diff --git a/tests/test_backtest.py b/tests/test_backtest.py
new file mode 100644
index 000000000..a81ce7778
--- /dev/null
+++ b/tests/test_backtest.py
@@ -0,0 +1,181 @@
+"""Backtesting: many single-shot decisions, scored by the decision log.
+
+A run already records its rating and later settles it with realized and alpha
+return against the regional benchmark. A backtest is that machinery over a grid
+of tickers and dates, aggregated. It evaluates decision quality; it does not
+simulate a portfolio, so there is no execution, no fees and no equity curve.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from tradingagents.agents.utils.memory import TradingMemoryLog
+from tradingagents.backtest import iter_grid, run_backtest, summarize
+
+DECISION = "Rating: Buy\n\nbuy it"
+
+
+@pytest.mark.unit
+def test_grid_spacing_and_canonical_dates():
+    assert iter_grid("2026-01-05", "2026-01-20", every_n_days=7) == ["2026-01-05", "2026-01-12", "2026-01-19"]
+
+
+@pytest.mark.unit
+def test_grid_stops_at_today(monkeypatch):
+    import tradingagents.backtest as bt
+
+    monkeypatch.setattr(bt, "get_current_date", lambda: "2026-01-10")
+    assert iter_grid("2026-01-05", "2026-02-20", every_n_days=5) == ["2026-01-05", "2026-01-10"]
+
+
+@pytest.mark.unit
+def test_grid_rejects_a_non_canonical_date():
+    with pytest.raises(ValueError, match="YYYY-MM-DD"):
+        iter_grid("2026-1-5", "2026-01-20")
+
+
+class _FakeGraph:
+    """Stands in for TradingAgentsGraph, writing to the log the harness gave it."""
+
+    instances: list = []
+    fail_on: set = set()
+
+    def __init__(self, selected_analysts=None, config=None, **kw):
+        self.analysts = list(selected_analysts) if selected_analysts else None
+        self.config = config
+        self.memory_log = TradingMemoryLog(config)
+        self.calls = []
+        self.settled = []
+        _FakeGraph.instances.append(self)
+
+    def propagate(self, ticker, trade_date, asset_type="stock", portfolio=None):
+        self.calls.append((ticker, trade_date))
+        if (ticker, trade_date) in _FakeGraph.fail_on:
+            raise RuntimeError("vendor exploded")
+        self.memory_log.store_decision(ticker, trade_date, DECISION)
+        return {"final_trade_decision": DECISION}, "Buy"
+
+    def settle_pending(self, ticker):
+        self.settled.append(ticker)
+
+
+@pytest.fixture(autouse=True)
+def _fake_graph(monkeypatch, tmp_path):
+    import tradingagents.backtest as bt
+
+    _FakeGraph.instances = []
+    _FakeGraph.fail_on = set()
+    monkeypatch.setattr(bt, "TradingAgentsGraph", _FakeGraph)
+    return _FakeGraph
+
+
+def _config(tmp_path):
+    return {"results_dir": str(tmp_path / "results"),
+            "memory_log_path": str(tmp_path / "live_trading_memory.md")}
+
+
+@pytest.mark.unit
+def test_the_live_decision_log_is_never_written(tmp_path):
+    config = _config(tmp_path)
+    result = run_backtest(["NVDA"], ["2026-01-05", "2026-01-12"], config)
+
+    assert not (tmp_path / "live_trading_memory.md").exists()
+    assert result.log_path.exists() and result.cells_run == 2
+
+
+@pytest.mark.unit
+def test_a_cell_already_in_the_log_is_not_run_again(tmp_path):
+    config = _config(tmp_path)
+    first = run_backtest(["NVDA"], ["2026-01-05"], config)
+
+    again = run_backtest(["NVDA"], ["2026-01-05", "2026-01-12"], config, run_id=first.run_id)
+
+    assert again.cells_run == 1 and again.skipped == 1
+    assert _FakeGraph.instances[-1].calls == [("NVDA", "2026-01-12")]
+
+
+@pytest.mark.unit
+def test_every_ticker_is_settled_after_the_grid(tmp_path):
+    """Settlement runs at the start of the next same-ticker run, so the last
+    date of each ticker would stay pending without an explicit pass."""
+    run_backtest(["NVDA", "AAPL"], ["2026-01-05", "2026-01-12"], _config(tmp_path))
+    assert sorted(_FakeGraph.instances[-1].settled) == ["AAPL", "NVDA"]
+
+
+@pytest.mark.unit
+def test_a_failed_cell_does_not_abort_the_sweep(tmp_path):
+    _FakeGraph.fail_on = {("NVDA", "2026-01-05")}
+    result = run_backtest(["NVDA"], ["2026-01-05", "2026-01-12"], _config(tmp_path))
+
+    assert result.cells_run == 1
+    assert result.failures == [("NVDA", "2026-01-05", "vendor exploded")]
+
+
+# --- reading the result ------------------------------------------------------
+
+def _log_with(tmp_path, rows):
+    log = TradingMemoryLog({"memory_log_path": str(tmp_path / "m.md")})
+    for ticker, date, decision, outcome in rows:
+        log.store_decision(ticker, date, decision)
+        if outcome is not None:
+            log.update_with_outcome(ticker, date, outcome[0], outcome[1], 5, "note", "2026-02-01")
+    return log
+
+
+@pytest.mark.unit
+def test_summary_scores_resolved_cells_and_keeps_pending_out_of_the_average(tmp_path):
+    log = _log_with(tmp_path, [
+        ("NVDA", "2026-01-05", "Rating: Buy\n\nx", (0.10, 0.04)),
+        ("NVDA", "2026-01-12", "Rating: Buy\n\nx", (-0.02, -0.02)),
+        ("AAPL", "2026-01-05", "Rating: Sell\n\nx", None),
+    ])
+
+    summary = summarize(log)
+
+    assert summary.resolved == 2 and summary.pending == 1
+    buys = summary.by_rating["Buy"]
+    assert buys.count == 2 and buys.hit_rate == 0.5 and round(buys.mean_alpha, 4) == 0.01
+    assert "Sell" not in summary.by_rating  # unsettled: nothing to score yet
+
+
+@pytest.mark.unit
+def test_summary_states_what_it_cannot_prove(tmp_path):
+    text = summarize(_log_with(tmp_path, [("NVDA", "2026-01-05", DECISION, (0.1, 0.05))])).render()
+    assert "not archived" in text
+    assert "one" in text.lower() and "sampl" in text.lower()
+
+
+@pytest.mark.unit
+def test_the_analyst_set_under_test_is_the_one_that_runs(tmp_path):
+    """A backtest of a two-analyst setup must not silently run four."""
+    run_backtest(["NVDA"], ["2026-01-05"], _config(tmp_path), selected_analysts=["market", "news"])
+    assert _FakeGraph.instances[-1].analysts == ["market", "news"]
+
+
+@pytest.mark.unit
+def test_a_run_id_cannot_escape_the_results_directory(tmp_path):
+    """run_id becomes a path segment, so it is validated like a ticker is."""
+    with pytest.raises(ValueError):
+        run_backtest(["NVDA"], ["2026-01-05"], _config(tmp_path), run_id="../../escaped")
+    with pytest.raises(ValueError):
+        run_backtest(["NVDA"], ["2026-01-05"], _config(tmp_path), run_id="/etc/cron.d/x")
+
+
+@pytest.mark.unit
+def test_a_failed_settlement_does_not_lose_the_remaining_tickers(tmp_path, monkeypatch):
+    """Settlement reflects with an LLM, so it can fail; the sweep still returns
+    its result and every other ticker still gets settled."""
+    settled = []
+
+    def _settle(self, ticker):
+        if ticker == "NVDA":
+            raise RuntimeError("reflector timed out")
+        settled.append(ticker)
+
+    monkeypatch.setattr(_FakeGraph, "settle_pending", _settle, raising=False)
+    result = run_backtest(["NVDA", "AAPL"], ["2026-01-05"], _config(tmp_path))
+
+    assert result.cells_run == 2
+    assert settled == ["AAPL"]
+    assert result.settlement_failures == [("NVDA", "reflector timed out")]
diff --git a/tradingagents/backtest.py b/tradingagents/backtest.py
new file mode 100644
index 000000000..6557f3f96
--- /dev/null
+++ b/tradingagents/backtest.py
@@ -0,0 +1,176 @@
+"""Run the graph over a grid of tickers and dates, and score what came back.
+
+One run yields one decision, so it cannot say whether the system decides well.
+This runs the same machinery over many (ticker, date) cells and reads the
+aggregate. The decision log is the results table: every run already records its
+rating and later settles it with realized and alpha return against the
+instrument's regional benchmark, so there is nothing to record separately.
+
+Scope: this evaluates decision quality. It is not a portfolio simulator, and
+must not grow one. Turning a rating into a filled order needs a quantity, a fill
+price and a cash ledger, none of which the system has; inventing them here would
+put an execution model behind an evaluation tool. Cells are therefore
+independent, and a portfolio, when given, is the same standing book for every
+cell rather than a position carried forward.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from datetime import datetime, timedelta
+from pathlib import Path
+
+from tradingagents.agents.utils.memory import TradingMemoryLog
+from tradingagents.dataflows.utils import get_current_date, safe_ticker_component
+from tradingagents.graph.trading_graph import TradingAgentsGraph
+
+logger = logging.getLogger(__name__)
+
+
+def iter_grid(start_date: str, end_date: str, every_n_days: int = 1) -> list[str]:
+    """Analysis dates from ``start_date``, never past today.
+
+    A future date has no outcome to settle against, and the graph rejects one, so
+    the grid stops at the present rather than producing cells that cannot score.
+    """
+    start, end = _canonical(start_date), _canonical(end_date)
+    if every_n_days < 1:
+        raise ValueError("every_n_days must be at least 1")
+
+    last = min(end, datetime.strptime(get_current_date(), "%Y-%m-%d"))
+    dates, cursor = [], start
+    while cursor <= last:
+        dates.append(cursor.strftime("%Y-%m-%d"))
+        cursor += timedelta(days=every_n_days)
+    return dates
+
+
+def _canonical(date: str) -> datetime:
+    """Parse a grid bound, rejecting anything the run date would also reject."""
+    try:
+        parsed = datetime.strptime(str(date), "%Y-%m-%d")
+    except (TypeError, ValueError) as exc:
+        raise ValueError(f"grid dates must be in YYYY-MM-DD format, got {date!r}") from exc
+    if parsed.strftime("%Y-%m-%d") != str(date):
+        raise ValueError(f"grid dates must be in YYYY-MM-DD format, got {date!r}")
+    return parsed
+
+
+def _alpha(entry: dict) -> float | None:
+    """Alpha return of a settled entry, or None when it has not settled.
+
+    The log stores it as a percentage rounded to one decimal, so aggregates here
+    are accurate to 0.1 of a percentage point, not to the raw quote.
+    """
+    text = (entry.get("alpha") or "").strip().rstrip("%")
+    try:
+        return float(text) / 100
+    except ValueError:
+        return None
+
+
+@dataclass
+class BacktestResult:
+    run_id: str
+    log_path: Path
+    cells_run: int = 0
+    skipped: int = 0
+    failures: list[tuple[str, str, str]] = field(default_factory=list)
+    settlement_failures: list[tuple[str, str]] = field(default_factory=list)
+
+
+@dataclass
+class RatingScore:
+    count: int
+    hit_rate: float
+    mean_alpha: float
+
+
+@dataclass
+class BacktestSummary:
+    resolved: int
+    pending: int
+    by_rating: dict[str, RatingScore]
+
+    def render(self) -> str:
+        lines = [f"Resolved cells: {self.resolved} · pending: {self.pending}"]
+        for rating, score in self.by_rating.items():
+            lines.append(
+                f"- {rating}: n={score.count}, beat the benchmark "
+                f"{score.hit_rate:.0%}, mean alpha {score.mean_alpha:+.2%}"
+            )
+        lines += [
+            "",
+            "Pending cells are not scored above; re-run to settle them.",
+            "One model sampling per cell, and text feeds are not archived, so "
+            "these figures are indicative rather than repeatable.",
+        ]
+        return "\n".join(lines)
+
+
+def run_backtest(
+    tickers: list[str],
+    dates: list[str],
+    config: dict,
+    asset_type: str = "stock",
+    portfolio=None,
+    selected_analysts=("market", "social", "news", "fundamentals"),
+    run_id: str | None = None,
+) -> BacktestResult:
+    """Analyze every ticker on every date, into a decision log of this run's own.
+
+    The live log stays untouched: a sweep would otherwise flood the context that
+    real runs read back. Cells already in this run's log are skipped, so an
+    interrupted sweep resumes by being run again.
+    """
+    # run_id becomes a path segment, so it is validated like a ticker: an
+    # absolute or dotted value would otherwise place the run outside results_dir.
+    run_id = safe_ticker_component(run_id or datetime.now().strftime("%Y%m%d_%H%M%S"))
+    run_dir = Path(config["results_dir"]) / "backtest" / run_id
+    run_dir.mkdir(parents=True, exist_ok=True)
+    run_config = {**config, "results_dir": str(run_dir),
+                  "memory_log_path": str(run_dir / "trading_memory.md")}
+
+    graph = TradingAgentsGraph(selected_analysts, config=run_config)
+    result = BacktestResult(run_id=run_id, log_path=Path(run_config["memory_log_path"]))
+    done = {(e["ticker"], e["date"]) for e in graph.memory_log.load_entries()}
+
+    for ticker in tickers:
+        for date in dates:
+            if (ticker, date) in done:
+                result.skipped += 1
+                continue
+            try:
+                graph.propagate(ticker, date, asset_type, portfolio=portfolio)
+                result.cells_run += 1
+            except Exception as exc:  # one unreachable vendor must not end the sweep
+                logger.warning("Backtest cell %s %s failed: %s", ticker, date, exc)
+                result.failures.append((ticker, date, str(exc)))
+
+    # Settlement runs at the start of the next run for a ticker, so each ticker's
+    # last cell would stay pending without this pass.
+    for ticker in tickers:
+        try:
+            graph.settle_pending(ticker)
+        except Exception as exc:  # reflection calls an LLM; one failure is not the sweep's
+            logger.warning("Settling %s failed: %s", ticker, exc)
+            result.settlement_failures.append((ticker, str(exc)))
+    return result
+
+
+def summarize(memory_log: TradingMemoryLog) -> BacktestSummary:
+    """Score the settled decisions in a log, by rating."""
+    entries = memory_log.load_entries()
+    resolved = [(e, _alpha(e)) for e in entries if not e["pending"]]
+    resolved = [(e, a) for e, a in resolved if a is not None]
+    by_rating: dict[str, RatingScore] = {}
+    for rating in dict.fromkeys(e["rating"] for e, _ in resolved):
+        alphas = [a for e, a in resolved if e["rating"] == rating]
+        by_rating[rating] = RatingScore(
+            count=len(alphas),
+            hit_rate=sum(a > 0 for a in alphas) / len(alphas),
+            mean_alpha=sum(alphas) / len(alphas),
+        )
+    return BacktestSummary(resolved=len(resolved), pending=len(entries) - len(resolved),
+                           by_rating=by_rating)