diff --git a/apps/api/src/ops/replication.service.ts b/apps/api/src/ops/replication.service.ts
index 2aa6e03..946f36b 100644
--- a/apps/api/src/ops/replication.service.ts
+++ b/apps/api/src/ops/replication.service.ts
@@ -4,6 +4,45 @@ import { promisify } from "node:util";
 
 const exec = promisify(execFile);
 
+/**
+ * How far the SQL thread is behind the I/O thread, in source binlog bytes.
+ *
+ * This is a different question from `secondsBehind`, and it answers the case
+ * that lag hides: while the SQL thread grinds through one huge transaction,
+ * `Seconds_Behind_Source` can sit still or even read 0, but the relay backlog
+ * is plainly shrinking (or not). It costs nothing extra — every field here
+ * comes out of the same `SHOW REPLICA STATUS` the panel already runs.
+ *
+ * Both positions are coordinates in the SOURCE's binlog, so they are only
+ * comparable while both threads are working on the SAME source file. When they
+ * are not, the replica is whole files behind and the byte delta is meaningless
+ * (positions restart at ~4 in each new file), so `backlogBytes` and `percent`
+ * are null and `sameFile` says why.
+ */
+export interface ApplyProgress {
+  /** Source binlog file the I/O thread is currently reading. */
+  sourceLogFile: string | null;
+  /** Position in `sourceLogFile` that the I/O thread has fetched up to. */
+  readPos: number;
+  /** Source binlog file the SQL thread is currently applying. */
+  relayLogFile: string | null;
+  /** Position in `relayLogFile` that the SQL thread has applied up to. */
+  execPos: number;
+  /** True while both threads are on the same source file. */
+  sameFile: boolean;
+  /** Fetched-but-not-yet-applied bytes. Null when the files differ. */
+  backlogBytes: number | null;
+  /**
+   * `execPos / readPos` as a percentage, null when the files differ.
+   *
+   * Deliberately never rounded up to 100 while any backlog remains: binlog
+   * positions are large, so a real backlog of a few KB is 99.99% of the file
+   * and would render as "caught up" when it is not. Read `backlogBytes === 0`
+   * for actually caught up.
+   */
+  percent: number | null;
+}
+
 export interface ReplicationStatus {
   /** false when the replica is not configured for this environment at all. */
   configured: boolean;
@@ -17,6 +56,8 @@ export interface ReplicationStatus {
   lastIoError: string | null;
   lastSqlError: string | null;
   sourceHost: string | null;
+  /** Relay-log apply progress. Null when the status output has no positions. */
+  apply: ApplyProgress | null;
   /** Human-readable reason when healthy is false. */
   problem: string | null;
   checkedAt: string;
@@ -57,6 +98,7 @@ export class ReplicationService {
       lastIoError: null,
       lastSqlError: null,
       sourceHost: null,
+      apply: null,
       problem: null,
       checkedAt: now,
     };
@@ -142,12 +184,60 @@ export class ReplicationService {
       lastIoError,
       lastSqlError,
       sourceHost: field("Source_Host"),
+      // Reported, never folded into `healthy`: a non-zero backlog is the normal
+      // state of a working replica for the instant between fetch and apply, so
+      // alarming on it would cry wolf. It is here to answer "is it moving?"
+      // when the lag counter is stuck.
+      apply: applyProgress(raw),
       problem,
       checkedAt: now,
     };
   }
 }
 
+/**
+ * Derive relay-apply progress from `SHOW REPLICA STATUS\G` output.
+ *
+ * Exported for testing. Free in query terms — it re-reads four more fields from
+ * the output the caller already has, with no second round trip to the replica
+ * and no connection to the source.
+ *
+ * @returns null when either position is missing or unparseable, which is what
+ *          happens on a server that is not a replica at all.
+ */
+export function applyProgress(raw: string): ApplyProgress | null {
+  const num = (name: string): number | null => {
+    const v = replicaField(raw, name);
+    if (v === null || v === "NULL") return null;
+    const n = Number(v);
+    return Number.isFinite(n) ? n : null;
+  };
+
+  const readPos = num("Read_Source_Log_Pos");
+  const execPos = num("Exec_Source_Log_Pos");
+  if (readPos === null || execPos === null) return null;
+
+  const sourceLogFile = replicaField(raw, "Source_Log_File");
+  const relayLogFile = replicaField(raw, "Relay_Source_Log_File");
+  const sameFile =
+    sourceLogFile !== null && relayLogFile !== null && sourceLogFile === relayLogFile;
+
+  // Clamped at 0: the SQL thread cannot be ahead of the I/O thread, but the two
+  // fields are sampled independently, so a rotation racing this read can print
+  // a momentarily negative delta. Zero is the honest floor, not a bug.
+  const backlogBytes = sameFile ? Math.max(0, readPos - execPos) : null;
+
+  let percent: number | null = null;
+  if (backlogBytes !== null && readPos > 0) {
+    // Truncate rather than round, and hold short of 100 while bytes remain —
+    // see the doc on ApplyProgress.percent.
+    const p = Math.floor((execPos / readPos) * 10_000) / 100;
+    percent = backlogBytes === 0 ? 100 : Math.min(p, 99.99);
+  }
+
+  return { sourceLogFile, readPos, relayLogFile, execPos, sameFile, backlogBytes, percent };
+}
+
 /**
  * Read one field out of `SHOW REPLICA STATUS\G` output.
  *
diff --git a/apps/api/src/ops/replication.spec.ts b/apps/api/src/ops/replication.spec.ts
index 9d75304..e026df2 100644
--- a/apps/api/src/ops/replication.spec.ts
+++ b/apps/api/src/ops/replication.spec.ts
@@ -1,4 +1,4 @@
-import { replicaField } from "./replication.service";
+import { applyProgress, replicaField } from "./replication.service";
 
 /**
  * Verbatim shape of `SHOW REPLICA STATUS\G` from the live replica, trimmed to
@@ -13,6 +13,10 @@ const HEALTHY = [
   "             Replica_IO_State: Waiting for source to send event",
   "                  Source_Host: 100.103.77.46",
   "                  Source_User: repl",
+  "              Source_Log_File: binlog.000042",
+  "          Read_Source_Log_Pos: 194884231",
+  "        Relay_Source_Log_File: binlog.000042",
+  "          Exec_Source_Log_Pos: 194884231",
   "           Replica_IO_Running: Yes",
   "          Replica_SQL_Running: Yes",
   "              Replicate_Do_DB: ",
@@ -85,3 +89,92 @@ describe("replicaField", () => {
     expect(replicaField(raw, "Last_SQL_Error")).toBe("boom");
   });
 });
+
+/** Builds the four position fields the apply-progress reader cares about. */
+function positions(
+  sourceFile: string,
+  readPos: number | string,
+  relayFile: string,
+  execPos: number | string,
+): string {
+  return [
+    `              Source_Log_File: ${sourceFile}`,
+    `          Read_Source_Log_Pos: ${readPos}`,
+    `        Relay_Source_Log_File: ${relayFile}`,
+    `          Exec_Source_Log_Pos: ${execPos}`,
+  ].join("\n");
+}
+
+describe("applyProgress", () => {
+  it("reports zero backlog and 100% when both positions match", () => {
+    const p = applyProgress(HEALTHY)!;
+    expect(p.sameFile).toBe(true);
+    expect(p.sourceLogFile).toBe("binlog.000042");
+    expect(p.readPos).toBe(194884231);
+    expect(p.execPos).toBe(194884231);
+    expect(p.backlogBytes).toBe(0);
+    expect(p.percent).toBe(100);
+  });
+
+  it("reports the byte delta when the SQL thread trails inside one file", () => {
+    const p = applyProgress(positions("binlog.000042", 2_000_000, "binlog.000042", 1_500_000))!;
+    expect(p.backlogBytes).toBe(500_000);
+    expect(p.percent).toBe(75);
+  });
+
+  /**
+   * The reason the byte delta exists at all. `Seconds_Behind_Source` holds at 0
+   * while the SQL thread is mid-transaction, so the backlog is the only field
+   * that moves — and the only one that says the replica is not caught up.
+   */
+  it("shows a backlog even when the lag counter reads zero", () => {
+    const raw = [
+      "        Seconds_Behind_Source: 0",
+      positions("binlog.000042", 900, "binlog.000042", 400),
+    ].join("\n");
+    expect(replicaField(raw, "Seconds_Behind_Source")).toBe("0");
+    expect(applyProgress(raw)!.backlogBytes).toBe(500);
+  });
+
+  /**
+   * Positions restart near 4 in every new binlog file, so subtracting across
+   * files produces a number that is not a backlog — here it would be a large
+   * NEGATIVE one, which would render as "ahead of the source".
+   */
+  it("refuses to compare positions across different binlog files", () => {
+    const p = applyProgress(positions("binlog.000043", 500, "binlog.000042", 194_000_000))!;
+    expect(p.sameFile).toBe(false);
+    expect(p.backlogBytes).toBeNull();
+    expect(p.percent).toBeNull();
+    expect(p.sourceLogFile).toBe("binlog.000043");
+    expect(p.relayLogFile).toBe("binlog.000042");
+  });
+
+  /**
+   * Percent must not round up to 100 while bytes remain: binlog positions are
+   * large, so a genuine backlog is a rounding error away from the whole file
+   * and would otherwise render as "caught up" on a replica that is not.
+   */
+  it("stops short of 100% while any backlog remains", () => {
+    const p = applyProgress(positions("binlog.000042", 194_884_231, "binlog.000042", 194_884_230))!;
+    expect(p.backlogBytes).toBe(1);
+    expect(p.percent).toBe(99.99);
+  });
+
+  /** Sampled independently, so a rotation racing the read can invert them. */
+  it("clamps a momentarily negative delta to zero", () => {
+    const p = applyProgress(positions("binlog.000042", 400, "binlog.000042", 500))!;
+    expect(p.backlogBytes).toBe(0);
+    expect(p.percent).toBe(100);
+  });
+
+  it("returns null when the server is not a replica and prints no positions", () => {
+    expect(applyProgress("")).toBeNull();
+    expect(applyProgress(BROKEN)).toBeNull();
+  });
+
+  /** A stopped thread makes MySQL print NULL, which is not a position. */
+  it("returns null when a position is NULL", () => {
+    expect(applyProgress(positions("binlog.000042", "NULL", "binlog.000042", 400))).toBeNull();
+  });
+});
diff --git a/apps/web/src/app/operaciones/page.tsx b/apps/web/src/app/operaciones/page.tsx
index 8736ef8..6e02e82 100644
--- a/apps/web/src/app/operaciones/page.tsx
+++ b/apps/web/src/app/operaciones/page.tsx
@@ -23,6 +23,7 @@ import {
 } from "@/lib/api";
 import type { UploadProgress } from "@/lib/api";
 import type {
+  ApplyProgress,
   BackupFile,
   IngestFile,
   OpsJob,
@@ -685,8 +686,58 @@ function ReplicationCard() {
           label="Retraso"
           value={status.secondsBehind === null ? "sin dato" : `${status.secondsBehind} s`}
         />
+        <KV label="Pendiente de aplicar" value={backlogLabel(status.apply)} />
         <KV label="Consultado" value={formatDateTime(status.checkedAt)} />
       </div>
+
+      <ApplyProgressBar apply={status.apply} />
+    </div>
+  );
+}
+
+/**
+ * Bytes the replica has fetched but not yet applied.
+ *
+ * Kept separate from the lag figure because it answers a question the lag
+ * cannot: while the SQL thread chews through one big transaction, the seconds
+ * counter can hold still, but this number visibly falls.
+ */
+function backlogLabel(apply: ApplyProgress | null): string {
+  if (!apply) return "sin dato";
+  // Different source binlog files means the replica is whole files behind and
+  // the byte delta is not a delta at all — positions restart in each new file.
+  if (!apply.sameFile) return "más de un archivo de binlog";
+  if (apply.backlogBytes === 0) return "al día";
+  return formatBytes(apply.backlogBytes);
+}
+
+/**
+ * Applied-vs-fetched bar. Rendered only when both threads are on the same
+ * source binlog file, because that is the only case where the percentage is
+ * arithmetic rather than a guess.
+ */
+function ApplyProgressBar({ apply }: { apply: ApplyProgress | null }) {
+  if (!apply || !apply.sameFile || apply.percent === null) return null;
+
+  return (
+    <div className="upload-progress" style={{ marginTop: 12 }}>
+      <div
+        className="progress-track"
+        role="progressbar"
+        aria-valuenow={apply.percent}
+        aria-valuemin={0}
+        aria-valuemax={100}
+        aria-label="Eventos aplicados de los recibidos"
+      >
+        <div className="progress-fill" style={{ width: `${apply.percent}%` }} />
+      </div>
+      <div className="upload-progress-stats mono">
+        <span>{apply.percent}% aplicado</span>
+        <span>
+          {apply.sourceLogFile} · {apply.execPos.toLocaleString("es-MX")} /{" "}
+          {apply.readPos.toLocaleString("es-MX")}
+        </span>
+      </div>
     </div>
   );
 }
diff --git a/apps/web/src/lib/types.ts b/apps/web/src/lib/types.ts
index c613514..ac793e5 100644
--- a/apps/web/src/lib/types.ts
+++ b/apps/web/src/lib/types.ts
@@ -99,10 +99,32 @@ export interface ReplicationStatus {
   lastIoError: string | null;
   lastSqlError: string | null;
   sourceHost: string | null;
+  apply: ApplyProgress | null;
   problem: string | null;
   checkedAt: string;
 }
 
+/**
+ * Relay-log apply progress, in source binlog bytes.
+ *
+ * Answers "is it moving?" when `secondsBehind` cannot: the lag counter sits
+ * still while the SQL thread works through one large transaction, but the
+ * backlog visibly shrinks. `backlogBytes === 0` is the only reading that means
+ * caught up — `percent` deliberately stops at 99.99 while bytes remain.
+ *
+ * Null fields when the two threads are on different source binlog files
+ * (`sameFile === false`), because the positions are then not comparable.
+ */
+export interface ApplyProgress {
+  sourceLogFile: string | null;
+  readPos: number;
+  relayLogFile: string | null;
+  execPos: number;
+  sameFile: boolean;
+  backlogBytes: number | null;
+  percent: number | null;
+}
+
 /** One of the four legacy Access files expected in the ingest folder. */
 export interface IngestFile {
   name: string;