diff --git a/examples/timeseries/anomaly.py b/examples/timeseries/anomaly.py
index 047692301..23d4b81c9 100644
--- a/examples/timeseries/anomaly.py
+++ b/examples/timeseries/anomaly.py
@@ -9,10 +9,19 @@
 
 time = torch.arange(200, device=device, dtype=torch.float32)
 x = torch.stack([torch.sin(time / 12), torch.cos(time / 20)], dim=-1)
-x[140:149] += 3  # Add anomalies.
+x[140:150] += 3  # Add anomalies.
 
 model = sdm.models.KumoAnomaly(device=device)
+out = model(x)
 
-# Higher reconstruction error scores indicate more likely anomalies:
-score = model(x).numerical.squeeze(-1)
-print(score[131:158])
+# Higher MAE scores indicate more likely anomalies:
+print("Reconstruction Error (MAE):")
+print(out["mae"].numerical[130:160, 0])
+
+# SCS flags anomalies using adaptive segments of the preprocessed series:
+print("Anomaly Labels (SCS):")
+print(out["scs"].categorical.decode()[130:160, 0])
+
+# MACS flags anomalies using variance-weighted score bounds:
+print("Anomaly Labels (MACS):")
+print(out["macs"].categorical.decode()[130:160, 0])
diff --git a/sdm/models/kumo/timeseries/anomaly/model.py b/sdm/models/kumo/timeseries/anomaly/model.py
index f15f06a94..e2d669e17 100644
--- a/sdm/models/kumo/timeseries/anomaly/model.py
+++ b/sdm/models/kumo/timeseries/anomaly/model.py
@@ -14,7 +14,7 @@
 import torch.nn.functional as F
 from torch import Tensor
 
-from sdm import Recipe, Stype, TableTensor
+from sdm import CategoricalTensor, Recipe, Stype, TableTensor
 from sdm._inference import inference_mode
 from sdm.models import Callback, Model
 from sdm.models._huggingface import download_checkpoint
@@ -24,6 +24,10 @@
 from sdm.models.kumo.timeseries.anomaly.recipe import default_recipe
 from sdm.models.kumo.timeseries.anomaly.sampling import DPMSolver
 from sdm.models.kumo.timeseries.anomaly.schedule import DiffusionSchedule
+from sdm.models.kumo.timeseries.anomaly.threshold import (
+    macs_threshold,
+    scs_threshold,
+)
 from sdm.processing.execution import MemberQuery, RecipeExecution
 
 
@@ -99,7 +103,7 @@ def forward(
         num_steps: int = 20,
         callbacks: Sequence[Callback] | None = None,
         generator: torch.Generator | None = None,
-    ) -> TableTensor:  # [..., R, 1]
+    ) -> TableTensor:  # [..., R, 3]
         r"""The forward pass.
 
         Args:
@@ -116,9 +120,13 @@ def forward(
                 pre-processing and model execution.
 
         Returns:
-            The anomaly scores with shape ``[..., R, 1]``, denoting the mean
-            absolute reconstruction error across preprocessed features at each
-            time step.
+            A table with shape ``[..., R, 3]`` containing numerical ``mae``
+            reconstruction scores and categorical boolean ``scs`` and ``macs``
+            anomaly labels at each time step. SCS (Segmented Confidence
+            Sequences) segments the preprocessed series and thresholds scores
+            using expanding bounds within each segment. MACS (Multi-Scale
+            Adaptive Confidence Segments) uses variance-weighted expanding
+            score bounds.
         """
         callbacks = () if callbacks is None else callbacks
         requires_grad = self.training
@@ -158,9 +166,19 @@ def forward(
                 generator=generator,
             )
 
+            scs = scs_threshold(out, query.x.numerical).long()
+            macs = macs_threshold(out).long()
+            categories = torch.tensor([False, True], device=out.device)
             out = TableTensor(
-                columns={Stype.numerical: ("mae",)},
+                columns={
+                    Stype.numerical: ("mae",),
+                    Stype.categorical: ("scs", "macs"),
+                },
                 numerical=out.unsqueeze(-1),
+                categorical=CategoricalTensor(
+                    code=torch.stack([scs, macs], dim=-1),
+                    categories=(categories, categories),
+                ),
             )
 
             for callback in callbacks:
diff --git a/sdm/models/kumo/timeseries/anomaly/threshold.py b/sdm/models/kumo/timeseries/anomaly/threshold.py
new file mode 100644
index 000000000..072e19ed3
--- /dev/null
+++ b/sdm/models/kumo/timeseries/anomaly/threshold.py
@@ -0,0 +1,226 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from NVIDIA/Kumo-TS, Kumo-Anomaly/utils/adaptive_threshold.py
+# at af74798bbf7d020d3298f545fb9be383560a18ce.
+
+import torch
+from torch import Tensor
+
+from sdm import TableTensor
+
+
+def _bounds(
+    scores: Tensor, starts: Tensor, confidence_level: float
+) -> tuple[Tensor, Tensor]:
+    """Compute the SDK's expanding bounds, restarting at segment starts."""
+    sums = torch.cat(
+        [torch.zeros_like(scores[..., :1]), scores.cumsum(-1)], dim=-1
+    )
+    squares = torch.cat(
+        [torch.zeros_like(scores[..., :1]), scores.square().cumsum(-1)], dim=-1
+    )
+    counts = (
+        torch.arange(1, scores.size(-1) + 1, device=scores.device) - starts
+    )
+    mean = (sums[..., 1:] - sums.gather(-1, starts)) / counts
+    mean = torch.where(counts == 1, scores, mean)
+    variance = (
+        squares[..., 1:] - squares.gather(-1, starts)
+    ) / counts - mean.square()
+    std = variance.clamp_min(0).sqrt()
+    width = torch.where(std > 0, 1.5 * std, 0.1)
+    if confidence_level > 0.95:
+        width = width * 1.2
+    elif confidence_level < 0.90:
+        width = width * 0.8
+    # The first score in each segment always has unscaled +/- 0.1 bounds.
+    width = torch.where(counts == 1, 0.1, width)
+    return mean - width, mean + width
+
+
+def _rolling_stats(scores: Tensor, window: int) -> tuple[Tensor, Tensor]:
+    """Return trailing means and sample variances with min_periods=1."""
+    # Center before prefix sums to reduce cancellation for offset scores.
+    centered = scores - scores[..., :1]
+    sums = torch.cat(
+        [torch.zeros_like(scores[..., :1]), centered.cumsum(-1)], dim=-1
+    )
+    squares = torch.cat(
+        [torch.zeros_like(scores[..., :1]), centered.square().cumsum(-1)],
+        dim=-1,
+    )
+    end = torch.arange(1, scores.size(-1) + 1, device=scores.device)
+    start = (end - window).clamp_min(0)
+    count = end - start
+    total = sums[..., end] - sums[..., start]
+    mean = total / count + scores[..., :1]
+    variance = squares[..., end] - squares[..., start] - total.square() / count
+    variance = variance.clamp_min(0) / (count - 1).clamp_min(1)
+    return mean, torch.where(count > 1, variance, 0)
+
+
+def _segment_starts(data: Tensor, min_segment_size: int) -> Tensor:
+    """Return each row's segment start using the published APCA rule."""
+    length = data.numel()
+    rows = torch.arange(length, device=data.device)
+    variation = data.std(correction=0) / (data.mean() + 1e-8)
+    # Preserve the signed coefficient of variation used by the SDK.
+    if variation < 0.1:
+        size = max(200, length // 15)
+        return rows // size * size
+
+    starts = data.new_zeros(length, dtype=torch.long)
+    sums = torch.cat([data.new_zeros(1), data.cumsum(0)])
+    squares = torch.cat([data.new_zeros(1), data.square().cumsum(0)])
+    start = 0
+    while start < length:
+        starts[start:] = start
+        if length - start < 2 * min_segment_size:
+            break
+        first = start + min_segment_size
+        last = length - min_segment_size
+        if last <= first:
+            break
+        step = max(1, (last - first) // 200)
+        splits = torch.arange(first, last, step, device=data.device)
+        left_sum = sums[splits] - sums[start]
+        right_sum = sums[-1] - sums[splits]
+        error = (
+            squares[-1]
+            - squares[start]
+            - left_sum.square() / (splits - start)
+            - right_sum.square() / (length - splits)
+        )
+        best = error.argmin()
+        no_split = (
+            squares[-1]
+            - squares[start]
+            - (sums[-1] - sums[start]).square() / (length - start)
+        )
+        improvement = 0.7 if variation > 0.3 else 0.5
+        if error[best] >= no_split * improvement:
+            break
+        start = int(splits[best])
+    return starts
+
+
+@torch.no_grad()
+def scs_threshold(
+    scores: Tensor,
+    target: Tensor | TableTensor,
+    *,
+    confidence_level: float = 0.99,
+    min_segment_size: int = 50,
+    percentile_filter: float = 95.0,
+) -> Tensor:
+    """Flag anomalies using Segmented Confidence Sequences (SCS).
+
+    Segment preprocessed observations with APCA, compute expanding score
+    bounds per segment, and retain violations above a whole-series percentile.
+    Defaults follow the Kumo-Anomaly SDK. Decisions use future observations;
+    the confidence setting is a heuristic, not a calibrated guarantee.
+    Segmentation copies feature means to CPU; flags stay on the score device.
+
+    Args:
+        scores: Finite reconstruction scores with shape ``[..., R]``.
+        target: Finite preprocessed observations of shape ``[..., R, D]``
+            without padded features. ``R`` is the number of time steps and
+            ``D`` the number of features.
+        confidence_level: Bound width setting.
+        min_segment_size: Minimum observations on each side of an APCA split.
+        percentile_filter: Global score percentile used as a strict cutoff.
+
+    Returns:
+        Boolean anomaly flags with shape ``[..., R]``.
+    """
+    scores = scores.double()
+    data = target.numerical if isinstance(target, TableTensor) else target
+    if data.shape[:-1] != scores.shape:
+        raise ValueError("Target batch and row dimensions must match scores")
+    if scores.numel() == 0:
+        return scores.bool()
+    data = data.mean(-1).to(device="cpu", dtype=torch.float64)
+    starts = (
+        torch.stack(
+            [
+                _segment_starts(series, min_segment_size)
+                for series in data.reshape(-1, data.size(-1))
+            ]
+        )
+        .reshape(scores.shape)
+        .to(scores.device)
+    )
+    lower, upper = _bounds(scores, starts, confidence_level)
+    cutoff = scores.quantile(percentile_filter / 100, dim=-1, keepdim=True)
+    return ((scores < lower) | (scores > upper)) & (scores > cutoff)
+
+
+@torch.no_grad()
+def macs_threshold(
+    scores: Tensor,
+    *,
+    short_window: int = 10,
+    medium_window: int = 50,
+    long_window: int = 250,
+    confidence_level: float = 0.95,
+    percentile_filter: float = 97.0,
+) -> Tensor:
+    """Flag anomalies using Multi-Scale Adaptive Confidence Segments (MACS).
+
+    Combine expanding score bounds with variance-based weights and a
+    regime-change check, then retain violations above a whole-series
+    percentile. Defaults follow the SDK, including its shared expanding bounds
+    across scales. Decisions use future observations; the confidence setting
+    is a heuristic, not a calibrated guarantee.
+
+    Args:
+        scores: Finite reconstruction scores with shape ``[..., R]``,
+            where ``R`` is the number of time steps.
+        short_window: Window used for variance-based attention weights.
+        medium_window: SDK scale setting; bounds remain expanding.
+        long_window: Window used to identify regime changes.
+        confidence_level: Bound width setting.
+        percentile_filter: Global score percentile used as a strict cutoff.
+
+    Returns:
+        Boolean anomaly flags with shape ``[..., R]``.
+    """
+    scores = scores.double()
+    if scores.numel() == 0:
+        return scores.bool()
+    starts = torch.zeros_like(scores, dtype=torch.long)
+    lower, upper = _bounds(scores, starts, confidence_level)
+    window = max(2, min(short_window, scores.size(-1) // 10))
+    _, variance = _rolling_stats(scores, window)
+    maximum = variance.amax(dim=-1, keepdim=True)
+    normalized = variance / torch.where(maximum > 0, maximum, 1)
+    group = torch.where(normalized > 0.7, 2, (normalized > 0.3).long())
+    weights = scores.new_tensor(
+        [[0.1, 0.3, 0.6], [0.2, 0.6, 0.2], [0.6, 0.3, 0.1]]
+    )[group]
+    short_weight, medium_weight, long_weight = weights.unbind(-1)
+    # Keep the SDK's weighted arithmetic even though the bounds coincide.
+    combined_lower = (
+        short_weight * lower + medium_weight * lower + long_weight * lower
+    )
+    combined_upper = (
+        short_weight * upper + medium_weight * upper + long_weight * upper
+    )
+    mean, variance = _rolling_stats(scores, long_window)
+    std = variance.sqrt()
+    previous_mean = torch.cat(
+        [torch.zeros_like(scores[..., :1]), mean[..., :-1]], dim=-1
+    )
+    previous_std = torch.cat(
+        [torch.zeros_like(scores[..., :1]), std[..., :-1]], dim=-1
+    )
+    denominator = previous_std + 1e-8
+    regime = ((mean - previous_mean).abs() / denominator > 2) | (
+        (std - previous_std).abs() / denominator > 1.5
+    )
+    violations = (scores < lower) | (scores > upper)
+    attention = (scores < combined_lower) | (scores > combined_upper)
+    anomalies = attention & (~regime | violations)
+    cutoff = scores.quantile(percentile_filter / 100, dim=-1, keepdim=True)
+    return anomalies & (scores > cutoff)
diff --git a/test/models/kumo/timeseries/anomaly/test_model.py b/test/models/kumo/timeseries/anomaly/test_model.py
index 849dd2874..6e1c6c71e 100644
--- a/test/models/kumo/timeseries/anomaly/test_model.py
+++ b/test/models/kumo/timeseries/anomaly/test_model.py
@@ -13,8 +13,13 @@ def test_forward(device: torch.device) -> None:
     x = torch.randn(2, 130, 3, device=device)
 
     out = model(x, num_samples=1, num_steps=1)
-    assert out.size() == (2, 130, 1)
+    assert out.size() == (2, 130, 3)
     assert out.columns["numerical"] == ("mae",)
+    assert out.columns["categorical"] == ("scs", "macs")
     assert out.device == device
     assert out.numerical.isfinite().all()
     assert (out.numerical >= 0).all()
+    for categories in out.categorical.categories:
+        torch.testing.assert_close(
+            categories, torch.tensor([False, True], device=device)
+        )
diff --git a/test/models/kumo/timeseries/anomaly/test_threshold.py b/test/models/kumo/timeseries/anomaly/test_threshold.py
new file mode 100644
index 000000000..700dc5d30
--- /dev/null
+++ b/test/models/kumo/timeseries/anomaly/test_threshold.py
@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+
+from sdm.models.kumo.timeseries.anomaly.threshold import (
+    macs_threshold,
+    scs_threshold,
+)
+from sdm.testing import withCUDA
+
+
+@withCUDA
+def test_threshold(device: torch.device) -> None:
+    scores = torch.full((2, 400), 0.1, device=device)
+    scores[0, [0, 50, 200, 350]] = 2
+    target = torch.ones(2, 400, 1, device=device)
+
+    expected = torch.zeros_like(scores, dtype=torch.bool)
+    expected[0, [50, 350]] = True
+    torch.testing.assert_close(scs_threshold(scores, target), expected)
+
+    expected[0, 200] = True
+    torch.testing.assert_close(macs_threshold(scores), expected)