diff --git a/vero/src/vero/harbor/cli.py b/vero/src/vero/harbor/cli.py
index e68ce4c..11e2f52 100644
--- a/vero/src/vero/harbor/cli.py
+++ b/vero/src/vero/harbor/cli.py
@@ -120,8 +120,14 @@ def finalize_cmd(token_file, output):
     from vero.harbor.auth import read_admin_token
 
     token = read_admin_token(token_file)
-    reward = _request("POST", "/finalize", headers={"Authorization": f"Bearer {token}"})
+    resp = _request("POST", "/finalize", headers={"Authorization": f"Bearer {token}"})
+    # finalize returns {"rewards": {...}, "baseline": {...}}. Only the rewards are
+    # the reward.json payload the outer harness consumes; the baseline outcome is
+    # echoed to stdout (the trial's stdout survives teardown, the admin volume does
+    # not) so a baseline skip or failure is durably recorded. Tolerate the older
+    # bare-rewards shape so a mixed-version sidecar still writes a valid reward.json.
+    rewards = resp["rewards"] if isinstance(resp, dict) and "rewards" in resp else resp
     out = Path(output)
     out.parent.mkdir(parents=True, exist_ok=True)
-    out.write_text(json.dumps(reward))
-    click.echo(json.dumps(reward, indent=2))
+    out.write_text(json.dumps(rewards))
+    click.echo(json.dumps(resp, indent=2))
diff --git a/vero/src/vero/harbor/protocol.py b/vero/src/vero/harbor/protocol.py
index 1f41ed7..20c0500 100644
--- a/vero/src/vero/harbor/protocol.py
+++ b/vero/src/vero/harbor/protocol.py
@@ -51,6 +51,10 @@ class StatusSummary:
     submit_enabled: bool
     # per (split, dataset_id): tier + whether the agent may evaluate it + remaining budget
     splits: list[dict] = field(default_factory=list)
+    # the seeded baseline sha and whether its one budget-free reference eval
+    # is still available (None/False when the task has no recorded baseline)
+    base_commit: str | None = None
+    free_baseline_available: bool = False
 
     def to_dict(self) -> dict:
         return asdict(self)
@@ -98,6 +102,8 @@ def build_status(
     submit_enabled: bool,
     budget: dict[tuple[str, str], SplitBudget],
     split_accesses: list[SplitAccess],
+    base_commit: str | None = None,
+    free_baseline_available: bool = False,
 ) -> StatusSummary:
     """Build the agent-facing status from the budget ledger + split tiers.
 
@@ -117,4 +123,9 @@ def build_status(
                 "remaining_run_budget": b.remaining_run_budget,
             }
         )
-    return StatusSummary(submit_enabled=submit_enabled, splits=splits)
+    return StatusSummary(
+        submit_enabled=submit_enabled,
+        splits=splits,
+        base_commit=base_commit,
+        free_baseline_available=free_baseline_available,
+    )
diff --git a/vero/src/vero/harbor/serve.py b/vero/src/vero/harbor/serve.py
index 85f6c28..962805a 100644
--- a/vero/src/vero/harbor/serve.py
+++ b/vero/src/vero/harbor/serve.py
@@ -71,6 +71,12 @@ class ServeConfig(BaseModel):
     # write it to <admin_volume>/baseline.json: makes regressions visible
     # (an optimized candidate can score WORSE than the untouched baseline).
     score_baseline: bool = False
+    # Total attempts for the finalize baseline eval (>=1): a transient nested-run
+    # failure once silently dropped the regression check.
+    baseline_score_attempts: int = 2
+    # auto_best never ships a candidate that fails to beat the untouched baseline
+    # on the selection split; it reverts to base_commit instead (needs base_commit).
+    auto_best_baseline_floor: bool = True
 
     # volumes / token
     agent_volume: str
@@ -221,6 +227,7 @@ async def build_components(config: ServeConfig) -> tuple[EvaluationSidecar, Veri
         agent_volume=Path(config.agent_volume),
         admin_volume=Path(config.admin_volume),
         submit_enabled=config.submit_enabled,
+        base_commit=config.base_commit,
     )
     verifier = Verifier(
         engine=engine,
@@ -232,6 +239,8 @@ async def build_components(config: ServeConfig) -> tuple[EvaluationSidecar, Veri
         selection_task=config.task,
         selection_dataset_id=config.dataset_id,
         score_baseline=config.score_baseline,
+        baseline_score_attempts=config.baseline_score_attempts,
+        auto_best_baseline_floor=config.auto_best_baseline_floor,
     )
 
     token = generate_token()
diff --git a/vero/src/vero/harbor/server.py b/vero/src/vero/harbor/server.py
index f75cf6e..8724960 100644
--- a/vero/src/vero/harbor/server.py
+++ b/vero/src/vero/harbor/server.py
@@ -55,6 +55,7 @@ def __init__(
         agent_volume: Path,
         admin_volume: Path,
         submit_enabled: bool = False,
+        base_commit: str | None = None,
     ):
         self.engine = engine
         self.split_accesses = split_accesses
@@ -62,6 +63,8 @@ def __init__(
         self.agent_volume = Path(agent_volume)
         self.admin_volume = Path(admin_volume)
         self.submit_enabled = submit_enabled
+        self.base_commit = base_commit
+        self._free_baseline_used = False
 
     # ------------------------------------------------------------------
     # Handlers (the HTTP layer resolves `admin` from auth and calls these)
@@ -69,7 +72,31 @@ def __init__(
 
     async def evaluate(self, req: EvalRequest, *, admin: bool = False) -> EvalSummary:
         sha = await self._transfer_commit(req.commit)
-        exp = await self.engine.evaluate(replace(req, commit=sha), admin=admin)
+        # The agent's FIRST eval of the seeded baseline is budget-free. The
+        # baseline is the reference every candidate is implicitly compared to,
+        # yet it can never win selection (auto_best excludes base_commit), so
+        # metering it forces a choice between optimizing blind and paying a
+        # budgeted eval for a commit that cannot be selected (observed live:
+        # an optimizer that skipped the reference could not tell a no-op edit
+        # from an improvement and quit with budget unspent). Capped at one:
+        # later baseline evals debit normally, so free compute is bounded.
+        free_baseline = (
+            not admin
+            and self.base_commit is not None
+            and sha == self.base_commit
+            and not self._free_baseline_used
+        )
+        exp = await self.engine.evaluate(
+            replace(req, commit=sha), admin=admin or free_baseline
+        )
+        # Consume the free slot only after the eval actually succeeds. Setting it
+        # before the await would burn the one free baseline on a transient engine
+        # failure (timeout, infra), forcing the agent to pay for the retry, which is
+        # the exact failure mode this feature prevents. Safe in the single-threaded
+        # asyncio loop: no await runs between the check above and this write.
+        if free_baseline:
+            self._free_baseline_used = True
+        # Route with the agent's real tier even when the eval was unmetered.
         result_path = self._route_results(exp, admin=admin)
         budget_remaining = None
         if not admin:
@@ -99,6 +126,10 @@ def status(self) -> StatusSummary:
             submit_enabled=self.submit_enabled,
             budget=self.engine.budget.status(),
             split_accesses=self.split_accesses,
+            base_commit=self.base_commit,
+            free_baseline_available=(
+                self.base_commit is not None and not self._free_baseline_used
+            ),
         )
 
     # ------------------------------------------------------------------
diff --git a/vero/src/vero/harbor/verifier.py b/vero/src/vero/harbor/verifier.py
index ddb2614..d0d996c 100644
--- a/vero/src/vero/harbor/verifier.py
+++ b/vero/src/vero/harbor/verifier.py
@@ -53,6 +53,8 @@ def __init__(
         selection_dataset_id: str | None = None,
         rescore_top_k: int = 3,
         score_baseline: bool = False,
+        baseline_score_attempts: int = 2,
+        auto_best_baseline_floor: bool = True,
     ):
         self.engine = engine
         self.admin_volume = Path(admin_volume)
@@ -67,9 +69,27 @@ def __init__(
         self.selection_dataset_id = selection_dataset_id
         self.rescore_top_k = rescore_top_k
         self.score_baseline = score_baseline
+        # auto_best selection floor: never ship a candidate that fails to beat the
+        # untouched baseline on the selection split. Without it, auto_best (which
+        # excludes base_commit from the candidate pool) selects the least-bad
+        # candidate even when every candidate regressed, shipping a regression
+        # (observed live: a weak inner model, every candidate below baseline).
+        self.auto_best_baseline_floor = auto_best_baseline_floor
+        # Baseline scoring is retried this many times total before its outcome is
+        # reported as an error; the nested eval can fail transiently (a nested
+        # harbor run crashing right after a large eval), and a single blip must
+        # not silently drop the regression check.
+        self._baseline_score_attempts = max(1, baseline_score_attempts)
 
-    async def finalize(self) -> dict[str, float]:
-        """Select the commit and score it on every target -> {reward_key: score}.
+    async def finalize(self) -> dict:
+        """Select the commit, score it on every target, and score the baseline.
+
+        Returns a wrapper ``{"rewards": {reward_key: score}, "baseline": {...}}``.
+        ``rewards`` is the reward.json payload the outer harness consumes (the CLI
+        writes only that to reward.json); ``baseline`` is the outcome of baseline
+        scoring, surfaced here because it is otherwise invisible: the admin volume
+        it used to be written to does not survive teardown, and the finalize
+        response echoed to the trial's stdout is the only host-durable channel.
 
         A run in which the optimizer produced no scorable candidate (never
         submitted in ``submit`` mode; no non-baseline experiments on the
@@ -89,7 +109,8 @@ async def finalize(self) -> dict[str, float]:
                 len(self.targets),
                 default_minimum_score,
             )
-            return {t.reward_key: float(default_minimum_score) for t in self.targets}
+            rewards = {t.reward_key: float(default_minimum_score) for t in self.targets}
+            return {"rewards": rewards, "baseline": {"skipped": "no candidate commit"}}
         logger.info(f"Verifier selected commit {sha} (mode={self.reward_mode})")
         rewards: dict[str, float] = {}
         for target in self.targets:
@@ -104,22 +125,29 @@ async def finalize(self) -> dict[str, float]:
             rewards[target.reward_key] = (
                 float(score) if score is not None else default_minimum_score
             )
-        await self._maybe_score_baseline(rewards)
-        return rewards
+        baseline = await self._maybe_score_baseline(rewards)
+        return {"rewards": rewards, "baseline": baseline}
 
-    async def _maybe_score_baseline(self, rewards: dict[str, float]) -> None:
-        """Admin-score the unmodified baseline on every target and persist it.
+    async def _maybe_score_baseline(self, rewards: dict[str, float]) -> dict:
+        """Admin-score the unmodified baseline on every target and report it.
 
         An optimized candidate can score WORSE than the untouched baseline
         (observed live: a weak inner model went 0.3 -> 0.2 after optimization);
         without this, the regression is invisible because auto_best excludes the
-        baseline from selection and nothing else ever scores it. Written to
-        <admin_volume>/baseline.json (NOT into reward.json, whose keys the outer
-        harness consumes) and logged next to the candidate's rewards. Failures
-        here never fail the trial.
+        baseline from selection and nothing else ever scores it.
+
+        Returns a structured outcome (``{"scores": ...}`` / ``{"error": ...}`` /
+        ``{"skipped": ...}``) that ``finalize`` surfaces in its response, so a
+        skip or failure is durably recorded rather than lost. A live trial once
+        skipped this silently: the nested baseline eval failed transiently and
+        the only record (a log line) died with the container at teardown. So the
+        eval is retried once, and any failure is returned instead of swallowed.
+        Baseline scoring still never fails the trial (reward.json is unaffected).
+        A best-effort copy is also written to <admin_volume>/baseline.json for
+        in-cluster debugging while the sidecar is alive.
         """
         if not self.score_baseline:
-            return
+            return {"skipped": "score_baseline is disabled"}
         if not self.base_commit:
             # Misconfiguration must not be a silent no-op: the operator asked
             # for baseline scoring and would otherwise never learn it is off.
@@ -127,33 +155,60 @@ async def _maybe_score_baseline(self, rewards: dict[str, float]) -> None:
                 "score_baseline=True but base_commit is not set; skipping "
                 "baseline scoring."
             )
-            return
-        try:
-            baselines: dict[str, float] = {}
-            for target in self.targets:
-                exp = await self.engine.evaluate_admin(
-                    task=target.task,
-                    dataset_id=target.dataset_id,
-                    split=target.split,
-                    commit=self.base_commit,
-                    sample_ids=target.sample_ids,
-                )
-                score = exp.result.score()
-                baselines[target.reward_key] = (
-                    float(score) if score is not None else default_minimum_score
-                )
-            self.admin_volume.mkdir(parents=True, exist_ok=True)
-            (self.admin_volume / "baseline.json").write_text(
-                json.dumps(baselines, indent=2)
-            )
-            for key, value in rewards.items():
-                base = baselines.get(key)
-                tag = " (REGRESSION vs baseline)" if base is not None and value < base else ""
-                logger.info(
-                    "finalize: %s=%s baseline=%s%s", key, value, base, tag
+            return {"skipped": "base_commit is not set"}
+
+        last_error: Exception | None = None
+        for attempt in range(1, self._baseline_score_attempts + 1):
+            try:
+                baselines: dict[str, float] = {}
+                for target in self.targets:
+                    exp = await self.engine.evaluate_admin(
+                        task=target.task,
+                        dataset_id=target.dataset_id,
+                        split=target.split,
+                        commit=self.base_commit,
+                        sample_ids=target.sample_ids,
+                    )
+                    score = exp.result.score()
+                    baselines[target.reward_key] = (
+                        float(score) if score is not None else default_minimum_score
+                    )
+                # Best-effort local copy (admin volume does not survive teardown;
+                # the return value is the durable record).
+                try:
+                    self.admin_volume.mkdir(parents=True, exist_ok=True)
+                    (self.admin_volume / "baseline.json").write_text(
+                        json.dumps(baselines, indent=2)
+                    )
+                except OSError:
+                    logger.warning("could not write baseline.json to the admin volume")
+                for key, value in rewards.items():
+                    base = baselines.get(key)
+                    tag = (
+                        " (REGRESSION vs baseline)"
+                        if base is not None and value < base
+                        else ""
+                    )
+                    logger.info("finalize: %s=%s baseline=%s%s", key, value, base, tag)
+                return {"scores": baselines, "attempts": attempt}
+            except Exception as exc:  # noqa: BLE001 - never fail the trial on baseline scoring
+                last_error = exc
+                logger.warning(
+                    "baseline scoring attempt %d/%d failed: %s",
+                    attempt,
+                    self._baseline_score_attempts,
+                    exc,
                 )
-        except Exception:
-            logger.exception("baseline scoring failed; reward.json is unaffected")
+        logger.exception(
+            "baseline scoring failed after %d attempt(s); reward.json is unaffected",
+            self._baseline_score_attempts,
+            exc_info=last_error,
+        )
+        return {
+            "error": str(last_error),
+            "error_type": type(last_error).__name__ if last_error else None,
+            "attempts": self._baseline_score_attempts,
+        }
 
     async def _select_commit(self) -> str:
         if self.reward_mode == "submit":
@@ -236,4 +291,36 @@ async def _best_from_db(self) -> str:
             )
         # Highest admin score wins; ties break to the earliest shortlist position.
         rescored.sort(key=lambda t: (-t[0], t[1]))
-        return rescored[0][2]
+        best_score, _, best_commit = rescored[0]
+
+        # Selection floor: never ship a candidate that fails to beat the untouched
+        # baseline on the selection split. auto_best excludes base_commit from the
+        # candidate pool, so without this it selects the least-bad candidate even
+        # when every candidate regressed. Revert to the seed instead. Strict '>' so
+        # a statistical tie also reverts: if the optimizer cannot show an
+        # improvement, shipping the seed is the safe outcome. Needs a base_commit to
+        # compare against; costs one extra admin eval on the selection split.
+        if self.auto_best_baseline_floor and self.base_commit is not None:
+            base_dataset_id = self.selection_dataset_id
+            if base_dataset_id is None:
+                base_dataset_id = shortlist.iloc[0].get("dataset_subset_dataset_id")
+            base_exp = await self.engine.evaluate_admin(
+                task=self.selection_task,
+                dataset_id=base_dataset_id,
+                split=self.selection_split,
+                commit=self.base_commit,
+            )
+            base_s = base_exp.result.score()
+            base_score = float(base_s) if base_s is not None else default_minimum_score
+            if best_score <= base_score:
+                logger.info(
+                    "auto_best floor: best candidate %s (admin_score=%s) does not beat "
+                    "baseline %s (admin_score=%s); reverting to base_commit.",
+                    best_commit, best_score, self.base_commit, base_score,
+                )
+                return self.base_commit
+            logger.info(
+                "auto_best floor: best candidate %s (%s) beats baseline (%s); keeping it.",
+                best_commit, best_score, base_score,
+            )
+        return best_commit
diff --git a/vero/tests/test_harbor_app.py b/vero/tests/test_harbor_app.py
index c1ae87f..59892b2 100644
--- a/vero/tests/test_harbor_app.py
+++ b/vero/tests/test_harbor_app.py
@@ -75,7 +75,10 @@ def test_budget_exceeded_maps_to_429(self):
 class TestAdminEndpoint:
     def test_finalize_requires_token(self):
         verifier = MagicMock()
-        verifier.finalize = AsyncMock(return_value={"reward": 1.0})
+        # Mock mirrors the real finalize contract: {"rewards": ..., "baseline": ...}
+        # (the CLI extracts "rewards" for reward.json and echoes the rest to stdout).
+        payload = {"rewards": {"reward": 1.0}, "baseline": {"scores": {"reward": 0.8}}}
+        verifier.finalize = AsyncMock(return_value=payload)
         client = _client(verifier=verifier)
 
         assert client.post("/finalize").status_code == 403  # no token
@@ -83,7 +86,7 @@ def test_finalize_requires_token(self):
         verifier.finalize.assert_not_awaited()
 
         r = client.post("/finalize", headers={"Authorization": f"Bearer {TOKEN}"})
-        assert r.status_code == 200 and r.json() == {"reward": 1.0}
+        assert r.status_code == 200 and r.json() == payload
         verifier.finalize.assert_awaited_once()
 
 
diff --git a/vero/tests/test_harbor_cli.py b/vero/tests/test_harbor_cli.py
index afab16c..bc0b125 100644
--- a/vero/tests/test_harbor_cli.py
+++ b/vero/tests/test_harbor_cli.py
@@ -79,4 +79,27 @@ def test_finalize_uses_token_and_writes_reward(monkeypatch, tmp_path):
     assert result.exit_code == 0
     assert cap["url"].endswith("/finalize")
     assert cap["headers"]["Authorization"] == "Bearer T0KEN"
+    # Back-compat: a bare-rewards response (older sidecar) still writes reward.json.
     assert json.loads(out.read_text()) == {"reward": 1.0}
+
+
+def test_finalize_writes_only_rewards_and_echoes_baseline(monkeypatch, tmp_path):
+    # New wrapper shape: reward.json gets only the rewards (the outer harness
+    # consumes its keys), while the baseline outcome is echoed to stdout, the one
+    # channel that survives teardown, so a baseline skip/failure is durably recorded.
+    monkeypatch.setenv("VERO_EVAL_URL", "http://sidecar:8000")
+    token_file = tmp_path / "tok"
+    token_file.write_text("T0KEN")
+    out = tmp_path / "reward.json"
+    cap: dict = {}
+    resp = {"rewards": {"accuracy": 0.4}, "baseline": {"scores": {"accuracy": 0.43}}}
+    _patch_httpx(monkeypatch, _Resp(200, resp), cap)
+
+    result = CliRunner().invoke(
+        harbor, ["finalize", "--token-file", str(token_file), "--output", str(out)]
+    )
+    assert result.exit_code == 0
+    # reward.json is only the rewards, not the baseline wrapper
+    assert json.loads(out.read_text()) == {"accuracy": 0.4}
+    # the baseline outcome is visible on stdout (captured into test-stdout.txt on the host)
+    assert "baseline" in result.output and "0.43" in result.output
diff --git a/vero/tests/test_harbor_serve.py b/vero/tests/test_harbor_serve.py
index df023e8..655c389 100644
--- a/vero/tests/test_harbor_serve.py
+++ b/vero/tests/test_harbor_serve.py
@@ -139,7 +139,7 @@ async def test_serve_assembles_and_evaluates_and_finalizes(fixture):
     assert exp.result.sample_results[0].score == 1.0
 
     # verifier selects the (only) candidate on "test" and scores it on the test target
-    rewards = await verifier.finalize()
+    rewards = (await verifier.finalize())["rewards"]
     assert rewards["reward"] == 1.0
 
 
@@ -197,7 +197,7 @@ async def test_finalize_does_not_run_agent_supplied_scorer(fixture):
     )
     assert exp.result.sample_results[0].score == 0.0
     # Finalize must reflect the TRUSTED score, not the agent's 1.0 scorer.
-    rewards = await verifier.finalize()
+    rewards = (await verifier.finalize())["rewards"]
     assert rewards["reward"] == 0.0
 
 
diff --git a/vero/tests/test_harbor_server.py b/vero/tests/test_harbor_server.py
index 16986a4..44b0078 100644
--- a/vero/tests/test_harbor_server.py
+++ b/vero/tests/test_harbor_server.py
@@ -41,7 +41,7 @@ def _experiment(split: str, commit: str = "abcdef123456") -> Experiment:
     )
 
 
-def _sidecar(tmp_path, *, split, submit_enabled=False):
+def _sidecar(tmp_path, *, split, submit_enabled=False, base_commit=None):
     engine = MagicMock()
     engine.evaluate = AsyncMock(return_value=_experiment(split))
     engine.budget = BudgetLedger(
@@ -54,6 +54,7 @@ def _sidecar(tmp_path, *, split, submit_enabled=False):
         agent_volume=tmp_path / "agent_vol",
         admin_volume=tmp_path / "admin_vol",
         submit_enabled=submit_enabled,
+        base_commit=base_commit,
     )
     # Stub the git transfer (integration-tested separately); pin the sha.
     sidecar._transfer_commit = AsyncMock(return_value="abcdef123456")
@@ -122,3 +123,80 @@ def test_status_reports_submit_and_splits(self, tmp_path):
         assert status.submit_enabled is True
         assert status.splits[0]["split"] == "train"
         assert status.splits[0]["remaining_run_budget"] == 5
+
+
+class TestFreeBaselineEval:
+    """The agent's first eval of the seeded baseline is budget-free: it is the
+    reference every candidate is compared to and can never win selection, so
+    metering it forced a choice between optimizing blind and wasting budget
+    (observed live: exp5's optimizer skipped the reference, could not tell a
+    no-op edit from an improvement, and quit with budget unspent)."""
+
+    @pytest.mark.asyncio
+    async def test_first_baseline_eval_is_unmetered(self, tmp_path):
+        sidecar = _sidecar(tmp_path, split="validation", base_commit="abcdef123456")
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        # engine.evaluate was called with admin=True (bypasses the ledger)
+        assert sidecar.engine.evaluate.await_args.kwargs["admin"] is True
+        # but results were routed with the agent tier (summary written)
+        dest = tmp_path / "agent_vol" / "results" / "validation__abcdef123456"
+        assert (dest / "summary.json").exists()
+
+    @pytest.mark.asyncio
+    async def test_second_baseline_eval_is_metered(self, tmp_path):
+        sidecar = _sidecar(tmp_path, split="validation", base_commit="abcdef123456")
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        assert sidecar.engine.evaluate.await_args.kwargs["admin"] is False
+
+    @pytest.mark.asyncio
+    async def test_non_baseline_commit_always_metered(self, tmp_path):
+        sidecar = _sidecar(tmp_path, split="validation", base_commit="other000000")
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        assert sidecar.engine.evaluate.await_args.kwargs["admin"] is False
+
+    @pytest.mark.asyncio
+    async def test_no_base_commit_never_free(self, tmp_path):
+        sidecar = _sidecar(tmp_path, split="validation")  # base_commit=None
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        assert sidecar.engine.evaluate.await_args.kwargs["admin"] is False
+
+    def test_status_surfaces_free_baseline(self, tmp_path):
+        sidecar = _sidecar(tmp_path, split="train", base_commit="abcdef123456")
+        s = sidecar.status()
+        assert s.base_commit == "abcdef123456"
+        assert s.free_baseline_available is True
+
+    @pytest.mark.asyncio
+    async def test_status_flips_after_use(self, tmp_path):
+        sidecar = _sidecar(tmp_path, split="validation", base_commit="abcdef123456")
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        assert sidecar.status().free_baseline_available is False
+
+    @pytest.mark.asyncio
+    async def test_admin_baseline_eval_does_not_consume_free_slot(self, tmp_path):
+        # An admin re-score of the base commit (finalize, score_baseline) must leave
+        # the agent's free slot intact: the guard is `not admin`, but pin it so a
+        # refactor that moves the flag outside the `if free_baseline` block regresses.
+        sidecar = _sidecar(tmp_path, split="validation", base_commit="abcdef123456")
+        await sidecar.evaluate(
+            EvalRequest(dataset_id="ds1", split="validation"), admin=True
+        )
+        assert sidecar._free_baseline_used is False
+        assert sidecar.status().free_baseline_available is True
+
+    @pytest.mark.asyncio
+    async def test_failed_baseline_eval_leaves_free_slot_available(self, tmp_path):
+        # A transient engine failure must NOT burn the one free baseline: the agent
+        # should be able to retry for free. The flag is consumed only after success.
+        sidecar = _sidecar(tmp_path, split="validation", base_commit="abcdef123456")
+        sidecar.engine.evaluate = AsyncMock(side_effect=RuntimeError("transient infra"))
+        with pytest.raises(RuntimeError):
+            await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        assert sidecar._free_baseline_used is False
+        assert sidecar.status().free_baseline_available is True
+        # The retry is still granted for free (admin=True bypasses the ledger).
+        sidecar.engine.evaluate = AsyncMock(return_value=_experiment("validation"))
+        await sidecar.evaluate(EvalRequest(dataset_id="ds1", split="validation"))
+        assert sidecar.engine.evaluate.await_args.kwargs["admin"] is True
+        assert sidecar._free_baseline_used is True
diff --git a/vero/tests/test_harbor_verifier.py b/vero/tests/test_harbor_verifier.py
index 4672e1a..735763a 100644
--- a/vero/tests/test_harbor_verifier.py
+++ b/vero/tests/test_harbor_verifier.py
@@ -28,7 +28,7 @@ async def test_finalize_submit_scores_nominated_commit(self, tmp_path):
             reward_mode="submit",
             targets=[VerificationTarget(task="t", dataset_id="ds1", split="test", reward_key="reward")],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"reward": 0.8}
         assert engine.evaluate_admin.await_args.kwargs["commit"] == "deadbeef"
         assert engine.evaluate_admin.await_args.kwargs["split"] == "test"
@@ -48,7 +48,7 @@ async def test_finalize_submit_no_submission_floors_rewards(self, tmp_path):
                 VerificationTarget(task="t2", dataset_id="ds2", split="test", reward_key="held_out"),
             ],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"reward": 0.0, "held_out": 0.0}
         engine.evaluate_admin.assert_not_awaited()
 
@@ -67,7 +67,7 @@ async def test_finalize_emits_multiple_reward_keys(self, tmp_path):
                 VerificationTarget(task="t2", dataset_id="ds2", split="test", reward_key="held_out"),
             ],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"in_domain": 0.9, "held_out": 0.4}
         assert engine.evaluate_admin.await_count == 2
 
@@ -103,7 +103,7 @@ async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
             selection_task="math",
             targets=[VerificationTarget(task="math", dataset_id="ds1", split="test", reward_key="reward")],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         # the final (target) eval is on the WINNER 'lo', chosen by admin re-score
         assert engine.evaluate_admin.await_args.kwargs["commit"] == "lo"
         assert engine.evaluate_admin.await_args.kwargs["split"] == "test"
@@ -111,7 +111,9 @@ async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
         assert rewards["reward"] == 0.95
 
     @pytest.mark.asyncio
-    async def test_auto_best_excludes_baseline_after_rescore(self, tmp_path):
+    async def test_auto_best_excludes_baseline_from_ranking(self, tmp_path):
+        # base_commit is excluded from the candidate ranking pool. Floor off here so
+        # the test isolates ranking-exclusion (the floor is covered separately below).
         engine = MagicMock()
         engine.db.get_experiments_df.return_value = pd.DataFrame(
             {
@@ -134,6 +136,7 @@ async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
             selection_split="validation",
             base_commit="base",
             selection_task="math",
+            auto_best_baseline_floor=False,
             targets=[VerificationTarget(task="math", dataset_id="ds1", split="test", reward_key="reward")],
         )
         await v.finalize()
@@ -143,6 +146,169 @@ async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
         assert engine.evaluate_admin.await_args.kwargs["commit"] == "agent"
 
 
+class TestAutoBestBaselineFloor:
+    """auto_best never ships a candidate that fails to beat the baseline.
+
+    auto_best excludes base_commit from the candidate pool, so without a floor it
+    selects the least-bad candidate even when every candidate regressed (observed
+    live: a weak inner model, every candidate below baseline, shipped a -0.10
+    regression despite the free baseline being available). The floor reverts to the
+    seed instead.
+    """
+
+    def _df(self):
+        return pd.DataFrame(
+            {
+                "dataset_subset_split": ["train", "train"],
+                "dataset_subset_dataset_id": ["ds1", "ds1"],
+                "candidate_commit": ["base", "agent"],
+                "mean_score": [0.3, 0.9],  # agent inflated its own recorded score
+                "candidate_created_at": [1, 2],
+            }
+        )
+
+    @pytest.mark.asyncio
+    async def test_reverts_to_base_when_no_candidate_beats_baseline(self, tmp_path):
+        engine = MagicMock()
+        engine.db.get_experiments_df.return_value = self._df()
+
+        # agent admin-scores 0.2 on the selection split; base admin-scores 0.3;
+        # the reverted base scores 0.35 on the target split (distinct values so the
+        # assertions can tell the target eval apart from the floor comparison).
+        async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
+            if commit == "base":
+                score = 0.35 if split == "validation" else 0.3
+            else:
+                score = 0.2
+            return MagicMock(result=MagicMock(score=MagicMock(return_value=score)))
+
+        engine.evaluate_admin = AsyncMock(side_effect=_admin)
+        v = Verifier(
+            engine=engine,
+            admin_volume=tmp_path,
+            reward_mode="auto_best",
+            selection_split="train",
+            base_commit="base",
+            selection_task="math",
+            targets=[VerificationTarget(task="math", dataset_id="ds1", split="validation", reward_key="reward")],
+        )
+        result = await v.finalize()
+        # winner reverted to base -> the emitted reward is the SEED's target-split
+        # score, not the regressed candidate's
+        assert result["rewards"] == {"reward": 0.35}
+        rescored = [c.kwargs["commit"] for c in engine.evaluate_admin.await_args_list]
+        assert "base" in rescored  # base was admin-scored for the floor comparison
+        # the final call is the target eval of the reverted commit (validation split),
+        # not the floor comparison (train split)
+        assert engine.evaluate_admin.await_args.kwargs["commit"] == "base"
+        assert engine.evaluate_admin.await_args.kwargs["split"] == "validation"
+
+    @pytest.mark.asyncio
+    async def test_exact_tie_reverts_to_base(self, tmp_path):
+        # The floor uses '<=': a statistical tie reverts. If the optimizer cannot
+        # show an improvement, shipping the seed is the safe outcome. Pins the
+        # boundary so a refactor to '<' regresses loudly.
+        engine = MagicMock()
+        engine.db.get_experiments_df.return_value = self._df()
+
+        async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
+            return MagicMock(result=MagicMock(score=MagicMock(return_value=0.3)))  # all equal
+
+        engine.evaluate_admin = AsyncMock(side_effect=_admin)
+        v = Verifier(
+            engine=engine,
+            admin_volume=tmp_path,
+            reward_mode="auto_best",
+            selection_split="train",
+            base_commit="base",
+            selection_task="math",
+            targets=[VerificationTarget(task="math", dataset_id="ds1", split="validation", reward_key="reward")],
+        )
+        await v.finalize()
+        assert engine.evaluate_admin.await_args.kwargs["commit"] == "base"
+
+    @pytest.mark.asyncio
+    async def test_floor_noop_without_base_commit(self, tmp_path):
+        # floor on (default) but base_commit=None: the floor must silently no-op,
+        # never issuing an eval with commit=None, and the best candidate ships.
+        engine = MagicMock()
+        engine.db.get_experiments_df.return_value = pd.DataFrame(
+            {
+                "dataset_subset_split": ["train"],
+                "dataset_subset_dataset_id": ["ds1"],
+                "candidate_commit": ["agent"],
+                "mean_score": [0.9],
+                "candidate_created_at": [1],
+            }
+        )
+
+        async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
+            return MagicMock(result=MagicMock(score=MagicMock(return_value=0.5)))
+
+        engine.evaluate_admin = AsyncMock(side_effect=_admin)
+        v = Verifier(
+            engine=engine,
+            admin_volume=tmp_path,
+            reward_mode="auto_best",
+            selection_split="train",
+            selection_task="math",
+            targets=[VerificationTarget(task="math", dataset_id="ds1", split="validation", reward_key="reward")],
+        )
+        await v.finalize()
+        commits = [c.kwargs["commit"] for c in engine.evaluate_admin.await_args_list]
+        assert None not in commits
+        assert engine.evaluate_admin.await_args.kwargs["commit"] == "agent"
+
+    @pytest.mark.asyncio
+    async def test_keeps_candidate_that_beats_baseline(self, tmp_path):
+        engine = MagicMock()
+        engine.db.get_experiments_df.return_value = self._df()
+
+        async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
+            score = 0.3 if commit == "base" else 0.6  # agent genuinely improves
+            return MagicMock(result=MagicMock(score=MagicMock(return_value=score)))
+
+        engine.evaluate_admin = AsyncMock(side_effect=_admin)
+        v = Verifier(
+            engine=engine,
+            admin_volume=tmp_path,
+            reward_mode="auto_best",
+            selection_split="train",
+            base_commit="base",
+            selection_task="math",
+            targets=[VerificationTarget(task="math", dataset_id="ds1", split="validation", reward_key="reward")],
+        )
+        await v.finalize()
+        # 'agent' beats base -> it is selected and target-scored
+        assert engine.evaluate_admin.await_args.kwargs["commit"] == "agent"
+
+    @pytest.mark.asyncio
+    async def test_floor_off_ships_least_bad_candidate(self, tmp_path):
+        # With the floor disabled, the old behavior stands: the best candidate is
+        # shipped even if it did not beat the baseline (base is never scored).
+        engine = MagicMock()
+        engine.db.get_experiments_df.return_value = self._df()
+
+        async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
+            return MagicMock(result=MagicMock(score=MagicMock(return_value=0.2)))
+
+        engine.evaluate_admin = AsyncMock(side_effect=_admin)
+        v = Verifier(
+            engine=engine,
+            admin_volume=tmp_path,
+            reward_mode="auto_best",
+            selection_split="train",
+            base_commit="base",
+            selection_task="math",
+            auto_best_baseline_floor=False,
+            targets=[VerificationTarget(task="math", dataset_id="ds1", split="validation", reward_key="reward")],
+        )
+        await v.finalize()
+        rescored = [c.kwargs["commit"] for c in engine.evaluate_admin.await_args_list]
+        assert "base" not in rescored
+        assert engine.evaluate_admin.await_args.kwargs["commit"] == "agent"
+
+
 class TestNoCandidateFallback:
     """finalize() floors rewards when the optimizer produced no candidate.
 
@@ -173,7 +339,7 @@ async def test_auto_best_baseline_only_floors_rewards(self, tmp_path):
             base_commit="base",
             targets=[VerificationTarget(task=None, dataset_id="ds1", split="validation", reward_key="accuracy")],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"accuracy": 0.0}
         # no candidate -> nothing re-scored, no target eval spent
         engine.evaluate_admin.assert_not_awaited()
@@ -190,7 +356,7 @@ async def test_auto_best_no_experiments_floors_rewards(self, tmp_path):
             selection_split="train",
             targets=[VerificationTarget(task=None, dataset_id="ds1", split="validation", reward_key="accuracy")],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"accuracy": 0.0}
         engine.evaluate_admin.assert_not_awaited()
 
@@ -212,7 +378,8 @@ async def test_auto_best_missing_db_still_raises(self, tmp_path):
 
     @pytest.mark.asyncio
     async def test_candidates_present_keeps_normal_selection(self, tmp_path):
-        # Regression guard: the fallback must not swallow the normal path.
+        # Regression guard: the fallback must not swallow the normal path. Floor off
+        # so this isolates candidate selection (the floor is covered separately).
         engine = MagicMock()
         engine.db.get_experiments_df.return_value = pd.DataFrame(
             {
@@ -234,9 +401,10 @@ async def _admin(*, task, dataset_id, split, commit, sample_ids=None):
             reward_mode="auto_best",
             selection_split="train",
             base_commit="base",
+            auto_best_baseline_floor=False,
             targets=[VerificationTarget(task=None, dataset_id="ds1", split="validation", reward_key="accuracy")],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"accuracy": 0.5}
         assert engine.evaluate_admin.await_args.kwargs["commit"] == "agent"
 
@@ -260,8 +428,11 @@ async def test_baseline_scored_and_persisted(self, tmp_path):
             score_baseline=True,
             targets=[VerificationTarget(task=None, dataset_id="ds", split="validation", reward_key="accuracy")],
         )
-        rewards = await v.finalize()
-        assert rewards == {"accuracy": 0.2}  # reward.json content unchanged
+        result = await v.finalize()
+        assert result["rewards"] == {"accuracy": 0.2}  # reward.json content unchanged
+        # the baseline outcome is surfaced in the finalize response (durable channel:
+        # echoed to the trial stdout, which survives teardown; the admin volume does not)
+        assert result["baseline"]["scores"] == {"accuracy": 0.3}
         data = json.loads((tmp_path / "baseline.json").read_text())
         assert data == {"accuracy": 0.3}
         # second admin eval was the baseline commit
@@ -278,18 +449,50 @@ async def test_default_off_no_extra_evals(self, tmp_path):
             base_commit="base",
             targets=[VerificationTarget(task=None, dataset_id="ds", split="validation", reward_key="accuracy")],
         )
-        rewards = await v.finalize()
+        rewards = (await v.finalize())["rewards"]
         assert rewards == {"accuracy": 0.9}
         assert engine.evaluate_admin.await_count == 1
         assert not (tmp_path / "baseline.json").exists()
 
     @pytest.mark.asyncio
-    async def test_baseline_failure_never_fails_trial(self, tmp_path):
+    async def test_baseline_failure_retries_then_reports_error_without_failing_trial(self, tmp_path):
+        # The baseline eval fails on every attempt (2 by default). The trial reward
+        # must survive, AND the failure must be surfaced in the finalize response
+        # (not silently swallowed): a live trial once lost its baseline check because
+        # the only record was a log line that died with the container at teardown.
         (tmp_path / "submission.json").write_text(json.dumps({"commit": "cand"}))
         engine = MagicMock()
         engine.evaluate_admin = AsyncMock(
             side_effect=[MagicMock(result=MagicMock(score=MagicMock(return_value=0.7))),
-                         RuntimeError("modal down")]
+                         RuntimeError("modal down"),   # baseline attempt 1
+                         RuntimeError("modal down")]   # baseline attempt 2 (retry)
+        )
+        v = Verifier(
+            engine=engine,
+            admin_volume=tmp_path,
+            reward_mode="submit",
+            base_commit="base",
+            score_baseline=True,
+            targets=[VerificationTarget(task=None, dataset_id="ds", split="validation", reward_key="accuracy")],
+        )
+        result = await v.finalize()
+        assert result["rewards"] == {"accuracy": 0.7}  # trial reward survives baseline failure
+        assert result["baseline"]["error_type"] == "RuntimeError"
+        assert result["baseline"]["attempts"] == 2  # tried twice before reporting
+        # 1 target eval + 2 baseline attempts
+        assert engine.evaluate_admin.await_count == 3
+        assert not (tmp_path / "baseline.json").exists()  # nothing persisted on failure
+
+    @pytest.mark.asyncio
+    async def test_baseline_transient_failure_recovers_on_retry(self, tmp_path):
+        # A single transient blip on the baseline eval must not drop the check: the
+        # retry succeeds and the baseline score is reported normally.
+        (tmp_path / "submission.json").write_text(json.dumps({"commit": "cand"}))
+        engine = MagicMock()
+        engine.evaluate_admin = AsyncMock(
+            side_effect=[MagicMock(result=MagicMock(score=MagicMock(return_value=0.7))),  # target
+                         RuntimeError("transient"),                                        # baseline attempt 1
+                         MagicMock(result=MagicMock(score=MagicMock(return_value=0.5)))]   # baseline attempt 2 ok
         )
         v = Verifier(
             engine=engine,
@@ -299,8 +502,10 @@ async def test_baseline_failure_never_fails_trial(self, tmp_path):
             score_baseline=True,
             targets=[VerificationTarget(task=None, dataset_id="ds", split="validation", reward_key="accuracy")],
         )
-        rewards = await v.finalize()
-        assert rewards == {"accuracy": 0.7}  # trial reward survives baseline failure
+        result = await v.finalize()
+        assert result["rewards"] == {"accuracy": 0.7}
+        assert result["baseline"]["scores"] == {"accuracy": 0.5}
+        assert result["baseline"]["attempts"] == 2
 
     @pytest.mark.asyncio
     async def test_missing_base_commit_warns(self, tmp_path, caplog):
@@ -316,7 +521,7 @@ async def test_missing_base_commit_warns(self, tmp_path, caplog):
             targets=[VerificationTarget(task=None, dataset_id="ds", split="validation", reward_key="accuracy")],
         )
         with caplog.at_level("WARNING", logger="vero.harbor.verifier"):
-            rewards = await v.finalize()
+            rewards = (await v.finalize())["rewards"]
         assert rewards == {"accuracy": 0.9}
         assert not (tmp_path / "baseline.json").exists()
         assert any("base_commit is not set" in m for m in caplog.messages)