diff --git a/.changeset/rollup-non-numeric-aggregand.md b/.changeset/rollup-non-numeric-aggregand.md
new file mode 100644
index 0000000000..1673c5baed
--- /dev/null
+++ b/.changeset/rollup-non-numeric-aggregand.md
@@ -0,0 +1,14 @@
+---
+"@objectstack/lint": minor
+---
+
+`os lint` now refuses a `min`/`max` roll-up whose answer cannot be stored in the column it rolls up into — `rollup/non-numeric-aggregand`, at `error`.
+
+`FieldSchema.summaryOperations` admits `min`/`max` over ANY child field, and the engine's `aggregateSummaryValue` returns the driver's answer verbatim (only an empty-set fallback stands between the backend and the stored value). A `summary` field is a member of the spec's `NUMERIC_VALUE_TYPES`, so `valueSchemaFor` answers `z.number().finite()` for it and `driver-sql`'s `createColumn` emits a float column. An ordinary "latest shipment" roll-up — `max` over a `datetime` child field — therefore computes an instant into a column the value contract says holds a finite number, and nothing between author and driver correlated the two. It is refused at authoring time rather than tolerated in a consumer (Prime Directive #12).
+
+- **The accept set** is the numeric class union the boolean class, read from `NUMERIC_VALUE_TYPES` and `BOOLEAN_VALUE_TYPES` rather than typed out. The first is the set that DEFINES the criterion — it is the membership `valueSchemaFor` consults to answer `z.number().finite()`, so a type joining it moves the value contract and this door together. The second is admitted on the authority of the `min(flag)=0` / `max(flag)=1` ruling pinned by the spec's own `AGGREGATION_CASES` (#11152): the answer is a number, so it fits.
+- **It is NOT `isAggregateCompatibleWithFieldType`.** That table deliberately accepts `min`/`max` over the temporal class, because there the answer is returned to a caller and "return[s] a value of the field's OWN type" (#15768). Reusing it here would accept the very declaration this rule exists to refuse. The two questions look alike and are not — "can every backend give one answer" versus "does that answer fit the column this roll-up is stored into" — so this predicate is that table's `min`/`max` row narrowed by exactly the temporal class, and a test pins the disagreement.
+- **Scope.** `min`/`max` only. `count` reads no value off the field; `sum`/`avg` over a non-numeric child is a different shape, whose accept set the aggregate table's own rows already exclude, and is not widened into here.
+- **Silent where it cannot resolve.** An unknown child object, a field the child does not declare, or a field with no declared type produce no finding — the aggregate table's own consumer tier ("a consumer that cannot resolve a field's type must NOT call the predicate with a guess"). A partially-loaded model cannot draw a false refusal.
+
+No export moves: the rule id is an inline literal inside the already-exported `lintDataModel`, beside `rollup/missing-summary`. Measured across this repository, no declaration trips the new refusal — all three `min`/`max` roll-ups aggregate a `number` child field — so this adds a door rather than migrating anything.
diff --git a/packages/lint/src/data-model-rules.summary-rollup.test.ts b/packages/lint/src/data-model-rules.summary-rollup.test.ts
new file mode 100644
index 0000000000..aec6893e22
--- /dev/null
+++ b/packages/lint/src/data-model-rules.summary-rollup.test.ts
@@ -0,0 +1,186 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+//
+// `rollup/non-numeric-aggregand` — the roll-up door.
+//
+// `FieldSchema.summaryOperations` admits `min`/`max` over ANY child field, and
+// `aggregateSummaryValue` (objectql) returns the driver's answer verbatim with
+// only an empty-set fallback. A `summary` field is a member of the spec's
+// `NUMERIC_VALUE_TYPES`, so `valueSchemaFor` answers `z.number().finite()` for
+// it and `driver-sql`'s `createColumn` emits `table.float(name)`. An ordinary
+// "latest shipment" roll-up — `max` over a `datetime` child field — therefore
+// computes an INSTANT into a column the value contract says holds a finite
+// number, and nothing between author and driver correlated the two.
+//
+// The load-bearing test in this file is `discriminates from the analytics
+// table`. The refusal CANNOT be `isAggregateCompatibleWithFieldType`: that
+// table deliberately accepts `min`/`max` over the temporal class, because
+// there the answer is RETURNED to a caller rather than stored, and it "return[s]
+// a value of the field's OWN type (#15768)". Reusing it here would accept the
+// very declaration this rule exists to refuse, and the rule would be green
+// because it never fires. The two questions look alike and are not:
+// "can every backend give one answer" vs "does that answer fit the column this
+// roll-up is stored into".
+import { describe, expect, it } from 'vitest';
+import {
+  BOOLEAN_VALUE_TYPES,
+  FieldType,
+  NUMERIC_VALUE_TYPES,
+  isAggregateCompatibleWithFieldType,
+} from '@objectstack/spec/data';
+
+import { lintDataModel } from './data-model-rules.js';
+
+const RULE = 'rollup/non-numeric-aggregand';
+
+/** A parent rolling up `fn(child.<field>)`, and a child declaring that field as `childType`. */
+const model = (childType: string | undefined, fn = 'max', overrides: Record<string, unknown> = {}) => [
+  {
+    name: 'invoice',
+    fields: {
+      name: { type: 'text' },
+      rolled_up: {
+        type: 'summary',
+        summaryOperations: { object: 'invoice_line', field: 'shipped_at', function: fn, ...overrides },
+      },
+    },
+  },
+  {
+    name: 'invoice_line',
+    fields: {
+      name: { type: 'text' },
+      invoice: { type: 'master_detail', reference: 'invoice', required: true, deleteBehavior: 'cascade' },
+      ...(childType ? { shipped_at: { type: childType } } : {}),
+    },
+  },
+];
+
+const findings = (objects: unknown[]) => lintDataModel(objects as any[]).filter((i) => i.rule === RULE);
+
+describe('rollup/non-numeric-aggregand — refuses a min/max roll-up whose answer cannot fit the column', () => {
+  it('refuses the card\'s own example: max over a datetime child field', () => {
+    const found = findings(model('datetime'));
+    expect(found).toHaveLength(1);
+    const issue = found[0];
+    expect(issue.severity).toBe('error');
+    expect(issue.rule).toBe(RULE);
+    // The message must let an author act without opening the source: WHICH
+    // child object, WHICH field, its DECLARED type, and why the answer cannot
+    // be stored.
+    expect(issue.message).toContain('invoice_line');
+    expect(issue.message).toContain('shipped_at');
+    expect(issue.message).toContain('datetime');
+    expect(issue.message).toContain('finite number');
+    expect(issue.path).toBe('objects[0].fields.rolled_up.summaryOperations.field');
+    expect(issue.fix).toBeTruthy();
+  });
+
+  it('refuses every member of the temporal class, under both min and max', () => {
+    for (const childType of ['date', 'datetime', 'time']) {
+      for (const fn of ['min', 'max']) {
+        expect(findings(model(childType, fn)), `${fn}(${childType})`).toHaveLength(1);
+      }
+    }
+  });
+
+  it('refuses a text / option / reference child field too', () => {
+    for (const childType of ['text', 'select', 'lookup', 'json', 'autonumber']) {
+      expect(findings(model(childType)), childType).toHaveLength(1);
+    }
+  });
+});
+
+describe('rollup/non-numeric-aggregand — accepts the classes whose answer IS a number', () => {
+  it('accepts the numeric class', () => {
+    for (const childType of NUMERIC_VALUE_TYPES) {
+      expect(findings(model(childType)), childType).toEqual([]);
+    }
+  });
+
+  // #11152 (maintainer ruling, 2026-08-28), pinned by the spec's own
+  // `AGGREGATION_CASES`: `min(flag)=0` / `max(flag)=1` on six backends. A
+  // careless predicate — "numeric only" — breaks exactly this leg.
+  it('accepts the boolean class', () => {
+    for (const childType of BOOLEAN_VALUE_TYPES) {
+      expect(findings(model(childType)), childType).toEqual([]);
+    }
+  });
+
+  it('accepts exactly the numeric ∪ boolean classes over every declared FieldType', () => {
+    const accepted = new Set([...NUMERIC_VALUE_TYPES, ...BOOLEAN_VALUE_TYPES]);
+    const refused = FieldType.options.filter((t) => findings(model(t)).length > 0);
+    expect(refused.sort()).toEqual(FieldType.options.filter((t) => !accepted.has(t)).sort());
+    // A floor, so a future edit that stops the rule firing at all cannot make
+    // the equality above vacuously true.
+    expect(refused.length).toBeGreaterThan(10);
+  });
+});
+
+describe('rollup/non-numeric-aggregand — discriminates from the analytics table', () => {
+  // ⭐ The assertion that stops someone "simplifying" this rule back into
+  // `isAggregateCompatibleWithFieldType`. If this ever fails because the table
+  // started refusing the temporal class, that is a spec change to read, not a
+  // test to update: the two predicates would then answer the same question.
+  it('the analytics table ACCEPTS the temporal pair this rule refuses', () => {
+    for (const childType of ['date', 'datetime', 'time']) {
+      for (const fn of ['min', 'max']) {
+        expect(isAggregateCompatibleWithFieldType(fn, childType), `${fn}(${childType})`).toBe(true);
+        expect(findings(model(childType, fn)), `${fn}(${childType})`).toHaveLength(1);
+      }
+    }
+  });
+
+  it('the two agree everywhere else on min/max — the difference is exactly the temporal class', () => {
+    const TEMPORAL = new Set(['date', 'datetime', 'time']);
+    for (const childType of FieldType.options) {
+      const tableAccepts = isAggregateCompatibleWithFieldType('max', childType);
+      const doorAccepts = findings(model(childType)).length === 0;
+      if (TEMPORAL.has(childType)) continue;
+      expect(doorAccepts, `max(${childType})`).toBe(tableAccepts);
+    }
+  });
+});
+
+describe('rollup/non-numeric-aggregand — stays silent where it cannot resolve, and outside its scope', () => {
+  // "A consumer that cannot resolve a field's type must NOT call the predicate
+  // with a guess" — `aggregate-field-type-compatibility.ts`. A refusal fired on
+  // a partially-loaded model would redden an app over metadata never seen.
+  it('is silent when the child object is not in the pass\'s object set', () => {
+    const objects = model('datetime').slice(0, 1); // parent only — no `invoice_line`
+    expect(findings(objects)).toEqual([]);
+  });
+
+  it('is silent when the named field is not declared on the child', () => {
+    expect(findings(model(undefined))).toEqual([]);
+  });
+
+  it('is silent when the child field declares no type', () => {
+    const objects = model('datetime') as any[];
+    objects[1].fields.shipped_at = { label: 'Shipped At' };
+    expect(findings(objects)).toEqual([]);
+  });
+
+  it('is silent when summaryOperations names no object or no field', () => {
+    expect(findings(model('datetime', 'max', { object: undefined }))).toEqual([]);
+    expect(findings(model('datetime', 'max', { field: undefined }))).toEqual([]);
+    expect(findings(model('datetime', 'max', { object: '' }))).toEqual([]);
+  });
+
+  it('is scoped to min/max — count reads no value off the field', () => {
+    expect(findings(model('datetime', 'count'))).toEqual([]);
+  });
+
+  // `sum` / `avg` over a non-numeric child is a different shape, whose accept
+  // set the analytics table's own rows already exclude, and is deliberately not
+  // this rule's. (Nothing in this tree consults that table — an open gap
+  // reported with this change, not one this rule silently absorbs.)
+  it('does not fire for sum or avg', () => {
+    expect(findings(model('datetime', 'sum'))).toEqual([]);
+    expect(findings(model('datetime', 'avg'))).toEqual([]);
+  });
+
+  it('is silent for a field that is not a summary at all', () => {
+    const objects = model('datetime') as any[];
+    objects[0].fields.rolled_up.type = 'number';
+    expect(findings(objects)).toEqual([]);
+  });
+});
diff --git a/packages/lint/src/data-model-rules.ts b/packages/lint/src/data-model-rules.ts
index 4b3c419928..ac2309969c 100644
--- a/packages/lint/src/data-model-rules.ts
+++ b/packages/lint/src/data-model-rules.ts
@@ -16,6 +16,8 @@
  * schema-valid AND lint-clean here.
  */
 
+import { BOOLEAN_VALUE_TYPES, NUMERIC_VALUE_TYPES } from '@objectstack/spec/data';
+
 export type Severity = 'error' | 'warning' | 'suggestion';
 
 export interface LintIssue {
@@ -105,6 +107,57 @@ const NUMERIC_TYPES = new Set([
   'number', 'currency', 'integer', 'decimal', 'percent', 'float', 'double',
 ]);
 const OPTION_FIELD_TYPES = new Set(['select', 'multiselect', 'radio', 'enum']);
+
+/**
+ * Does a `min` / `max` roll-up over a child field of `childFieldType` produce
+ * an answer that fits the column a `summary` field is stored in?
+ *
+ * This is the roll-up door's OWN predicate, and it is deliberately NOT
+ * `isAggregateCompatibleWithFieldType`
+ * (`packages/spec/src/data/aggregate-field-type-compatibility.ts`). The two
+ * answer different questions about the same pair:
+ *
+ *   - That table asks **"can every backend give one answer"** — and for
+ *     `min` / `max` it ACCEPTS the temporal class on purpose, because there
+ *     they "return a value of the field's OWN type (#15768)". A `DatasetMeasure`
+ *     hands that answer straight to the caller, so a temporal answer is fine.
+ *   - This rule asks **"does that answer fit the column this roll-up is
+ *     STORED into"**. A roll-up's answer is not returned, it is persisted in
+ *     the `summary` field, whose value contract is `z.number().finite()`
+ *     (`valueSchemaFor`, ADR-0104 D1) because `summary` is a member of
+ *     `NUMERIC_VALUE_TYPES`.
+ *
+ * So this predicate is the analytics table's `min` / `max` row NARROWED by
+ * exactly the temporal class. Reusing that table here would accept
+ * `max(child.shipped_at)` — the very declaration this rule exists to refuse —
+ * and the gate would be green because it never fires. ⛔ Do not "simplify"
+ * the two into one call; `data-model-rules.summary-rollup.test.ts` pins the
+ * disagreement.
+ *
+ * Membership is READ from the spec's own value classes rather than typed out
+ * here, and from these two specifically:
+ *
+ *   - `NUMERIC_VALUE_TYPES` is the set that DEFINES the acceptance criterion —
+ *     it is the very membership `valueSchemaFor` consults to answer
+ *     `z.number().finite()`. A type joining that class changes the `summary`
+ *     value contract and this door with it, in one edit.
+ *   - `BOOLEAN_VALUE_TYPES` is admitted on the authority of maintainer ruling
+ *     #11152: booleans aggregate as NUMBERS on every backend with no
+ *     per-aggregate exception, pinned by the spec's own `AGGREGATION_CASES`
+ *     (`min(flag)=0`, `max(flag)=1`, enrolled on six backends) and implemented
+ *     by `driver-sql`'s `int` cast on Postgres (#11635). The answer is a
+ *     number, so it fits.
+ *
+ * ⛔ Deliberately NOT read: `NON_TEXT_STORED_VALUE_TYPES`, whose membership is
+ * these same two classes TODAY. It is defined by a third question — "is the
+ * stored value never text" (#14079) — and excludes the temporal class for a
+ * DIALECT reason, not for this one. Composing the union here states why each
+ * half is in, so a future member of that set cannot widen this door as a side
+ * effect.
+ */
+function summaryRollupAnswerFitsColumn(childFieldType: string): boolean {
+  return NUMERIC_VALUE_TYPES.has(childFieldType) || BOOLEAN_VALUE_TYPES.has(childFieldType);
+}
 /**
  * Field names that give an object a title FACE, for R9
  * (`object/missing-name-field`).
@@ -491,6 +544,12 @@ export function lintDataModel(objects: any[]): LintIssue[] {
   ];
   if (!Array.isArray(objects) || objects.length === 0) return issues;
 
+  // Index: object name → the object, for resolving a roll-up's child object.
+  const objectsByName: Record<string, any> = {};
+  for (const o of objects) {
+    if (o?.name && !(o.name in objectsByName)) objectsByName[o.name] = o;
+  }
+
   // Index: parent object name → child relationships pointing at it.
   const childrenByParent: Record<string, Array<{ child: any; fieldName: string; def: any }>> = {};
   for (const child of objects) {
@@ -574,6 +633,71 @@ export function lintDataModel(objects: any[]): LintIssue[] {
         }
       }
 
+      // R13 — a `min`/`max` roll-up must aggregate a child field whose ANSWER
+      // fits the column the roll-up is stored into.
+      //
+      // `FieldSchema.summaryOperations` admits `min`/`max` over ANY child
+      // field, and `aggregateSummaryValue` (objectql) returns the driver's
+      // answer verbatim — only an empty-set fallback stands between the
+      // backend and the stored value. So an ordinary "latest shipment"
+      // roll-up, `max` over a `datetime` child field, computes an INSTANT into
+      // a field whose value contract says finite number. Nothing between
+      // author and driver correlated the two, so it is refused here, at
+      // authoring time, at `error` (Prime Directive #12: reject at authoring,
+      // never tolerate in a consumer).
+      //
+      // Scoped to `min`/`max` deliberately: `count` ignores the field
+      // entirely, and `sum`/`avg` over a non-numeric child is a different
+      // shape, whose accept set the analytics table's own rows already
+      // EXCLUDE. Measured on this tree, nothing consults that table on the
+      // roll-up door (or anywhere else), so that is an open gap, not a
+      // refusal this rule may lean on — and widening here to cover it would
+      // be a second account of a pair the table already rules on.
+      if (type === 'summary') {
+        const ops = def.summaryOperations;
+        const fn = ops?.function;
+        const childName = ops?.object;
+        const childFieldName = ops?.field;
+        if (
+          (fn === 'min' || fn === 'max') &&
+          typeof childName === 'string' && childName !== '' &&
+          typeof childFieldName === 'string' && childFieldName !== ''
+        ) {
+          // SILENCE, never a guess, on anything this pass cannot resolve: a
+          // child object contributed by another package, a partially-loaded
+          // stack, or a field name that resolves to no declaration. The spec's
+          // own aggregate table states the tier — "a consumer that cannot
+          // resolve a field's type … must NOT call the predicate with a guess;
+          // 'cannot answer, do not block' is the consumer's tier". A refusal
+          // fired on an unresolvable model would redden an app for metadata
+          // this pass simply never saw.
+          const child = objectsByName[childName];
+          const childField = child
+            ? fieldEntries(child.fields).find((f) => f.name === childFieldName)
+            : undefined;
+          const childType = childField?.def?.type;
+          if (typeof childType === 'string' && !summaryRollupAnswerFitsColumn(childType)) {
+            issues.push({
+              severity: 'error',
+              rule: 'rollup/non-numeric-aggregand',
+              message:
+                `summary field "${obj.name}.${fieldName}" rolls up ` +
+                `${fn}(${childName}.${childFieldName}), but "${childName}.${childFieldName}" is ` +
+                `a ${childType} field — ${fn} answers with a value of the CHILD field's own type, ` +
+                `while a summary field's value contract is a finite number (it is a member of the ` +
+                `spec's NUMERIC_VALUE_TYPES class), so the answer does not fit the column the ` +
+                `roll-up is stored in`,
+              path: `${fieldPath}.summaryOperations.field`,
+              fix:
+                `Aggregate a numeric or boolean child field instead (min/max over those answer ` +
+                `with a number), or use function: 'count' — which reads no value off the field. ` +
+                `To carry a ${childType} on "${obj.name}", declare a ${childType} field and ` +
+                `maintain it from a flow; a roll-up cannot store one.`,
+            });
+          }
+        }
+      }
+
       if (!RELATIONSHIP_TYPES.has(type)) continue;
       const parent = refOf(def);