glitchlings 0.2.2__tar.gz → 0.2.3__tar.gz

{glitchlings-0.2.2 → glitchlings-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: glitchlings
-Version: 0.2.2
+Version: 0.2.3
 Summary: Monsters for your language games.
 Author: osoleve
 License:                                  Apache License
@@ -215,6 +215,8 @@ Classifier: Intended Audience :: Developers
 Classifier: License :: OSI Approved :: Apache Software License
 Classifier: Programming Language :: Python
 Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Rust
 Classifier: Operating System :: MacOS :: MacOS X
@@ -223,7 +225,7 @@ Classifier: Operating System :: POSIX :: Linux
 Classifier: Operating System :: OS Independent
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development :: Testing
-Requires-Python: >=3.12
+Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: confusable-homoglyphs>=3.3.1
@@ -280,14 +282,16 @@ After all, what good is general intelligence if it can't handle a little chaos?
 pip install -U glitchlings
 ```
 
+> Glitchlings requires Python 3.10 or newer.
+
 ```python
 from glitchlings import Gaggle, SAMPLE_TEXT, Typogre, Mim1c, Reduple, Rushmore
 
 gaggle = Gaggle([
-    Typogre(max_change_rate=0.03),
-    Mim1c(replacement_rate=0.02),
+    Typogre(rate=0.03),
+    Mim1c(rate=0.02),
     Reduple(seed=404),
-    Rushmore(max_deletion_rate=0.02),
+    Rushmore(rate=0.02),
 ])
 
 print(gaggle(SAMPLE_TEXT))
@@ -295,41 +299,10 @@ print(gaggle(SAMPLE_TEXT))
 
 > Onҽ m‎ھ‎rning, wһen Gregor Samƽa woke from trouble𝐝 𝑑reams, he found himself transformed in his bed into a horrible vermin‎٠‎ He l   lay on his armour-like back, and if he lifted his head a little he could see his brown belly, slightlh domed and divided by arches ino stiff sections. The bedding was adly able to cover it and and seemed ready to slide off any  moment. His many legxs, pitifully thin compared with the size of the the rest of him, waved about helplessly ashe looked looked.
 
-## Usage
-
-Need detailed usage patterns, dataset workflows, or tips for enabling the
-Rust accelerator? Consult the [Glitchlings Usage Guide](docs/index.md)
-for end-to-end instructions spanning the Python API, CLI, Hugging Face
+Consult the [Glitchlings Usage Guide](docs/index.md)
+for end-to-end instructions spanning the Python API, CLI, HuggingFace and Prime Intellect
 integrations, and the feature-flagged Rust pipeline.
 
-### Prime Intellect environments
-
-After `pip install -e .[prime]`, the `glitchlings.dlc.prime.load_environment` helper mirrors `verifiers.load_environment` for Prime Intellect scenarios while optionally applying glitchlings before returning the environment:
-
-```python
-from glitchlings import Mim1c, Typogre
-from glitchlings.dlc.prime import echo_chamber, load_environment
-
-env = load_environment(
-    "osoleve/syllabify-en",
-    glitchlings=[Mim1c(replacement_rate=0.01), Typogre(max_change_rate=0.02)],
-    seed=404,
-)
-
-# Spin up an echo chamber that corrupts a dataset column and
-# rewards models for perfectly restoring it
-practice_env = echo_chamber(
-    "osoleve/clean-room",
-    column="text",
-    glitchlings=["Typogre", "Mim1c"],
-    reward_function=lambda prompt, completion, answer: float(completion == answer),
-)
-```
-
-Skip the `glitchlings` argument to receive an untouched verifier dataset, and
-override `reward_function` when you want to evaluate completions with a custom
-scoring routine.
-
 ## Motivation
 
 If your model performs well on a particular task, but not when `Glitchling`s are present, it's a sign that it hasn't actually generalized to the problem.
@@ -344,8 +317,8 @@ Glitchlings are standard Python classes, so you can instantiate them with whatev
 ```python
 from glitchlings import Gaggle, Typogre, Mim1c
 
-custom_typogre = Typogre(max_change_rate=0.1)
-selective_mimic = Mim1c(replacement_rate=0.05, classes=["LATIN", "GREEK"])
+custom_typogre = Typogre(rate=0.1)
+selective_mimic = Mim1c(rate=0.05, classes=["LATIN", "GREEK"])
 
 gaggle = Gaggle([custom_typogre, selective_mimic], seed=99)
 print(gaggle("Summoned heroes do not fear the glitch."))
@@ -376,7 +349,7 @@ glitchlings --list
 glitchlings -g typogre --file documents/report.txt --diff
 
 # Configure glitchlings inline by passing keyword arguments.
-glitchlings -g "Typogre(max_change_rate=0.05)" "Ghouls just wanna have fun"
+glitchlings -g "Typogre(rate=0.05)" "Ghouls just wanna have fun"
 
 # Pipe text straight into the CLI for an on-the-fly corruption.
 echo "Beware LLM-written flavor-text" | glitchlings -g mim1c
@@ -400,7 +373,7 @@ _What a nice word, would be a shame if something happened to it._
 >
 > Args
 >
-> - `max_change_rate (float)`: The maximum number of edits to make as a percentage of the length (default: 0.02, 2%).
+> - `rate (float)`: The maximum number of edits to make as a percentage of the length (default: 0.02, 2%).
 > - `keyboard (str)`: Keyboard layout key-neighbor map to use (default: "CURATOR_QWERTY"; also accepts "QWERTY", "DVORAK", "COLEMAK", and "AZERTY").
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
@@ -412,7 +385,7 @@ _Wait, was that...?_
 >
 > Args
 >
-> - `replacement_rate (float)`: The maximum proportion of characters to replace (default: 0.02, 2%).
+> - `rate (float)`: The maximum proportion of characters to replace (default: 0.02, 2%).
 > - `classes (list[str] | "all")`: Restrict replacements to these Unicode script classes (default: ["LATIN", "GREEK", "CYRILLIC"]).
 > - `banned_characters (Collection[str])`: Characters that must never appear as replacements (default: none).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
@@ -425,7 +398,7 @@ _How can a computer need reading glasses?_
 >
 > Args
 >
-> - `error_rate (float)`: The maximum proportion of eligible confusion spans to replace (default: 0.02, 2%).
+> - `rate (float)`: The maximum proportion of eligible confusion spans to replace (default: 0.02, 2%).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
 ### Jargoyle
@@ -436,7 +409,7 @@ _Uh oh. The worst person you know just bought a thesaurus._
 >
 > Args
 >
-> - `replacement_rate (float)`: The maximum proportion of words to replace (default: 0.1, 10%).
+> - `rate (float)`: The maximum proportion of words to replace (default: 0.1, 10%).
 > - `part_of_speech`: The WordNet part(s) of speech to target (default: nouns). Accepts `wn.NOUN`, `wn.VERB`, `wn.ADJ`, `wn.ADV`, any iterable of those tags, or the string `"any"` to include them all.
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
@@ -448,7 +421,7 @@ _Did you say that or did I?_
 >
 > Args
 >
-> - `reduplication_rate (float)`: The maximum proportion of words to reduplicate (default: 0.05, 5%).
+> - `rate (float)`: The maximum proportion of words to reduplicate (default: 0.05, 5%).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
 ### Rushmore
@@ -459,7 +432,7 @@ _I accidentally an entire word._
 >
 > Args
 >
-> - `max_deletion_rate (float)`: The maximum proportion of words to delete (default: 0.01, 1%).
+> - `rate (float)`: The maximum proportion of words to delete (default: 0.01, 1%).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
 ### Redactyl
@@ -471,7 +444,7 @@ _Oops, that was my black highlighter._
 > ### Args
 >
 > - `replacement_char (str)`: The character to use for redaction (default: █).
-> - `redaction_rate (float)`: The maximum proportion of words to redact (default: 0.05, 5%).
+> - `rate (float)`: The maximum proportion of words to redact (default: 0.05, 5%).
 > - `merge_adjacent (bool)`: Whether to redact the space between adjacent redacted words (default: False).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 

{glitchlings-0.2.2 → glitchlings-0.2.3}/README.md

@@ -36,14 +36,16 @@ After all, what good is general intelligence if it can't handle a little chaos?
 pip install -U glitchlings
 ```
 
+> Glitchlings requires Python 3.10 or newer.
+
 ```python
 from glitchlings import Gaggle, SAMPLE_TEXT, Typogre, Mim1c, Reduple, Rushmore
 
 gaggle = Gaggle([
-    Typogre(max_change_rate=0.03),
-    Mim1c(replacement_rate=0.02),
+    Typogre(rate=0.03),
+    Mim1c(rate=0.02),
     Reduple(seed=404),
-    Rushmore(max_deletion_rate=0.02),
+    Rushmore(rate=0.02),
 ])
 
 print(gaggle(SAMPLE_TEXT))
@@ -51,41 +53,10 @@ print(gaggle(SAMPLE_TEXT))
 
 > Onҽ m‎ھ‎rning, wһen Gregor Samƽa woke from trouble𝐝 𝑑reams, he found himself transformed in his bed into a horrible vermin‎٠‎ He l   lay on his armour-like back, and if he lifted his head a little he could see his brown belly, slightlh domed and divided by arches ino stiff sections. The bedding was adly able to cover it and and seemed ready to slide off any  moment. His many legxs, pitifully thin compared with the size of the the rest of him, waved about helplessly ashe looked looked.
 
-## Usage
-
-Need detailed usage patterns, dataset workflows, or tips for enabling the
-Rust accelerator? Consult the [Glitchlings Usage Guide](docs/index.md)
-for end-to-end instructions spanning the Python API, CLI, Hugging Face
+Consult the [Glitchlings Usage Guide](docs/index.md)
+for end-to-end instructions spanning the Python API, CLI, HuggingFace and Prime Intellect
 integrations, and the feature-flagged Rust pipeline.
 
-### Prime Intellect environments
-
-After `pip install -e .[prime]`, the `glitchlings.dlc.prime.load_environment` helper mirrors `verifiers.load_environment` for Prime Intellect scenarios while optionally applying glitchlings before returning the environment:
-
-```python
-from glitchlings import Mim1c, Typogre
-from glitchlings.dlc.prime import echo_chamber, load_environment
-
-env = load_environment(
-    "osoleve/syllabify-en",
-    glitchlings=[Mim1c(replacement_rate=0.01), Typogre(max_change_rate=0.02)],
-    seed=404,
-)
-
-# Spin up an echo chamber that corrupts a dataset column and
-# rewards models for perfectly restoring it
-practice_env = echo_chamber(
-    "osoleve/clean-room",
-    column="text",
-    glitchlings=["Typogre", "Mim1c"],
-    reward_function=lambda prompt, completion, answer: float(completion == answer),
-)
-```
-
-Skip the `glitchlings` argument to receive an untouched verifier dataset, and
-override `reward_function` when you want to evaluate completions with a custom
-scoring routine.
-
 ## Motivation
 
 If your model performs well on a particular task, but not when `Glitchling`s are present, it's a sign that it hasn't actually generalized to the problem.
@@ -100,8 +71,8 @@ Glitchlings are standard Python classes, so you can instantiate them with whatev
 ```python
 from glitchlings import Gaggle, Typogre, Mim1c
 
-custom_typogre = Typogre(max_change_rate=0.1)
-selective_mimic = Mim1c(replacement_rate=0.05, classes=["LATIN", "GREEK"])
+custom_typogre = Typogre(rate=0.1)
+selective_mimic = Mim1c(rate=0.05, classes=["LATIN", "GREEK"])
 
 gaggle = Gaggle([custom_typogre, selective_mimic], seed=99)
 print(gaggle("Summoned heroes do not fear the glitch."))
@@ -132,7 +103,7 @@ glitchlings --list
 glitchlings -g typogre --file documents/report.txt --diff
 
 # Configure glitchlings inline by passing keyword arguments.
-glitchlings -g "Typogre(max_change_rate=0.05)" "Ghouls just wanna have fun"
+glitchlings -g "Typogre(rate=0.05)" "Ghouls just wanna have fun"
 
 # Pipe text straight into the CLI for an on-the-fly corruption.
 echo "Beware LLM-written flavor-text" | glitchlings -g mim1c
@@ -156,7 +127,7 @@ _What a nice word, would be a shame if something happened to it._
 >
 > Args
 >
-> - `max_change_rate (float)`: The maximum number of edits to make as a percentage of the length (default: 0.02, 2%).
+> - `rate (float)`: The maximum number of edits to make as a percentage of the length (default: 0.02, 2%).
 > - `keyboard (str)`: Keyboard layout key-neighbor map to use (default: "CURATOR_QWERTY"; also accepts "QWERTY", "DVORAK", "COLEMAK", and "AZERTY").
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
@@ -168,7 +139,7 @@ _Wait, was that...?_
 >
 > Args
 >
-> - `replacement_rate (float)`: The maximum proportion of characters to replace (default: 0.02, 2%).
+> - `rate (float)`: The maximum proportion of characters to replace (default: 0.02, 2%).
 > - `classes (list[str] | "all")`: Restrict replacements to these Unicode script classes (default: ["LATIN", "GREEK", "CYRILLIC"]).
 > - `banned_characters (Collection[str])`: Characters that must never appear as replacements (default: none).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
@@ -181,7 +152,7 @@ _How can a computer need reading glasses?_
 >
 > Args
 >
-> - `error_rate (float)`: The maximum proportion of eligible confusion spans to replace (default: 0.02, 2%).
+> - `rate (float)`: The maximum proportion of eligible confusion spans to replace (default: 0.02, 2%).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
 ### Jargoyle
@@ -192,7 +163,7 @@ _Uh oh. The worst person you know just bought a thesaurus._
 >
 > Args
 >
-> - `replacement_rate (float)`: The maximum proportion of words to replace (default: 0.1, 10%).
+> - `rate (float)`: The maximum proportion of words to replace (default: 0.1, 10%).
 > - `part_of_speech`: The WordNet part(s) of speech to target (default: nouns). Accepts `wn.NOUN`, `wn.VERB`, `wn.ADJ`, `wn.ADV`, any iterable of those tags, or the string `"any"` to include them all.
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
@@ -204,7 +175,7 @@ _Did you say that or did I?_
 >
 > Args
 >
-> - `reduplication_rate (float)`: The maximum proportion of words to reduplicate (default: 0.05, 5%).
+> - `rate (float)`: The maximum proportion of words to reduplicate (default: 0.05, 5%).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
 ### Rushmore
@@ -215,7 +186,7 @@ _I accidentally an entire word._
 >
 > Args
 >
-> - `max_deletion_rate (float)`: The maximum proportion of words to delete (default: 0.01, 1%).
+> - `rate (float)`: The maximum proportion of words to delete (default: 0.01, 1%).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 
 ### Redactyl
@@ -227,7 +198,7 @@ _Oops, that was my black highlighter._
 > ### Args
 >
 > - `replacement_char (str)`: The character to use for redaction (default: █).
-> - `redaction_rate (float)`: The maximum proportion of words to redact (default: 0.05, 5%).
+> - `rate (float)`: The maximum proportion of words to redact (default: 0.05, 5%).
 > - `merge_adjacent (bool)`: Whether to redact the space between adjacent redacted words (default: False).
 > - `seed (int)`: The random seed for reproducibility (default: 151).
 

{glitchlings-0.2.2 → glitchlings-0.2.3}/pyproject.toml

@@ -1,9 +1,9 @@
 [project]
 name = "glitchlings"
-version = "0.2.2"
+version = "0.2.3"
 description = "Monsters for your language games."
 readme = "README.md"
-requires-python = ">=3.12"
+requires-python = ">=3.10"
 
 dependencies = [
     "confusable-homoglyphs>=3.3.1",
@@ -22,6 +22,8 @@ classifiers = [
     "License :: OSI Approved :: Apache Software License",
     "Programming Language :: Python",
     "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Rust",
     "Operating System :: MacOS :: MacOS X",

{glitchlings-0.2.2 → glitchlings-0.2.3}/rust/zoo/build.rs

@@ -23,7 +23,13 @@ fn configured_python() -> Option<OsString> {
 }
 
 fn detect_python() -> Option<OsString> {
-    const CANDIDATES: &[&str] = &["python3.12", "python3", "python"];
+    const CANDIDATES: &[&str] = &[
+        "python3.12",
+        "python3.11",
+        "python3.10",
+        "python3",
+        "python",
+    ];
 
     for candidate in CANDIDATES {
         let status = Command::new(candidate)

{glitchlings-0.2.2 → glitchlings-0.2.3}/src/glitchlings/dlc/prime.py

@@ -79,8 +79,8 @@ def tutorial_level(
 ) -> vf.Environment:
     """Create a low-corruption environment using tuned defaults."""
 
-    tuned_mim1c = Mim1c(replacement_rate=0.01 * difficulty.value)
-    tuned_typogre = Typogre(max_change_rate=0.025 * difficulty.value)
+    tuned_mim1c = Mim1c(rate=0.01 * difficulty.value)
+    tuned_typogre = Typogre(rate=0.025 * difficulty.value)
 
     return load_environment(
         env,
@@ -220,32 +220,54 @@ def echo_chamber(
             "Specify which split to use when the dataset loads as a DatasetDict."
         )
 
-    prompts: list[list[dict[str, str]]] = []
-    answers: list[str] = []
+    filtered_dataset = hf_dataset.filter(
+        lambda row: row.get(column) is not None,
+        load_from_cache_file=False,
+    )
 
-    for row in hf_dataset:
-        value = row.get(column)
-        if value is None:
-            continue
+    source_column_names = list(filtered_dataset.column_names)
 
-        text = str(value)
-        prompts.append(
-            [
-                {"role": "system", "content": instructions},
-                {"role": "user", "content": f"Corrupted text:\n{text}"},
-            ]
-        )
-        answers.append(text)
+    def _build_prompt(row: dict[str, Any]) -> dict[str, Any]:
+        text = str(row[column])
+        prompt = [
+            {"role": "system", "content": instructions},
+            {"role": "user", "content": f"Corrupted text:\n{text}"},
+        ]
+        return {"prompt": prompt, "answer": text}
 
-    if not prompts:
-        raise ValueError(
-            f"Column '{column}' did not yield any textual entries in dataset '{dataset_id}'."
-        )
+    base_dataset = filtered_dataset.map(
+        _build_prompt,
+        remove_columns=source_column_names,
+        load_from_cache_file=False,
+    )
 
-    dataset = HFDataset.from_dict({"prompt": prompts, "answer": answers})
+    try:
+        dataset_length = len(base_dataset)  # type: ignore[arg-type]
+    except TypeError:
+        preview_rows: list[dict[str, Any]]
+        take_fn = getattr(base_dataset, "take", None)
+        if callable(take_fn):
+            preview_rows = list(take_fn(1))
+        else:
+            iterator = iter(base_dataset)
+            try:
+                first_row = next(iterator)
+            except StopIteration:
+                preview_rows = []
+            else:
+                preview_rows = [first_row]
+        if not preview_rows:
+            raise ValueError(
+                f"Column '{column}' did not yield any textual entries in dataset '{dataset_id}'."
+            )
+    else:
+        if dataset_length == 0:
+            raise ValueError(
+                f"Column '{column}' did not yield any textual entries in dataset '{dataset_id}'."
+            )
 
     gaggle = _as_gaggle(glitchlings, seed=seed)
-    glitched_dataset = gaggle.corrupt_dataset(dataset, ["prompt"])
+    glitched_dataset = gaggle.corrupt_dataset(base_dataset, ["prompt"])
 
     rubric_func = reward_function or symmetric_damerau_levenshtein_similarity
     rubric = vf.Rubric(funcs=[rubric_func], weights=[1.0])

{glitchlings-0.2.2 → glitchlings-0.2.3}/src/glitchlings/main.py

@@ -46,7 +46,7 @@ def build_parser() -> argparse.ArgumentParser:
         metavar="SPEC",
         help=(
             "Glitchling to apply, optionally with parameters like "
-            "Typogre(max_change_rate=0.05). Repeat for multiples; defaults to all built-ins."
+            "Typogre(rate=0.05). Repeat for multiples; defaults to all built-ins."
         ),
     )
     parser.add_argument(

glitchlings-0.2.3/src/glitchlings/zoo/_rate.py

@@ -0,0 +1,21 @@
+from __future__ import annotations
+
+
+def resolve_rate(
+    *,
+    rate: float | None,
+    legacy_value: float | None,
+    default: float,
+    legacy_name: str,
+) -> float:
+    """Return the effective rate while enforcing mutual exclusivity."""
+
+    if rate is not None and legacy_value is not None:
+        raise ValueError(
+            f"Specify either 'rate' or '{legacy_name}', not both."
+        )
+    if rate is not None:
+        return rate
+    if legacy_value is not None:
+        return legacy_value
+    return default

{glitchlings-0.2.2 → glitchlings-0.2.3}/src/glitchlings/zoo/core.py

@@ -107,6 +107,7 @@ class Glitchling:
         scope: AttackWave,
         order: AttackOrder = AttackOrder.NORMAL,
         seed: int | None = None,
+        pipeline_operation: Callable[["Glitchling"], dict[str, Any] | None] | None = None,
         **kwargs: Any,
     ) -> None:
         """Initialize a glitchling.
@@ -128,31 +129,76 @@ class Glitchling:
         self.corruption_function: CorruptionCallable = corruption_function
         self.level: AttackWave = scope
         self.order: AttackOrder = order
+        self._pipeline_descriptor_factory = pipeline_operation
         self.kwargs: dict[str, Any] = {}
+        self._cached_rng_callable: CorruptionCallable | None = None
+        self._cached_rng_expectation: bool | None = None
         for kw, val in kwargs.items():
             self.set_param(kw, val)
 
     def set_param(self, key: str, value: Any) -> None:
         """Persist a parameter for use by the corruption callable."""
 
-        setattr(self, key, value)
-        self.kwargs[key] = value
-        if key == "seed":
+        aliases = getattr(self, "_param_aliases", {})
+        canonical = aliases.get(key, key)
+
+        # Drop stale alias keys so we only forward canonical kwargs.
+        self.kwargs.pop(key, None)
+        for alias, target in aliases.items():
+            if target == canonical:
+                self.kwargs.pop(alias, None)
+
+        self.kwargs[canonical] = value
+        setattr(self, canonical, value)
+
+        if canonical == "seed":
             self.reset_rng(value)
 
-    def __corrupt(self, text: str, *args: Any, **kwargs: Any) -> str:
-        """Execute the corruption callable, injecting the RNG when required."""
+        for alias, target in aliases.items():
+            if target == canonical:
+                setattr(self, alias, value)
 
-        # Pass rng to underlying corruption function if it expects it.
+    def pipeline_operation(self) -> dict[str, Any] | None:
+        """Return the Rust pipeline operation descriptor for this glitchling."""
+
+        factory = self._pipeline_descriptor_factory
+        if factory is None:
+            return None
+
+        return factory(self)
+
+    def _corruption_expects_rng(self) -> bool:
+        """Return `True` when the corruption function accepts an rng keyword."""
+
+        cached_callable = self._cached_rng_callable
+        cached_expectation = self._cached_rng_expectation
+        corruption_function = self.corruption_function
+
+        if (
+            cached_callable is corruption_function
+            and cached_expectation is not None
+        ):
+            return cached_expectation
+
+        expects_rng = False
         try:
-            signature = inspect.signature(self.corruption_function)
+            signature = inspect.signature(corruption_function)
         except (TypeError, ValueError):
             signature = None
 
-        expects_rng = False
         if signature is not None:
             expects_rng = "rng" in signature.parameters
 
+        self._cached_rng_callable = corruption_function
+        self._cached_rng_expectation = expects_rng
+        return expects_rng
+
+    def __corrupt(self, text: str, *args: Any, **kwargs: Any) -> str:
+        """Execute the corruption callable, injecting the RNG when required."""
+
+        # Pass rng to underlying corruption function if it expects it.
+        expects_rng = self._corruption_expects_rng()
+
         if expects_rng:
             corrupted = self.corruption_function(text, *args, rng=self.rng, **kwargs)
         else:
@@ -231,53 +277,14 @@ class Glitchling:
                 self.corruption_function,
                 self.level,
                 self.order,
+                pipeline_operation=self._pipeline_descriptor_factory,
                 **filtered_kwargs,
             )
 
         return cls(**filtered_kwargs)
 
 
-def _pipeline_operation_reduplicate(glitchling: "Glitchling") -> dict[str, Any] | None:
-    rate = glitchling.kwargs.get("reduplication_rate")
-    if rate is None:
-        return None
-    return {"type": "reduplicate", "reduplication_rate": float(rate)}
-
 
-def _pipeline_operation_delete(glitchling: "Glitchling") -> dict[str, Any] | None:
-    rate = glitchling.kwargs.get("max_deletion_rate")
-    if rate is None:
-        return None
-    return {"type": "delete", "max_deletion_rate": float(rate)}
-
-
-def _pipeline_operation_redact(glitchling: "Glitchling") -> dict[str, Any] | None:
-    replacement_char = glitchling.kwargs.get("replacement_char")
-    redaction_rate = glitchling.kwargs.get("redaction_rate")
-    merge_adjacent = glitchling.kwargs.get("merge_adjacent")
-    if replacement_char is None or redaction_rate is None or merge_adjacent is None:
-        return None
-    return {
-        "type": "redact",
-        "replacement_char": str(replacement_char),
-        "redaction_rate": float(redaction_rate),
-        "merge_adjacent": bool(merge_adjacent),
-    }
-
-
-def _pipeline_operation_ocr(glitchling: "Glitchling") -> dict[str, Any] | None:
-    error_rate = glitchling.kwargs.get("error_rate")
-    if error_rate is None:
-        return None
-    return {"type": "ocr", "error_rate": float(error_rate)}
-
-
-_PIPELINE_OPERATION_BUILDERS: dict[str, Callable[["Glitchling"], dict[str, Any] | None]] = {
-    "Reduple": _pipeline_operation_reduplicate,
-    "Rushmore": _pipeline_operation_delete,
-    "Redactyl": _pipeline_operation_redact,
-    "Scannequin": _pipeline_operation_ocr,
-}
 
 
 class Gaggle(Glitchling):
@@ -359,10 +366,7 @@ class Gaggle(Glitchling):
 
         descriptors: list[dict[str, Any]] = []
         for glitchling in self.apply_order:
-            builder = _PIPELINE_OPERATION_BUILDERS.get(glitchling.name)
-            if builder is None:
-                return None
-            operation = builder(glitchling)
+            operation = glitchling.pipeline_operation()
             if operation is None:
                 return None