diff --git a/tests/test_clean_input_selection.py b/tests/test_clean_input_selection.py
index b527a3a..04319b5 100644
--- a/tests/test_clean_input_selection.py
+++ b/tests/test_clean_input_selection.py
@@ -47,7 +47,7 @@ def _fake_run_sql(input_files, sql_query, output_path, read_cfg=None, **_kwargs)
         seen["output_path"] = output_path
         output_path.parent.mkdir(parents=True, exist_ok=True)
         output_path.write_bytes(b"PAR1")
-        return ("strict", {"delim": ";", "decimal": ",", "encoding": "utf-8"})
+        return ("strict", {"delim": ";", "decimal": ",", "encoding": "utf-8"}, 42)
 
     monkeypatch.setattr("toolkit.clean.run._run_sql", _fake_run_sql)
     run_clean("demo", 2024, str(tmp_path), clean_cfg, logger or _NoopLogger())
diff --git a/tests/test_run_context.py b/tests/test_run_context.py
index 15dbc87..d753f97 100644
--- a/tests/test_run_context.py
+++ b/tests/test_run_context.py
@@ -3,6 +3,8 @@
 import json
 from pathlib import Path
 
+import time
+
 from toolkit.core.run_context import RunContext, get_run_dir, read_run_record
 
 
@@ -80,6 +82,67 @@ def test_read_run_record_marks_absolute_paths_outside_root_as_non_portable(tmp_p
     assert record["_portability"]["warnings"] == ["/outside/root/file.csv"]
 
 
+def test_layer_metrics_default_to_null(tmp_path: Path) -> None:
+    ctx = RunContext("ds", 2030, root=str(tmp_path))
+    stored = _read_context(ctx.path)
+    for layer in ("raw", "clean", "mart"):
+        metrics = stored["layers"][layer]["metrics"]
+        assert metrics["output_rows"] is None
+        assert metrics["output_bytes"] is None
+        assert metrics["tables_count"] is None
+
+
+def test_set_layer_metrics_persists(tmp_path: Path) -> None:
+    ctx = RunContext("ds", 2030, root=str(tmp_path))
+    ctx.set_layer_metrics("clean", output_rows=1000, output_bytes=204800)
+    stored = _read_context(ctx.path)
+    m = stored["layers"]["clean"]["metrics"]
+    assert m["output_rows"] == 1000
+    assert m["output_bytes"] == 204800
+    assert m["tables_count"] is None
+
+
+def test_set_layer_metrics_mart_with_tables_count(tmp_path: Path) -> None:
+    ctx = RunContext("ds", 2030, root=str(tmp_path))
+    ctx.set_layer_metrics("mart", output_rows=5000, output_bytes=409600, tables_count=3)
+    stored = _read_context(ctx.path)
+    m = stored["layers"]["mart"]["metrics"]
+    assert m["output_rows"] == 5000
+    assert m["output_bytes"] == 409600
+    assert m["tables_count"] == 3
+
+
+def test_duration_seconds_computed_after_complete(tmp_path: Path) -> None:
+    ctx = RunContext("ds", 2030, root=str(tmp_path))
+    ctx.start_layer("raw")
+    time.sleep(0.05)
+    ctx.complete_layer("raw")
+    ctx.complete_run()
+    stored = _read_context(ctx.path)
+    assert stored["layers"]["raw"]["duration_seconds"] is not None
+    assert stored["layers"]["raw"]["duration_seconds"] >= 0
+    assert stored["duration_seconds"] is not None
+    assert stored["duration_seconds"] >= 0
+
+
+def test_duration_seconds_null_while_running(tmp_path: Path) -> None:
+    ctx = RunContext("ds", 2030, root=str(tmp_path))
+    stored = _read_context(ctx.path)
+    assert stored["duration_seconds"] is None
+    assert stored["layers"]["raw"]["duration_seconds"] is None
+
+
+def test_metrics_survive_json_round_trip(tmp_path: Path) -> None:
+    ctx = RunContext("ds", 2030, root=str(tmp_path))
+    ctx.set_layer_metrics("raw", output_bytes=8192)
+    ctx.set_layer_metrics("clean", output_rows=500, output_bytes=16384)
+    ctx.set_layer_metrics("mart", output_rows=200, output_bytes=4096, tables_count=2)
+    record = read_run_record(get_run_dir(tmp_path, "ds", 2030), ctx.run_id)
+    assert record["layers"]["raw"]["metrics"]["output_bytes"] == 8192
+    assert record["layers"]["clean"]["metrics"]["output_rows"] == 500
+    assert record["layers"]["mart"]["metrics"]["tables_count"] == 2
+
+
 def test_read_run_record_does_not_treat_error_message_as_path(tmp_path: Path) -> None:
     run_dir = get_run_dir(tmp_path, "demo_ds", 2022)
     run_dir.mkdir(parents=True, exist_ok=True)
diff --git a/toolkit/clean/run.py b/toolkit/clean/run.py
index 5ebbaf4..72e5028 100644
--- a/toolkit/clean/run.py
+++ b/toolkit/clean/run.py
@@ -188,7 +188,7 @@ def _run_sql(
     read_cfg: dict[str, Any] | None = None,
     read_mode: str = "fallback",
     logger=None,
-) -> tuple[str, dict[str, Any]]:
+) -> tuple[str, dict[str, Any], int]:
     con = duckdb.connect(":memory:")
     try:
         read_info = read_raw_to_relation(con, input_files, read_cfg, read_mode, logger)
@@ -197,7 +197,8 @@ def _run_sql(
         con.execute(
             f"COPY clean_out TO '{sql_path(output_path)}' (FORMAT PARQUET);"
         )
-        return read_info.source, read_info.params_used
+        row_count: int = con.execute("SELECT count(*) FROM clean_out").fetchone()[0]
+        return read_info.source, read_info.params_used, row_count
     finally:
         con.close()
 
@@ -253,7 +254,7 @@ def run_clean(
     )
 
     output_path = out_dir / f"{dataset}_{year}_clean.parquet"
-    read_source_used, read_params_used = _run_sql(
+    read_source_used, read_params_used, output_rows = _run_sql(
         input_files,
         sql,
         output_path,
@@ -261,6 +262,7 @@ def run_clean(
         read_mode=read_mode,
         logger=logger,
     )
+    output_bytes: int | None = output_path.stat().st_size if output_path.exists() else None
 
     outputs = [file_record(output_path)]
     metadata_payload = _clean_metadata_payload(
@@ -294,3 +296,4 @@ def run_clean(
         warnings_count=None,
     )
     logger.info(f"CLEAN -> {output_path}")
+    return {"output_rows": output_rows, "output_bytes": output_bytes}
diff --git a/toolkit/cli/cmd_run.py b/toolkit/cli/cmd_run.py
index 7f76bca..9255996 100644
--- a/toolkit/cli/cmd_run.py
+++ b/toolkit/cli/cmd_run.py
@@ -217,8 +217,10 @@ def _execute_layer(layer_name: str, target, *args, **kwargs) -> None:
         layer_logger = bind_logger(base_logger, layer=layer_name)
         context.start_layer(layer_name)
         try:
-            target(*args, logger=layer_logger, **kwargs)
+            metrics = target(*args, logger=layer_logger, **kwargs)
             context.complete_layer(layer_name)
+            if isinstance(metrics, dict):
+                context.set_layer_metrics(layer_name, **metrics)
 
             summary = _validation_runner(layer_name)(cfg, year, layer_logger)
             context.set_validation(layer_name, summary)
diff --git a/toolkit/core/run_context.py b/toolkit/core/run_context.py
index 88c2dbd..c12dc30 100644
--- a/toolkit/core/run_context.py
+++ b/toolkit/core/run_context.py
@@ -5,7 +5,7 @@
 import uuid
 from datetime import datetime, timezone
 from pathlib import Path, PurePath, PurePosixPath, PureWindowsPath
-from typing import Any, Dict
+from typing import Any, Dict, Optional
 
 from toolkit.core.paths import to_root_relative
 
@@ -22,6 +22,21 @@ def _now_iso() -> str:
     return datetime.now(timezone.utc).isoformat()
 
 
+def _duration_seconds(started: Optional[str], finished: Optional[str]) -> Optional[float]:
+    if started is None or finished is None:
+        return None
+    try:
+        s = datetime.fromisoformat(started)
+        f = datetime.fromisoformat(finished)
+        return round((f - s).total_seconds(), 3)
+    except Exception:
+        return None
+
+
+def _empty_layer_metrics() -> Dict[str, Any]:
+    return {"output_rows": None, "output_bytes": None, "tables_count": None}
+
+
 def get_run_dir(root: Path, dataset: str, year: int) -> Path:
     return root / "data" / "_runs" / dataset / str(year)
 
@@ -155,7 +170,7 @@ def __init__(
         self.finished_at: str | None = None
         self.status = "RUNNING"
         self.layers = {
-            layer: {"status": "PENDING", "started_at": None, "finished_at": None}
+            layer: {"status": "PENDING", "started_at": None, "finished_at": None, "metrics": _empty_layer_metrics()}
             for layer in _LAYER_NAMES
         }
         self.validations = {layer: {} for layer in _LAYER_NAMES}
@@ -169,14 +184,21 @@ def path(self) -> Path:
         return self._path
 
     def to_dict(self) -> Dict[str, Any]:
+        layers_out = {}
+        for layer, info in self.layers.items():
+            layers_out[layer] = {
+                **info,
+                "duration_seconds": _duration_seconds(info.get("started_at"), info.get("finished_at")),
+            }
         return {
             "dataset": self.dataset,
             "year": self.year,
             "run_id": self.run_id,
             "started_at": self.started_at,
             "finished_at": self.finished_at,
+            "duration_seconds": _duration_seconds(self.started_at, self.finished_at),
             "status": self.status,
-            "layers": self.layers,
+            "layers": layers_out,
             "validations": self.validations,
             "resumed_from": self.resumed_from,
             "error": self.error,
@@ -220,6 +242,22 @@ def set_validation(self, layer: str, summary: Dict[str, Any]) -> None:
         self.validations[layer] = summary
         self.save()
 
+    def set_layer_metrics(
+        self,
+        layer: str,
+        *,
+        output_rows: Optional[int] = None,
+        output_bytes: Optional[int] = None,
+        tables_count: Optional[int] = None,
+    ) -> None:
+        info = self._layer(layer)
+        info["metrics"] = {
+            "output_rows": output_rows,
+            "output_bytes": output_bytes,
+            "tables_count": tables_count,
+        }
+        self.save()
+
     def complete_run(self, *, success_with_warnings: bool = False) -> None:
         self.finished_at = _now_iso()
         if self.status != "FAILED":
diff --git a/toolkit/mart/run.py b/toolkit/mart/run.py
index be17aa4..327fc63 100644
--- a/toolkit/mart/run.py
+++ b/toolkit/mart/run.py
@@ -77,6 +77,7 @@ def run_mart(
         written: list[Path] = []
         executed: list[dict[str, Any]] = []
         debug_tables: list[dict[str, Any]] = []
+        total_rows = 0
 
         for i, table in enumerate(tables, start=1):
             if not isinstance(table, dict):
@@ -102,6 +103,7 @@ def run_mart(
 
             # Create table and export
             con.execute(f"CREATE OR REPLACE TABLE {name} AS {sql}")
+            total_rows += con.execute(f"SELECT count(*) FROM {name}").fetchone()[0]
 
             out = mart_dir / f"{name}.parquet"
             con.execute(f"COPY {name} TO '{out}' (FORMAT PARQUET);")
@@ -158,4 +160,6 @@ def run_mart(
         errors_count=None,
         warnings_count=None,
     )
+    total_bytes = sum(p.stat().st_size for p in written if p.exists())
     logger.info(f"MART -> {mart_dir}")
+    return {"output_rows": total_rows, "output_bytes": total_bytes, "tables_count": len(written)}
diff --git a/toolkit/raw/run.py b/toolkit/raw/run.py
index dce6a9c..7ae8b2e 100644
--- a/toolkit/raw/run.py
+++ b/toolkit/raw/run.py
@@ -324,3 +324,6 @@ def run_raw(
         raise RuntimeError(f"RAW validation failed for {dataset} {year}. See {vpath}")
     else:
         logger.info(f"RAW QA OK ({dataset} {year}) -> {vpath.name}")
+
+    output_bytes = sum(f.get("bytes", 0) for f in files_written) if files_written else None
+    return {"output_bytes": output_bytes}