+      expect(body.data).toMatchObject({ value: 95, unit: "mg/dL" });
+    },
+  );
+
+  it("does not compare a qualitative reading, which has no unit", async () => {
+    await seedGlucose();
+    const { status } = await postLab({
+      analyte: "Glucose",
+      valueText: "negative",
+      unit: "mmol/L",
+    });
+    expect(status).toBe(201);
+  });
+
+  it("lets a new marker adopt the unit of its first reading", async () => {
+    const { status, body } = await postLab({
+      analyte: "Creatinine",
+      value: 80,
+      unit: "µmol/L",
+    });
+    expect(status).toBe(201);
+    expect(body.data.unit).toBe("µmol/L");
+    // …and then holds the next reading in another unit to it.
+    const next = await postLab({
+      analyte: "Creatinine",
+      value: 0.9,
+      unit: "mg/dL",
+    });
+    expect(next.status).toBe(422);
+    expect(await labCount()).toBe(1);
+  });
+
+  it("never judges the refused value against the marker's range", async () => {
+    // The case from the report: 5.6 mmol/L is an ordinary glucose, and 5.6
+    // mg/dL would read as far below range.
+    await seedGlucose();
+    await postLab({ analyte: "Glucose", value: 5.6, unit: "mmol/L" });
+    const { GET } = await import("@/app/api/labs/route");
+    const res = await GET(
+      new NextRequest("http://localhost/api/labs?analyte=Glucose"),
+    );
+    expect((await res.json()).data.results).toEqual([]);
+  });
+});
+
+describe("POST /api/labs/ocr/commit", () => {
+  it("writes the matching row, skips the one in another unit, and reports a partial outcome", async () => {
+    await seedGlucose();
+    const { status, body } = await commitRows([
+      { analyte: "Glucose", value: 95, unit: "mg/dL" },
+      {
+        analyte: "Glucose",
+        value: 5.6,
+        unit: "mmol/L",
+        takenAt: "2026-08-10T12:00:00.000Z",
+      },
+    ]);
+    expect(status).toBe(200);
+    expect(body.data.inserted).toHaveLength(1);
+    expect(body.data.inserted[0]).toMatchObject({ value: 95, unit: "mg/dL" });
+    expect(body.data.skipped).toEqual([
+      { analyte: "Glucose", reason: "unit_mismatch" },
+    ]);
+    expect(body.data.outcome).toBe("partial");
+    expect(await labCount()).toBe(1);
+  });
+
+  it("reports a failed outcome when every row is in another unit", async () => {
+    await seedGlucose();
+    const { body } = await commitRows([
+      { analyte: "Glucose", value: 5.6, unit: "mmol/L" },
+    ]);
+    expect(body.data.inserted).toHaveLength(0);
+    expect(body.data.outcome).toBe("failed");
+    expect(await labCount()).toBe(0);
+  });
+
+  it("accepts a different spelling of the marker's unit", async () => {
+    await seedGlucose();
+    const { body } = await commitRows([
+      { analyte: "Glucose", value: 95, unit: "MG/DL" },
+    ]);
+    expect(body.data.outcome).toBe("success");
+    expect(body.data.inserted[0].unit).toBe("mg/dL");
+  });
+
+  it("does not compare a qualitative row", async () => {
+    await seedGlucose();
+    const { body } = await commitRows([
+      { analyte: "Glucose", valueText: "negative", unit: "mmol/L" },
+    ]);
+    expect(body.data.outcome).toBe("success");
+  });
+});
+
+describe("the scan review's view of a row", () => {
+  const scanned = (analyte: string, unit: string) => ({
+    analyte,
+    value: 5.6,
+    valueText: null,
+    unit,
+    referenceLow: null,
+    referenceHigh: null,
+    referenceText: null,
+    takenAt: null,
+    confidence: { analyte: 1, value: 1, unit: 1, range: 1 },
+  });
+
+  it("carries the matched marker's unit, so the screen can say so before Save", async () => {
+    await seedGlucose();
+    const { annotateRow } = await import("@/lib/labs/ocr-extract");
+    const known = await annotateRow(
+      USER_ID,
+      scanned("Glucose", "mmol/L"),
+      "2026-07-10",
+      "UTC",
+    );
+    expect(known).toMatchObject({
+      biomarkerMatch: "existing",
+      markerUnit: "mg/dL",
+    });
+
+    const fresh = await annotateRow(
+      USER_ID,
+      scanned("Creatinine", "µmol/L"),
+      "2026-07-10",
+      "UTC",
+    );
+    expect(fresh).toMatchObject({ biomarkerMatch: "new", markerUnit: null });
+  });
+});