diff --git a/packages/loopover-engine/src/index.ts b/packages/loopover-engine/src/index.ts
index 6de93bd8bd..3d31fb9c87 100644
--- a/packages/loopover-engine/src/index.ts
+++ b/packages/loopover-engine/src/index.ts
@@ -594,6 +594,16 @@ export {
   type ResultChangedFile,
   type ResultsPayload,
 } from "./results-payload.js";
+// `LoopConsumptionOutcome` is deliberately its own name, not loop-escalation.ts's `LoopRunOutcome` re-exported
+// below: that one is a loop's HEALTH state (running/converged/abandoned/error), whereas a consumption entry
+// only exists for a run that already stopped and only distinguishes finished work from work cut short.
+export {
+  buildLoopConsumptionEntry,
+  totalConsumptionForTenant,
+  type LoopConsumptionEntry,
+  type LoopConsumptionOutcome,
+  type LoopRunFacts,
+} from "./loop-consumption.js";
 export {
   evaluateTenantQuota,
   type QuotaDimension,
diff --git a/packages/loopover-engine/src/loop-consumption.ts b/packages/loopover-engine/src/loop-consumption.ts
new file mode 100644
index 0000000000..44daf1aefd
--- /dev/null
+++ b/packages/loopover-engine/src/loop-consumption.ts
@@ -0,0 +1,108 @@
+// Per-loop compute consumption ledger entry (pure) — #4792, part of the Rent-a-Loop path #4778.
+//
+// Deterministic and side-effect-free: given ONE finished loop run's already-metered raw facts, it produces the
+// consumption entry a rental ledger records for that run — the tenant it belongs to, the elapsed wall-clock it
+// occupied, and the compute units it burned. It is the upstream counterpart to tenant-quota.ts's
+// evaluateTenantQuota: summing these entries over a period yields exactly that function's TenantUsage
+// (computeUnitsUsed / wallClockMsUsed), so allocation can be reconciled against real consumption.
+//
+// It computes an entry only: it does NOT write to a ledger, meter a running loop, or price anything. Persisting
+// the entry is the separate, blocked-on-#4789/#4790 integration (and per #5669 must target whatever storage
+// abstraction #4940/#5216 lands on, not raw SQLite a second time) — the decision core below has no storage
+// opinion at all, so it stays correct whichever datastore that turns out to be.
+//
+// A KILLED run is a first-class case, not an error path: a loop stopped mid-run still consumed real compute and
+// real wall-clock, so it MUST still bill accurately (#4792's second acceptance criterion). It produces the same
+// shape as a completed run, flagged so a caller can tell a full run from a truncated one without inferring it.
+// Every numeric input is normalized first, so clock skew, a non-finite reading, or an end-before-start timestamp
+// can never make an entry negative, fractional, or NaN — a ledger that bills a tenant for -1 ms, or for NaN
+// units, is worse than one that bills 0. Mirrors tenant-quota.ts's own normalization discipline.
+
+/** How a loop run ended, for billing. Distinct from loop-escalation.ts's LoopRunOutcome, which describes a
+ *  loop's health state (running/converged/abandoned/error); a consumption entry only exists for a run that has
+ *  already stopped, and only cares whether it finished its work or was cut short. */
+export type LoopConsumptionOutcome = "completed" | "killed";
+
+/** One finished loop run's raw, already-metered facts — the input, never mutated. */
+export type LoopRunFacts = {
+  /** The tenant the run is billed to. */
+  tenantId: string;
+  /** The run's own identifier, carried through so an entry is traceable back to its loop. */
+  loopId: string;
+  /** Epoch-ms the loop started occupying compute. */
+  startedAtMs: number;
+  /** Epoch-ms it stopped — its own completion, or the moment it was killed. */
+  endedAtMs: number;
+  outcome: LoopConsumptionOutcome;
+  /** Compute units the run actually burned, as metered by the caller. 0 when nothing metered it — never fabricated. */
+  computeUnitsMetered: number;
+};
+
+/** One rental-ledger row: what a single loop run consumed, ready to sum into a period's TenantUsage. */
+export type LoopConsumptionEntry = {
+  tenantId: string;
+  loopId: string;
+  outcome: LoopConsumptionOutcome;
+  /** Wall-clock ms the run occupied. Never negative, whatever the input timestamps say. */
+  wallClockMs: number;
+  /** Compute units consumed. Never negative/fractional/NaN. */
+  computeUnits: number;
+  /** False for a run killed mid-work — the entry is still accurate, just not a full run. */
+  complete: boolean;
+};
+
+// Normalize any numeric input to a non-negative integer (a non-finite or negative value becomes 0), so no
+// reading can make an entry NaN, fractional, or negative. Same rule as tenant-quota.ts's own inputs.
+function finiteNonNegativeInt(value: number): number {
+  return Number.isFinite(value) ? Math.max(0, Math.floor(value)) : 0;
+}
+
+/**
+ * Build the rental-ledger consumption entry for one finished loop run. Pure: reads only the run it is handed
+ * and returns an entry without mutating or storing anything.
+ *
+ * Elapsed wall-clock is `endedAtMs - startedAtMs`, floored at 0: a non-finite timestamp, or an end that
+ * precedes its start (clock skew, or a kill recorded against a stale start), yields 0 rather than a negative
+ * charge. Compute units are taken as metered and normalized the same way — never inferred from elapsed time,
+ * because a loop that idled and one that saturated a core for the same duration did not consume the same
+ * compute, and guessing would bill a tenant for work that never happened.
+ *
+ * A `killed` run yields the same shape as a `completed` one, with `complete: false`: it really did consume the
+ * compute and time it occupied before being stopped, so it bills exactly like any other run (#4792) — the flag
+ * only records that the work was truncated.
+ */
+export function buildLoopConsumptionEntry(facts: LoopRunFacts): LoopConsumptionEntry {
+  const startedAtMs = finiteNonNegativeInt(facts.startedAtMs);
+  const endedAtMs = finiteNonNegativeInt(facts.endedAtMs);
+
+  return {
+    tenantId: facts.tenantId,
+    loopId: facts.loopId,
+    outcome: facts.outcome,
+    wallClockMs: Math.max(0, endedAtMs - startedAtMs),
+    computeUnits: finiteNonNegativeInt(facts.computeUnitsMetered),
+    complete: facts.outcome === "completed",
+  };
+}
+
+/**
+ * Sum a period's consumption entries into the shape tenant-quota.ts's evaluateTenantQuota reads, so an
+ * allocation can be reconciled against what was really consumed (#4792's "queryable against allocation").
+ * Pure. Entries for other tenants are ignored rather than silently mixed in: billing one tenant for another's
+ * compute is the one mistake a rental ledger must never make, so the caller's filtering is not trusted here.
+ * `activeLoops` is NOT derived — a finished run's entry says nothing about what is running right now, and
+ * inventing a count would make evaluateTenantQuota's concurrency dimension decide on a fabricated number.
+ */
+export function totalConsumptionForTenant(
+  entries: readonly LoopConsumptionEntry[],
+  tenantId: string,
+): { computeUnitsUsed: number; wallClockMsUsed: number } {
+  let computeUnitsUsed = 0;
+  let wallClockMsUsed = 0;
+  for (const entry of entries) {
+    if (entry.tenantId !== tenantId) continue;
+    computeUnitsUsed += finiteNonNegativeInt(entry.computeUnits);
+    wallClockMsUsed += finiteNonNegativeInt(entry.wallClockMs);
+  }
+  return { computeUnitsUsed, wallClockMsUsed };
+}
diff --git a/test/unit/loop-consumption.test.ts b/test/unit/loop-consumption.test.ts
new file mode 100644
index 0000000000..b9235dcf19
--- /dev/null
+++ b/test/unit/loop-consumption.test.ts
@@ -0,0 +1,138 @@
+import { describe, expect, it } from "vitest";
+
+import {
+  buildLoopConsumptionEntry,
+  totalConsumptionForTenant,
+  type LoopConsumptionEntry,
+  type LoopRunFacts,
+} from "../../packages/loopover-engine/src/loop-consumption";
+import { evaluateTenantQuota } from "../../packages/loopover-engine/src/tenant-quota";
+
+const facts = (over: Partial<LoopRunFacts> = {}): LoopRunFacts => ({
+  tenantId: "acme",
+  loopId: "loop-1",
+  startedAtMs: 1_000,
+  endedAtMs: 61_000,
+  outcome: "completed",
+  computeUnitsMetered: 42,
+  ...over,
+});
+
+const entry = (over: Partial<LoopConsumptionEntry> = {}): LoopConsumptionEntry => ({
+  tenantId: "acme",
+  loopId: "loop-1",
+  outcome: "completed",
+  wallClockMs: 60_000,
+  computeUnits: 42,
+  complete: true,
+  ...over,
+});
+
+describe("buildLoopConsumptionEntry (#4792)", () => {
+  // Acceptance criterion 1: a completed loop produces an entry with accurate elapsed compute/time.
+  it("a completed run bills its real elapsed wall-clock and metered compute", () => {
+    expect(buildLoopConsumptionEntry(facts())).toEqual({
+      tenantId: "acme",
+      loopId: "loop-1",
+      outcome: "completed",
+      wallClockMs: 60_000,
+      computeUnits: 42,
+      complete: true,
+    });
+  });
+
+  // Acceptance criterion 2: a killed-mid-run loop ALSO produces an accurate, consistent entry.
+  it("a killed run bills identically for what it consumed, flagged incomplete rather than dropped", () => {
+    const killed = buildLoopConsumptionEntry(facts({ outcome: "killed", endedAtMs: 31_000, computeUnitsMetered: 20 }));
+    expect(killed).toEqual({
+      tenantId: "acme",
+      loopId: "loop-1",
+      outcome: "killed",
+      wallClockMs: 30_000,
+      computeUnits: 20,
+      complete: false,
+    });
+    // Same shape as a completed run — a killed run is billed, never silently zeroed or discarded.
+    expect(Object.keys(killed).sort()).toEqual(Object.keys(buildLoopConsumptionEntry(facts())).sort());
+  });
+
+  it("carries tenant/loop identity through unchanged, so an entry stays traceable to its run", () => {
+    const out = buildLoopConsumptionEntry(facts({ tenantId: "globex", loopId: "loop-9" }));
+    expect(out.tenantId).toBe("globex");
+    expect(out.loopId).toBe("loop-9");
+  });
+
+  describe("never emits a nonsensical charge", () => {
+    it("an end before its start (clock skew, stale start on a kill) floors at 0, never a negative charge", () => {
+      expect(buildLoopConsumptionEntry(facts({ startedAtMs: 61_000, endedAtMs: 1_000 })).wallClockMs).toBe(0);
+    });
+
+    it("non-finite timestamps and compute normalize to 0 rather than NaN", () => {
+      const out = buildLoopConsumptionEntry(
+        facts({ startedAtMs: Number.NaN, endedAtMs: Number.POSITIVE_INFINITY, computeUnitsMetered: Number.NaN }),
+      );
+      expect(out.wallClockMs).toBe(0);
+      expect(out.computeUnits).toBe(0);
+    });
+
+    it("negative and fractional metered compute normalize to a non-negative integer", () => {
+      expect(buildLoopConsumptionEntry(facts({ computeUnitsMetered: -5 })).computeUnits).toBe(0);
+      expect(buildLoopConsumptionEntry(facts({ computeUnitsMetered: 7.9 })).computeUnits).toBe(7);
+    });
+
+    it("negative timestamps normalize before subtracting, so elapsed stays sane", () => {
+      expect(buildLoopConsumptionEntry(facts({ startedAtMs: -1_000, endedAtMs: 5_000 })).wallClockMs).toBe(5_000);
+    });
+
+    it("an unmetered run bills 0 compute — never inferred from elapsed time", () => {
+      const out = buildLoopConsumptionEntry(facts({ computeUnitsMetered: 0 }));
+      expect(out.computeUnits).toBe(0);
+      expect(out.wallClockMs).toBe(60_000); // time still real; compute is not guessed from it
+    });
+  });
+});
+
+describe("totalConsumptionForTenant (#4792)", () => {
+  it("sums a tenant's entries into evaluateTenantQuota's TenantUsage shape", () => {
+    const usage = totalConsumptionForTenant([entry(), entry({ loopId: "loop-2", wallClockMs: 10_000, computeUnits: 8 })], "acme");
+    expect(usage).toEqual({ computeUnitsUsed: 50, wallClockMsUsed: 70_000 });
+  });
+
+  it("INVARIANT: never bills a tenant for another tenant's compute", () => {
+    const usage = totalConsumptionForTenant(
+      [entry(), entry({ tenantId: "globex", loopId: "loop-3", wallClockMs: 999_000, computeUnits: 999 })],
+      "acme",
+    );
+    expect(usage).toEqual({ computeUnitsUsed: 42, wallClockMsUsed: 60_000 });
+  });
+
+  it("an empty period, and a tenant with no entries, total to zero rather than undefined", () => {
+    expect(totalConsumptionForTenant([], "acme")).toEqual({ computeUnitsUsed: 0, wallClockMsUsed: 0 });
+    expect(totalConsumptionForTenant([entry()], "nobody")).toEqual({ computeUnitsUsed: 0, wallClockMsUsed: 0 });
+  });
+
+  it("normalizes a corrupt stored entry instead of propagating NaN into the total", () => {
+    const usage = totalConsumptionForTenant([entry({ computeUnits: Number.NaN, wallClockMs: -5 })], "acme");
+    expect(usage).toEqual({ computeUnitsUsed: 0, wallClockMsUsed: 0 });
+  });
+
+  it("counts a killed run's consumption toward the period like any other", () => {
+    expect(totalConsumptionForTenant([entry({ outcome: "killed", complete: false })], "acme")).toEqual({
+      computeUnitsUsed: 42,
+      wallClockMsUsed: 60_000,
+    });
+  });
+
+  // The reason this primitive exists: its output is exactly what the sibling quota evaluator reads, so a
+  // period's real consumption can be reconciled against the tenant's allocation (#4792 ↔ #4796).
+  it("composes with evaluateTenantQuota: summed consumption drives the allocation decision", () => {
+    const entries = [entry({ computeUnits: 90, wallClockMs: 30_000 })];
+    const usage = totalConsumptionForTenant(entries, "acme");
+    const quota = { computeUnits: 100, wallClockMs: 60_000, maxConcurrentLoops: 2 };
+
+    expect(evaluateTenantQuota({ ...usage, activeLoops: 0 }, quota)).toMatchObject({ allowed: true, exceeded: null });
+
+    const overspent = totalConsumptionForTenant([...entries, entry({ loopId: "loop-2", computeUnits: 10, wallClockMs: 0 })], "acme");
+    expect(evaluateTenantQuota({ ...overspent, activeLoops: 0 }, quota)).toMatchObject({ allowed: false, exceeded: "compute" });
+  });
+});