diff --git a/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/sf.py b/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/sf.py
index 6a1a1228..a9e4a607 100644
--- a/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/sf.py
+++ b/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/sf.py
@@ -1,4 +1,3 @@
-# ty: ignore
 # ruff: ignore
 """Cromer-Mann coefficents and related constants
 
diff --git a/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/transformer.py b/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/transformer.py
index 90f92cae..d8367179 100644
--- a/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/transformer.py
+++ b/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/transformer.py
@@ -1,5 +1,3 @@
-# ty: ignore
-
 """Minimal components of qFit's transformer module necessary to implement
 qFit's volume.py
 
diff --git a/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/unitcell.py b/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/unitcell.py
index 582d2beb..23988e6b 100644
--- a/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/unitcell.py
+++ b/src/sampleworks/core/forward_models/xray/real_space_density_deps/qfit/unitcell.py
@@ -1,6 +1,3 @@
-# type: ignore
-# ty: ignore
-
 """Classes for handling unit cell transformation."""
 
 from itertools import product
diff --git a/src/sampleworks/core/samplers/edm.py b/src/sampleworks/core/samplers/edm.py
index 81456170..7b94fdaa 100644
--- a/src/sampleworks/core/samplers/edm.py
+++ b/src/sampleworks/core/samplers/edm.py
@@ -4,7 +4,7 @@
 
 from collections.abc import Mapping
 from dataclasses import dataclass
-from typing import TYPE_CHECKING
+from typing import cast, TYPE_CHECKING
 
 import einx
 import torch
@@ -395,7 +395,8 @@ def step(
 
         t_hat = context.t_effective
         dt = context.dt
-        eps_scale = context.noise_scale
+        # check_context() guarantees noise_scale is present for sampling steps.
+        eps_scale = cast(float, context.noise_scale)
         allow_gradients = True if scaler and getattr(scaler, "requires_gradients", False) else False
 
         centroid = einx.mean("... [n] c", state)
@@ -415,7 +416,7 @@ def step(
 
         # Store eps separately for proper frame transformation
         # eps_scale will be float if check_context didn't raise
-        eps = torch.randn_like(maybe_augmented_state) * eps_scale  # ty: ignore[unsupported-operator]
+        eps = torch.randn_like(maybe_augmented_state) * eps_scale
         noisy_state = maybe_augmented_state + eps
         noisy_state = torch.as_tensor(noisy_state).detach().requires_grad_(allow_gradients)
 
@@ -507,7 +508,7 @@ def step(
             # multiple particles.
             # Only compute when noise_var > 0 to avoid division by near-zero
             # (matching Boltz behavior)
-            noise_var = eps_scale**2  # ty: ignore[unsupported-operator]
+            noise_var = eps_scale**2
             if noise_var > 0:
                 log_proposal_correction = einx.sum(
                     "... [b n c]", eps_working_frame**2 - (eps_working_frame + proposal_shift) ** 2
diff --git a/src/sampleworks/metrics/lddt.py b/src/sampleworks/metrics/lddt.py
index 8602e755..ca49617f 100644
--- a/src/sampleworks/metrics/lddt.py
+++ b/src/sampleworks/metrics/lddt.py
@@ -336,11 +336,13 @@ def compute(
         # Set token ids to something useful for residue-level LDDT (chain ID + residue number)
         # Note: add_global_token_id_annotation works with both AtomArray and
         # AtomArrayStack at runtime
-        predicted_atom_array_stack = add_global_token_id_annotation(
-            predicted_atom_array_stack  # ty: ignore[invalid-argument-type]
+        predicted_atom_array_stack = cast(
+            AtomArrayStack,
+            add_global_token_id_annotation(cast(Any, predicted_atom_array_stack)),
         )
-        ground_truth_atom_array_stack = add_global_token_id_annotation(
-            ground_truth_atom_array_stack  # ty: ignore[invalid-argument-type]
+        ground_truth_atom_array_stack = cast(
+            AtomArrayStack,
+            add_global_token_id_annotation(cast(Any, ground_truth_atom_array_stack)),
         )
 
         # restrict to atoms that are present in both structures
@@ -477,11 +479,13 @@ def compute(
 
             # set the token ids, to avoid any possible confusion later on
             # Note: add_global_token_id_annotation works with AtomArrayStack at runtime
-            filtered_predicted = add_global_token_id_annotation(
-                filtered_predicted  # ty: ignore[invalid-argument-type]
+            filtered_predicted = cast(
+                AtomArrayStack,
+                add_global_token_id_annotation(cast(Any, filtered_predicted)),
             )
-            filtered_ground_truth = add_global_token_id_annotation(
-                filtered_ground_truth  # ty: ignore[invalid-argument-type]
+            filtered_ground_truth = cast(
+                AtomArrayStack,
+                add_global_token_id_annotation(cast(Any, filtered_ground_truth)),
             )
 
             lddt_features = extract_lddt_features_from_atom_arrays(
diff --git a/src/sampleworks/metrics/rmsd.py b/src/sampleworks/metrics/rmsd.py
index e3362650..7bd90020 100644
--- a/src/sampleworks/metrics/rmsd.py
+++ b/src/sampleworks/metrics/rmsd.py
@@ -92,11 +92,13 @@ def compute(
             - ``best_of_{N}_segment_rmsd``: minimum of ``segment_rmsd``.
         """
         # 1. Annotate token IDs so atoms can be grouped into residues downstream.
-        predicted_atom_array_stack = add_global_token_id_annotation(
-            predicted_atom_array_stack  # ty: ignore[invalid-argument-type]  (accepts AtomArray|AtomArrayStack at runtime; stub is narrower)
+        predicted_atom_array_stack = cast(
+            AtomArrayStack,
+            add_global_token_id_annotation(cast(Any, predicted_atom_array_stack)),
         )
-        ground_truth_atom_array_stack = add_global_token_id_annotation(
-            ground_truth_atom_array_stack  # ty: ignore[invalid-argument-type]  (accepts AtomArray|AtomArrayStack at runtime; stub is narrower)
+        ground_truth_atom_array_stack = cast(
+            AtomArrayStack,
+            add_global_token_id_annotation(cast(Any, ground_truth_atom_array_stack)),
         )
 
         # 2. Restrict both stacks to atoms present in both structures, in matching order.
diff --git a/src/sampleworks/models/boltz/wrapper.py b/src/sampleworks/models/boltz/wrapper.py
index d75b6ab5..5254341a 100644
--- a/src/sampleworks/models/boltz/wrapper.py
+++ b/src/sampleworks/models/boltz/wrapper.py
@@ -454,7 +454,7 @@ def create_boltz_input_from_structure(
             sequence_to_chains.setdefault(seq, []).append(chain_id)
 
         unique_chain_sequences = {chains[0]: seq for seq, chains in sequence_to_chains.items()}
-        msa_paths_unique = msa_manager.get_msa(unique_chain_sequences, msa_pairing_strategy)  # ty: ignore[invalid-argument-type]
+        msa_paths_unique = msa_manager.get_msa(unique_chain_sequences, msa_pairing_strategy)
 
         msa_paths = {}
         for seq, chains_with_seq in sequence_to_chains.items():
@@ -739,7 +739,9 @@ def featurize(self, structure: dict) -> GenerativeModelInput[BoltzConditioning]:
                 "x_init from prior. This means align_to_input will not work properly,"
                 " and reward functions dependent on this won't be accurate."
             )
-            temp_features = GenerativeModelInput(x_init=None, conditioning=conditioning)  # ty: ignore[invalid-argument-type]
+            temp_features = GenerativeModelInput[BoltzConditioning](
+                x_init=cast(Any, None), conditioning=conditioning
+            )
             x_init = self.initialize_from_prior(batch_size=ensemble_size, features=temp_features)
 
         return GenerativeModelInput(x_init=x_init, conditioning=conditioning)
@@ -810,7 +812,7 @@ def _pairformer_pass(
             z = z + msa_module(z, s_inputs, features, use_kernels=self.model.use_kernels)
 
             if self.model.is_pairformer_compiled:
-                pairformer_module = self.model.pairformer_module._orig_mod
+                pairformer_module = cast(Any, self.model.pairformer_module)._orig_mod
             else:
                 pairformer_module = self.model.pairformer_module
 
@@ -1078,8 +1080,9 @@ def _setup_data_module(
         )
 
         processed_dir = out_dir / "processed"
+        load_manifest = cast(Any, Manifest.load)
         processed = BoltzProcessedInput(
-            manifest=Manifest.load(processed_dir / "manifest.json"),
+            manifest=load_manifest(processed_dir / "manifest.json"),
             targets_dir=processed_dir / "structures",
             msa_dir=processed_dir / "msa",
             constraints_dir=(processed_dir / "constraints")
@@ -1194,7 +1197,9 @@ def featurize(self, structure: dict) -> GenerativeModelInput[BoltzConditioning]:
                 "x_init from prior. This means align_to_input will not work properly,"
                 " and reward functions dependent on this won't be accurate."
             )
-            temp_features = GenerativeModelInput(x_init=None, conditioning=conditioning)  # ty: ignore[invalid-argument-type]
+            temp_features = GenerativeModelInput[BoltzConditioning](
+                x_init=cast(Any, None), conditioning=conditioning
+            )
             x_init = self.initialize_from_prior(batch_size=ensemble_size, features=temp_features)
 
         return GenerativeModelInput(x_init=x_init, conditioning=conditioning)
@@ -1368,7 +1373,7 @@ def _pairformer_pass(
                 )
 
             if self.model.is_pairformer_compiled:
-                pairformer_module = self.model.pairformer_module._orig_mod
+                pairformer_module = cast(Any, self.model.pairformer_module)._orig_mod
             else:
                 pairformer_module = self.model.pairformer_module
 
diff --git a/src/sampleworks/models/rf3/wrapper.py b/src/sampleworks/models/rf3/wrapper.py
index 023371f6..f4fc6771 100644
--- a/src/sampleworks/models/rf3/wrapper.py
+++ b/src/sampleworks/models/rf3/wrapper.py
@@ -365,8 +365,10 @@ def featurize(self, structure: dict) -> GenerativeModelInput[RF3Conditioning]:
             InferenceInput, input_batch[0]
         )  # since we're not batching, the loader returns a list of length 1
 
-        # (Hydra instantiation of pipeline means it is going to be hard to type check here)
-        pipeline_output = self.inference_engine.pipeline(input_spec.to_pipeline_input())  # ty: ignore[call-non-callable]
+        # Hydra instantiation leaves pipeline imprecisely typed even though it is callable
+        # at runtime.
+        pipeline = cast(Any, self.inference_engine.pipeline)
+        pipeline_output = pipeline(input_spec.to_pipeline_input())
         pipeline_output = trainer.fabric.to_device(pipeline_output)
 
         features = trainer._assemble_network_inputs(pipeline_output)
diff --git a/src/sampleworks/utils/framework_utils.py b/src/sampleworks/utils/framework_utils.py
index fa5d61b3..143be4c0 100644
--- a/src/sampleworks/utils/framework_utils.py
+++ b/src/sampleworks/utils/framework_utils.py
@@ -1,6 +1,6 @@
 from collections.abc import Callable
 from functools import wraps
-from typing import Any, overload, ParamSpec, TYPE_CHECKING, TypeVar
+from typing import Any, cast, overload, ParamSpec, TYPE_CHECKING, TypeVar
 
 import numpy as np
 
@@ -184,8 +184,8 @@ def match_batch(array: Array | np.ndarray, target_batch_size: int) -> Array | np
 
     # singleton: lazy broadcast (no copy)
     if b == 1:
-        return _broadcast_to(array, (n, *array.shape[1:]))  # ty: ignore[invalid-argument-type]
+        return _broadcast_to(cast(Any, array), (n, *array.shape[1:]))
     # divisible: tile
     if n % b:
         raise ValueError(f"batch {b} not divisible into target {n}")
-    return _tile(array, (n // b, *(1,) * (array.ndim - 1)))  # ty: ignore[invalid-argument-type]
+    return _tile(cast(Any, array), (n // b, *(1,) * (array.ndim - 1)))
diff --git a/src/sampleworks/utils/msa.py b/src/sampleworks/utils/msa.py
index 83e28052..5ede9910 100644
--- a/src/sampleworks/utils/msa.py
+++ b/src/sampleworks/utils/msa.py
@@ -1,3 +1,4 @@
+from collections.abc import Mapping
 from hashlib import sha3_256
 from pathlib import Path
 
@@ -18,6 +19,14 @@
 
 MAX_PAIRED_SEQS = 8192
 MAX_MSA_SEQS = 16384
+MSAData = Mapping[str, str] | Mapping[int, str] | Mapping[str | int, str]
+
+
+def _msa_data_key_sort_key(key: str | int) -> tuple[int, int | str]:
+    """Return a deterministic sort key for supported MSA sequence keys."""
+    if isinstance(key, int):
+        return (0, key)
+    return (1, key)
 
 
 def _validate_msa_cache_contents(msa_hash: str, msa_dir: Path) -> None:
@@ -104,7 +113,7 @@ def _validate_msa_cache_contents(msa_hash: str, msa_dir: Path) -> None:
 #   For the love of decent code, don't copy this and use it somewhere else and respect the
 #   leading underscore!
 def _compute_msa(
-    data: dict[str | int, str],
+    data: MSAData,
     target_id: str,
     msa_dir: Path,
     msa_server_url: str,
@@ -118,7 +127,7 @@ def _compute_msa(
 
     Parameters
     ----------
-    data : dict[str | int, str]
+    data : MSAData
         The input protein sequences.
     target_id : str
         The target id.
@@ -192,7 +201,7 @@ def _compute_msa(
     #  order as they are in `data`, and furthermore just returns a list of strings, the content
     #  of each string being a single sequence alignment. It's some weird file parsing that we
     #  should clean up so users don't break it or have to worry about it.
-    outputs = {}
+    outputs: dict[str | int, Path] = {}
     for idx, name in enumerate(data):
         # Get paired sequences
         paired = paired_msas[idx].strip().splitlines()
@@ -308,14 +317,14 @@ def __init__(
         self._cache_hits = 0
 
     @staticmethod
-    def _hash_arguments(data: dict[str | int, str], msa_pairing_strategy: str) -> str:
+    def _hash_arguments(data: MSAData, msa_pairing_strategy: str) -> str:
         encoded_sequence_tuple = str.encode(str(tuple(data.values())) + msa_pairing_strategy)
         hexdigest = sha3_256(encoded_sequence_tuple).hexdigest()
         return hexdigest
 
     def get_msa(
         self,
-        data: dict[str | int, str],
+        data: MSAData,
         msa_pairing_strategy: str,
         structure_predictor: str | StructurePredictor = StructurePredictor.BOLTZ_2,
     ) -> dict[str | int, Path]:
@@ -323,7 +332,7 @@ def get_msa(
 
         Parameters
         ----------
-        data : dict[str | int, str]
+        data : MSAData
             A dictionary mapping target (usu. chain or index) names to protein sequences.
         msa_pairing_strategy : str
             The MSA pairing strategy to use (usually "greedy").
@@ -382,8 +391,9 @@ def get_msa(
             protenix_dir.mkdir(parents=True, exist_ok=True)
             # Protenix adds extra information, easiest just to use their pipeline.
             # make sure sort order stays the same:
-            data_keys = sorted(data.keys())
-            sequences = [data[key] for key in data_keys]
+            data_items = sorted(data.items(), key=lambda item: _msa_data_key_sort_key(item[0]))
+            data_keys = [key for key, _ in data_items]
+            sequences = [sequence for _, sequence in data_items]
             out_dir = self.msa_dir / "protenix" / hash_key
 
             msa_directories = [out_dir / str(idx) for idx in data_keys]
@@ -392,7 +402,7 @@ def get_msa(
                 (out_dir / str(idx) / fn).exists() for idx in data_keys for fn in reqd_files
             )
             if need_msas:
-                msa_directories = protenix_msa_search(sequences, out_dir, mode="protenix")
+                msa_directories = protenix_msa_search(sequences, str(out_dir), mode="protenix")
                 self._api_calls += 1
             else:
                 self._cache_hits += 1
diff --git a/tests/conftest.py b/tests/conftest.py
index 4fba62e4..e84d889e 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -1089,5 +1089,7 @@ def perturbed_coords(
     """
     torch.manual_seed(42)
     base = converging_mock_wrapper.target
-    perturbation = torch.randn_like(base) * 0.1  # ty: ignore[invalid-argument-type]
-    return base, base + perturbation  # ty: ignore[invalid-return-type, unsupported-operator]
+    if base is None:
+        raise ValueError("converging_mock_wrapper must provide target coordinates")
+    perturbation = torch.randn_like(base) * 0.1
+    return base, base + perturbation
diff --git a/tests/utils/test_atom_array_utils.py b/tests/utils/test_atom_array_utils.py
index 76f0f2b1..47271c2a 100644
--- a/tests/utils/test_atom_array_utils.py
+++ b/tests/utils/test_atom_array_utils.py
@@ -1,6 +1,6 @@
 """Tests for atom_array_utils module."""
 
-from typing import cast
+from typing import Any, cast
 
 import numpy as np
 import pytest
@@ -200,12 +200,12 @@ def test_invalid_type_raises_error(self):
         invalid_input = "not an atom array"
 
         with pytest.raises(TypeError, match="can only accept AtomArray or AtomArrayStack"):
-            select_altloc(invalid_input, "A")  # ty: ignore[invalid-argument-type]
+            select_altloc(cast(Any, invalid_input), "A")
 
     def test_none_input_raises_error(self):
         """Test that None input raises TypeError."""
         with pytest.raises(TypeError, match="can only accept AtomArray or AtomArrayStack"):
-            select_altloc(None, "A")  # ty: ignore[invalid-argument-type]
+            select_altloc(cast(Any, None), "A")
 
 
 class TestSelectAltlocEdgeCases:
@@ -358,7 +358,7 @@ def test_invalid_type_first_arg(self):
         array.coord = np.random.rand(3, 3)
 
         with pytest.raises(TypeError, match="must be AtomArray or AtomArrayStack"):
-            filter_to_common_atoms("not an array", array)  # ty: ignore[invalid-argument-type]
+            filter_to_common_atoms(cast(Any, "not an array"), array)
 
     def test_invalid_type_second_arg(self):
         """Test that invalid second argument raises TypeError."""
@@ -366,7 +366,7 @@ def test_invalid_type_second_arg(self):
         array.coord = np.random.rand(3, 3)
 
         with pytest.raises(TypeError, match="must be AtomArray or AtomArrayStack"):
-            filter_to_common_atoms(array, None)  # ty: ignore[invalid-argument-type]
+            filter_to_common_atoms(array, cast(Any, None))
 
     def test_preserves_coordinates(self, atom_array_partial_overlap):
         """Test that coordinates are preserved for common atoms."""
@@ -443,7 +443,7 @@ def test_is_idempotent(self, structure_fixture, request):
 
     def test_invalid_type_raises_error(self):
         with pytest.raises(TypeError, match="can only accept AtomArray or AtomArrayStack"):
-            remove_hydrogens("bad input")  # ty:ignore[invalid-argument-type]
+            remove_hydrogens(cast(Any, "bad input"))
 
 
 class TestKeepPolymer:
@@ -467,7 +467,7 @@ def test_atom_array_and_stack_give_same_atom_count(self, test_structure):
 
     def test_invalid_type_raises_error(self):
         with pytest.raises(TypeError, match="can only accept AtomArray or AtomArrayStack"):
-            keep_polymer(42)  # ty:ignore[invalid-argument-type]
+            keep_polymer(cast(Any, 42))
 
 
 class TestKeepAminoAcids:
@@ -494,7 +494,7 @@ def test_atom_array_and_stack_give_same_atom_count(self, test_structure):
 
     def test_invalid_type_raises_error(self):
         with pytest.raises(TypeError, match="can only accept AtomArray or AtomArrayStack"):
-            keep_amino_acids(None)  # ty:ignore[invalid-argument-type]
+            keep_amino_acids(cast(Any, None))
 
 
 class TestFilterFunctionsIntegration:
diff --git a/tests/utils/test_framework_utils.py b/tests/utils/test_framework_utils.py
index 1738b6ae..5905cf3d 100644
--- a/tests/utils/test_framework_utils.py
+++ b/tests/utils/test_framework_utils.py
@@ -1,5 +1,7 @@
 """Tests for framework conversion utilities from sampleworks.utils.framework_utils."""
 
+from typing import Any, cast
+
 import numpy as np
 import pytest
 from sampleworks.utils.framework_utils import (
@@ -527,7 +529,7 @@ def test_raises_on_incompatible_batch_sizes(self):
     def test_raises_on_scalar_input(self):
         scalar = np.float64(1.0)
         with pytest.raises(ValueError, match="ndim >= 1"):
-            match_batch(scalar, target_batch_size=2)  # ty: ignore[no-matching-overload]
+            match_batch(cast(Any, scalar), target_batch_size=2)
 
     def test_1d_array(self):
         array = np.array([42.0])
@@ -549,4 +551,4 @@ def test_preserves_dtype(self):
 
     def test_raises_on_unsupported_type(self):
         with pytest.raises(TypeError, match="unsupported array type"):
-            match_batch([1, 2, 3], target_batch_size=2)  # ty: ignore[no-matching-overload]
+            match_batch(cast(Any, [1, 2, 3]), target_batch_size=2)
diff --git a/tests/utils/test_msa.py b/tests/utils/test_msa.py
index 412fb5d0..25f41519 100644
--- a/tests/utils/test_msa.py
+++ b/tests/utils/test_msa.py
@@ -8,7 +8,7 @@
 import pytest
 from sampleworks.utils import msa as msa_module
 from sampleworks.utils.guidance_constants import StructurePredictor
-from sampleworks.utils.msa import _compute_msa, MSAManager
+from sampleworks.utils.msa import _compute_msa, _msa_data_key_sort_key, MSAManager
 
 
 # ============================================================================
@@ -30,6 +30,13 @@ def test_hash_arguments_is_deterministic_and_input_sensitive():
     assert base != MSAManager._hash_arguments({"B": "GGGAA", "A": "MKTAY"}, "greedy")
 
 
+def test_msa_data_key_sort_preserves_numeric_order_and_handles_mixed_keys():
+    """Ensure Protenix MSA key sorting preserves numeric order and mixed-key stability."""
+    keys = [10, 2, "1", "A"]
+
+    assert sorted(keys, key=_msa_data_key_sort_key) == [2, 10, "1", "A"]
+
+
 # ============================================================================
 # __init__ cache directory handling (test 4)
 # ============================================================================