circuitcheck 0.1.0__tar.gz

circuitcheck-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 albertlieyingadrian
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

circuitcheck-0.1.0/PKG-INFO

@@ -0,0 +1,171 @@
+Metadata-Version: 2.4
+Name: circuitcheck
+Version: 0.1.0
+Summary: Tools on top of circuit-tracer: diff attribution graphs across training, validate circuits with controls, and label features cheaply.
+Keywords: interpretability,circuits,attribution-graphs,transcoders,circuit-tracer,autointerp
+Author: albertlieyingadrian
+License-Expression: MIT
+License-File: LICENSE
+Classifier: Development Status :: 2 - Pre-Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Typing :: Typed
+Requires-Dist: numpy>=1.26
+Requires-Dist: pydantic>=2.7,<3
+Requires-Dist: typer>=0.12
+Requires-Dist: anthropic>=0.69 ; extra == 'anthropic'
+Requires-Dist: pytest>=9.0 ; extra == 'dev'
+Requires-Dist: pytest-cov>=7.0 ; extra == 'dev'
+Requires-Dist: ruff>=0.17 ; extra == 'dev'
+Requires-Dist: mypy>=2.0 ; extra == 'dev'
+Requires-Dist: nbclient>=0.10 ; extra == 'dev'
+Requires-Dist: nbformat>=5.10 ; extra == 'dev'
+Requires-Dist: ipykernel>=6.29 ; extra == 'dev'
+Requires-Dist: openai>=1.100 ; extra == 'openai'
+Requires-Dist: torch>=2.4 ; extra == 'tracer'
+Requires-Dist: transformers>=4.51 ; extra == 'tracer'
+Requires-Python: >=3.11
+Project-URL: Homepage, https://github.com/albertlieyingadrian/circuitcheck
+Project-URL: Source, https://github.com/albertlieyingadrian/circuitcheck
+Project-URL: Documentation, https://github.com/albertlieyingadrian/circuitcheck/tree/main/docs
+Project-URL: Changelog, https://github.com/albertlieyingadrian/circuitcheck/blob/main/CHANGELOG.md
+Provides-Extra: anthropic
+Provides-Extra: dev
+Provides-Extra: openai
+Provides-Extra: tracer
+Description-Content-Type: text/markdown
+
+# circuitcheck
+
+<p align="center"><img src="docs/images/demo.gif" alt="circuitcheck validate --demo: the toy model's planted switch comes back as a switch, its two readouts as readouts and its redundant pair as backups" width="760"></p>
+
+**Test what an attribution graph claims.** circuitcheck is three tools on top of
+[circuit-tracer](https://github.com/decoderesearch/circuit-tracer) attribution graphs:
+
+| Tool | Question it answers | Runs without a GPU on |
+|---|---|---|
+| `circuitcheck validate` | Is a group of graph features a **switch** (it causes the behaviour) or a **readout** (it only reflects it)? | the built-in toy model |
+| `circuitcheck diff` | Which features and edges did fine-tuning or RL add or remove, and is the change real? | saved graph summaries |
+| `circuitcheck label` | What does each feature do, in a few words, and how far can that label be trusted? | cached model replies |
+
+▶ [Watch the 2.5-minute walkthrough](docs/videos/walkthrough.mp4) · [Use cases](docs/use-cases.mdx) ·
+[circuitcheck vs circuit-tracer](docs/vs-circuit-tracer.mdx) · [Docs](docs/index.mdx)
+
+It grew out of [steerlab's circuits episodes](https://albertlieyingadrian.github.io/steerlab/experiments/#circuits).
+In [a password lock](https://albertlieyingadrian.github.io/steerlab/experiments/lock-circuit/report/) and
+[an RL reward hack](https://albertlieyingadrian.github.io/steerlab/experiments/rl-hack-circuit/report/), the graph's
+most influential features all turned out to be readouts. circuitcheck turns the checks behind that finding into
+reusable code.
+
+## Install
+
+```bash
+pip install circuitcheck        # after the first PyPI release; until then:
+pip install "git+https://github.com/albertlieyingadrian/circuitcheck"
+# real models: circuit-tracer from GitHub, plus torch
+pip install "git+https://github.com/decoderesearch/circuit-tracer"
+# labels: pip install "circuitcheck[anthropic]" (Claude Haiku 5.5, the default) or "[openai]"
+```
+
+## Five-minute tour (no GPU, no key)
+
+```bash
+circuitcheck validate --demo --html validate.html
+circuitcheck diff examples/data/rl_hack_graphs.json.gz --before installed/before --after installed/after \
+  --control-before control/before --control-after control/after --html diff.html
+```
+
+```python
+from circuitcheck.validate import logit_diff, validate
+from circuitcheck.validate.toy import demo, demo_nodes
+
+backend = demo()
+metric = logit_diff(backend.token_id(" Austin"), backend.token_id(" Sacramento"))
+report = validate(backend, "dallas", demo_nodes(backend), metric=metric, counterfactual="oakland")
+print([(g.name, g.verdict.value) for g in report.groups])
+# [('top-5', 'switch'), ('L5/100@5', 'switch'), ('L7/200@5', 'readout'), ('L8/201@5', 'readout'), ...]
+```
+
+On a real model, `circuitcheck.tracer.TracerBackend` wraps a circuit-tracer `ReplacementModel`:
+
+```python
+from circuit_tracer import ReplacementModel
+from circuitcheck.tracer import TracerBackend
+
+model = ReplacementModel.from_pretrained("google/gemma-2-2b", "gemma")
+report = validate(TracerBackend(model), prompt, nodes, metric=metric, counterfactual=other_prompt)
+```
+
+## What each tool checks
+
+**diff.** Graphs of the same prompts before and after training are reduced to per-feature influence and
+feature-to-feature edges, summed over positions. Each feature gets a status (gained, lost, up, down, stable). The
+statuses are backed by:
+
+- a paired swap null: before and after labels are swapped at random within each prompt pair, with a
+  Benjamini-Hochberg correction across features;
+- a bootstrap interval and a split-half check;
+- optionally, a control run (did the change happen only in the run that learned the behaviour?) and a replication
+  seed.
+
+The validity check compares replacement and completeness scores, error-node influence and per-layer reconstruction
+error (`circuitcheck.tracer.reconstruction_error`) before and after. The transcoders were trained on the base model,
+and the report says so every time.
+
+**validate.** For each group of features:
+
+- ablation (set them to 0);
+- steering (multiply them by 3);
+- a counterfactual patch (copy their values from a prompt with a different answer);
+- each of the above on 20 random feature sets matched by layer and position, and on 20 sets matched by write size
+  (|activation| x decoder norm).
+
+The verdict follows printed rules: switch, partial, backup, nonspecific or readout. An optional explanation swap (Gao's
+test) runs label rules written as code in a checked sandbox. Output is JSON plus an HTML page.
+
+**labels.** The pipeline:
+
+1. Explain each feature from its top activating records.
+2. Score the label on held-out records: detection AUROC (activating and silent excerpts) and, optionally, simulation
+   correlation.
+3. Report calibration: stated confidence against the earned score, and the detector's expected calibration error.
+
+Every reply is cached. Labels export in Neuronpedia's explanation format, and Neuronpedia's own labels can be scored by
+the same detector for a fair comparison.
+
+## Results that reproduce the episodes
+
+- **Graph diff.** On the RL-hack study's 48 graphs, `diff_graphs` picks the same ten features the study did, in the
+  same order. 9 of the 10 are specific to the run that hacked and 6 come back in another seed (test:
+  `tests/test_diff.py::test_reproduces_rl_hack_selection`). Split halves of the prompts agree on 70% of the top 10, and
+  reconstruction error is unchanged by RL (0.500 vs 0.499).
+- **Labels.** Claude Haiku 5.5 labelled 50 Gemma-2-2B transcoder features for $0.26 in total. Its labels match
+  Neuronpedia's on detection (AUROC 0.80 vs 0.78 and 0.80) and beat them on simulation (r = 0.45 vs 0.37 and 0.38).
+  The explainer's stated confidence barely predicts its score (r = 0.11), so use the score, not the confidence.
+- **Live validator runs** on the RL-hack model and the password-locked model: see `examples/data/live/`.
+
+## Layout
+
+```
+src/circuitcheck/
+  diff/       graphs.py (summaries and loaders: circuit-tracer .pt, Neuronpedia JSON), core.py (the diff)
+  validate/   core.py (checks and verdicts), controls.py, metrics.py, sandbox.py, toy.py
+  labels/     autointerp.py (explain, detect, simulate, calibrate), providers.py (Claude, OpenAI, cache), neuronpedia.py
+  tracer.py   circuit-tracer adapter and reconstruction error
+  report.py   HTML pages
+  cli.py      the circuitcheck command
+examples/     notebooks 01-03, data/, scripts/ (how the bundled data was made), modal/ (live GPU checks)
+```
+
+## Development
+
+```bash
+uv venv && uv pip install -e ".[dev]"
+pytest && ruff check && mypy src
+```
+
+MIT licence.

circuitcheck-0.1.0/README.md

@@ -0,0 +1,130 @@
+# circuitcheck
+
+<p align="center"><img src="docs/images/demo.gif" alt="circuitcheck validate --demo: the toy model's planted switch comes back as a switch, its two readouts as readouts and its redundant pair as backups" width="760"></p>
+
+**Test what an attribution graph claims.** circuitcheck is three tools on top of
+[circuit-tracer](https://github.com/decoderesearch/circuit-tracer) attribution graphs:
+
+| Tool | Question it answers | Runs without a GPU on |
+|---|---|---|
+| `circuitcheck validate` | Is a group of graph features a **switch** (it causes the behaviour) or a **readout** (it only reflects it)? | the built-in toy model |
+| `circuitcheck diff` | Which features and edges did fine-tuning or RL add or remove, and is the change real? | saved graph summaries |
+| `circuitcheck label` | What does each feature do, in a few words, and how far can that label be trusted? | cached model replies |
+
+▶ [Watch the 2.5-minute walkthrough](docs/videos/walkthrough.mp4) · [Use cases](docs/use-cases.mdx) ·
+[circuitcheck vs circuit-tracer](docs/vs-circuit-tracer.mdx) · [Docs](docs/index.mdx)
+
+It grew out of [steerlab's circuits episodes](https://albertlieyingadrian.github.io/steerlab/experiments/#circuits).
+In [a password lock](https://albertlieyingadrian.github.io/steerlab/experiments/lock-circuit/report/) and
+[an RL reward hack](https://albertlieyingadrian.github.io/steerlab/experiments/rl-hack-circuit/report/), the graph's
+most influential features all turned out to be readouts. circuitcheck turns the checks behind that finding into
+reusable code.
+
+## Install
+
+```bash
+pip install circuitcheck        # after the first PyPI release; until then:
+pip install "git+https://github.com/albertlieyingadrian/circuitcheck"
+# real models: circuit-tracer from GitHub, plus torch
+pip install "git+https://github.com/decoderesearch/circuit-tracer"
+# labels: pip install "circuitcheck[anthropic]" (Claude Haiku 5.5, the default) or "[openai]"
+```
+
+## Five-minute tour (no GPU, no key)
+
+```bash
+circuitcheck validate --demo --html validate.html
+circuitcheck diff examples/data/rl_hack_graphs.json.gz --before installed/before --after installed/after \
+  --control-before control/before --control-after control/after --html diff.html
+```
+
+```python
+from circuitcheck.validate import logit_diff, validate
+from circuitcheck.validate.toy import demo, demo_nodes
+
+backend = demo()
+metric = logit_diff(backend.token_id(" Austin"), backend.token_id(" Sacramento"))
+report = validate(backend, "dallas", demo_nodes(backend), metric=metric, counterfactual="oakland")
+print([(g.name, g.verdict.value) for g in report.groups])
+# [('top-5', 'switch'), ('L5/100@5', 'switch'), ('L7/200@5', 'readout'), ('L8/201@5', 'readout'), ...]
+```
+
+On a real model, `circuitcheck.tracer.TracerBackend` wraps a circuit-tracer `ReplacementModel`:
+
+```python
+from circuit_tracer import ReplacementModel
+from circuitcheck.tracer import TracerBackend
+
+model = ReplacementModel.from_pretrained("google/gemma-2-2b", "gemma")
+report = validate(TracerBackend(model), prompt, nodes, metric=metric, counterfactual=other_prompt)
+```
+
+## What each tool checks
+
+**diff.** Graphs of the same prompts before and after training are reduced to per-feature influence and
+feature-to-feature edges, summed over positions. Each feature gets a status (gained, lost, up, down, stable). The
+statuses are backed by:
+
+- a paired swap null: before and after labels are swapped at random within each prompt pair, with a
+  Benjamini-Hochberg correction across features;
+- a bootstrap interval and a split-half check;
+- optionally, a control run (did the change happen only in the run that learned the behaviour?) and a replication
+  seed.
+
+The validity check compares replacement and completeness scores, error-node influence and per-layer reconstruction
+error (`circuitcheck.tracer.reconstruction_error`) before and after. The transcoders were trained on the base model,
+and the report says so every time.
+
+**validate.** For each group of features:
+
+- ablation (set them to 0);
+- steering (multiply them by 3);
+- a counterfactual patch (copy their values from a prompt with a different answer);
+- each of the above on 20 random feature sets matched by layer and position, and on 20 sets matched by write size
+  (|activation| x decoder norm).
+
+The verdict follows printed rules: switch, partial, backup, nonspecific or readout. An optional explanation swap (Gao's
+test) runs label rules written as code in a checked sandbox. Output is JSON plus an HTML page.
+
+**labels.** The pipeline:
+
+1. Explain each feature from its top activating records.
+2. Score the label on held-out records: detection AUROC (activating and silent excerpts) and, optionally, simulation
+   correlation.
+3. Report calibration: stated confidence against the earned score, and the detector's expected calibration error.
+
+Every reply is cached. Labels export in Neuronpedia's explanation format, and Neuronpedia's own labels can be scored by
+the same detector for a fair comparison.
+
+## Results that reproduce the episodes
+
+- **Graph diff.** On the RL-hack study's 48 graphs, `diff_graphs` picks the same ten features the study did, in the
+  same order. 9 of the 10 are specific to the run that hacked and 6 come back in another seed (test:
+  `tests/test_diff.py::test_reproduces_rl_hack_selection`). Split halves of the prompts agree on 70% of the top 10, and
+  reconstruction error is unchanged by RL (0.500 vs 0.499).
+- **Labels.** Claude Haiku 5.5 labelled 50 Gemma-2-2B transcoder features for $0.26 in total. Its labels match
+  Neuronpedia's on detection (AUROC 0.80 vs 0.78 and 0.80) and beat them on simulation (r = 0.45 vs 0.37 and 0.38).
+  The explainer's stated confidence barely predicts its score (r = 0.11), so use the score, not the confidence.
+- **Live validator runs** on the RL-hack model and the password-locked model: see `examples/data/live/`.
+
+## Layout
+
+```
+src/circuitcheck/
+  diff/       graphs.py (summaries and loaders: circuit-tracer .pt, Neuronpedia JSON), core.py (the diff)
+  validate/   core.py (checks and verdicts), controls.py, metrics.py, sandbox.py, toy.py
+  labels/     autointerp.py (explain, detect, simulate, calibrate), providers.py (Claude, OpenAI, cache), neuronpedia.py
+  tracer.py   circuit-tracer adapter and reconstruction error
+  report.py   HTML pages
+  cli.py      the circuitcheck command
+examples/     notebooks 01-03, data/, scripts/ (how the bundled data was made), modal/ (live GPU checks)
+```
+
+## Development
+
+```bash
+uv venv && uv pip install -e ".[dev]"
+pytest && ruff check && mypy src
+```
+
+MIT licence.

circuitcheck-0.1.0/pyproject.toml

@@ -0,0 +1,134 @@
+[build-system]
+requires = ["uv_build>=0.11.7,<0.14"]
+build-backend = "uv_build"
+
+[project]
+name = "circuitcheck"
+version = "0.1.0"
+description = "Tools on top of circuit-tracer: diff attribution graphs across training, validate circuits with controls, and label features cheaply."
+readme = "README.md"
+requires-python = ">=3.11"
+license = "MIT"
+license-files = ["LICENSE"]
+keywords = [
+    "interpretability",
+    "circuits",
+    "attribution-graphs",
+    "transcoders",
+    "circuit-tracer",
+    "autointerp",
+]
+classifiers = [
+    "Development Status :: 2 - Pre-Alpha",
+    "Intended Audience :: Science/Research",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Typing :: Typed",
+]
+dependencies = [
+    "numpy>=1.26",
+    "pydantic>=2.7,<3",
+    "typer>=0.12",
+]
+
+[[project.authors]]
+name = "albertlieyingadrian"
+
+[project.optional-dependencies]
+tracer = [
+    "torch>=2.4",
+    "transformers>=4.51",
+]
+anthropic = ["anthropic>=0.69"]
+openai = ["openai>=1.100"]
+dev = [
+    "pytest>=9.0",
+    "pytest-cov>=7.0",
+    "ruff>=0.17",
+    "mypy>=2.0",
+    "nbclient>=0.10",
+    "nbformat>=5.10",
+    "ipykernel>=6.29",
+]
+
+[project.scripts]
+circuitcheck = "circuitcheck.cli:app"
+
+[project.urls]
+Homepage = "https://github.com/albertlieyingadrian/circuitcheck"
+Source = "https://github.com/albertlieyingadrian/circuitcheck"
+Documentation = "https://github.com/albertlieyingadrian/circuitcheck/tree/main/docs"
+Changelog = "https://github.com/albertlieyingadrian/circuitcheck/blob/main/CHANGELOG.md"
+
+[dependency-groups]
+dev = [
+    "pytest>=9.0",
+    "pytest-cov>=7.0",
+    "ruff>=0.17",
+    "mypy>=2.0",
+]
+
+[tool.uv.build-backend]
+module-name = "circuitcheck"
+source-include = ["tests/**"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-ra --strict-markers --strict-config -m 'not slow'"
+markers = ["slow: real-model or network runs, excluded by default (run with -m slow)"]
+
+[tool.ruff]
+line-length = 120
+target-version = "py311"
+extend-exclude = [
+    ".agents",
+    ".claude",
+]
+
+[tool.ruff.lint]
+select = [
+    "E",
+    "F",
+    "W",
+    "I",
+    "B",
+    "UP",
+    "SIM",
+    "RUF",
+]
+ignore = [
+    "RUF001",
+    "RUF002",
+    "RUF003",
+]
+
+[tool.mypy]
+python_version = "3.11"
+strict = true
+plugins = ["pydantic.mypy"]
+
+[[tool.mypy.overrides]]
+module = [
+    "torch",
+    "torch.*",
+    "transformers",
+    "transformers.*",
+    "circuit_tracer",
+    "circuit_tracer.*",
+    "anthropic",
+    "anthropic.*",
+    "openai",
+    "openai.*",
+]
+ignore_missing_imports = true
+
+[tool.coverage.run]
+source = ["circuitcheck"]
+branch = true
+
+[tool.coverage.report]
+show_missing = true
+skip_covered = true

circuitcheck-0.1.0/pyproject.toml.orig

@@ -0,0 +1,95 @@
+[build-system]
+# Keep the upper bound: uv_build follows uv's versioning and may change between minors.
+requires = ["uv_build>=0.11.7,<0.14"]
+build-backend = "uv_build"
+
+[project]
+name = "circuitcheck"
+# The ONLY place the version lives. Code reads it with importlib.metadata.
+version = "0.1.0"
+description = "Tools on top of circuit-tracer: diff attribution graphs across training, validate circuits with controls, and label features cheaply."
+readme = "README.md"
+requires-python = ">=3.11"
+license = "MIT"
+license-files = ["LICENSE"]
+authors = [{ name = "albertlieyingadrian" }]
+keywords = ["interpretability", "circuits", "attribution-graphs", "transcoders", "circuit-tracer", "autointerp"]
+classifiers = [
+    "Development Status :: 2 - Pre-Alpha",
+    "Intended Audience :: Science/Research",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Typing :: Typed",
+]
+# The base install stays light: no torch. Everything runs on saved graphs, the toy model or cached labels.
+dependencies = [
+    "numpy>=1.26",
+    "pydantic>=2.7,<3",
+    "typer>=0.12",
+]
+
+[project.optional-dependencies]
+# circuit-tracer is installed from GitHub (it is not on PyPI): see the README.
+tracer = ["torch>=2.4", "transformers>=4.51"]
+anthropic = ["anthropic>=0.69"]
+openai = ["openai>=1.100"]
+dev = ["pytest>=9.0", "pytest-cov>=7.0", "ruff>=0.17", "mypy>=2.0", "nbclient>=0.10", "nbformat>=5.10", "ipykernel>=6.29"]
+
+[project.scripts]
+circuitcheck = "circuitcheck.cli:app"
+
+[project.urls]
+Homepage = "https://github.com/albertlieyingadrian/circuitcheck"
+Source = "https://github.com/albertlieyingadrian/circuitcheck"
+Documentation = "https://github.com/albertlieyingadrian/circuitcheck/tree/main/docs"
+Changelog = "https://github.com/albertlieyingadrian/circuitcheck/blob/main/CHANGELOG.md"
+
+[dependency-groups]
+dev = [
+    "pytest>=9.0",
+    "pytest-cov>=7.0",
+    "ruff>=0.17",
+    "mypy>=2.0",
+]
+
+[tool.uv.build-backend]
+module-name = "circuitcheck"
+# The wheel ships only src/circuitcheck; the sdist also carries the tests.
+source-include = ["tests/**"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-ra --strict-markers --strict-config -m 'not slow'"
+markers = [
+    "slow: real-model or network runs, excluded by default (run with -m slow)",
+]
+
+[tool.ruff]
+line-length = 120
+target-version = "py311"
+# Agent skills are third-party tooling, not this project's code.
+extend-exclude = [".agents", ".claude"]
+
+[tool.ruff.lint]
+select = ["E", "F", "W", "I", "B", "UP", "SIM", "RUF"]
+ignore = ["RUF001", "RUF002", "RUF003"]
+
+[tool.mypy]
+python_version = "3.11"
+strict = true
+plugins = ["pydantic.mypy"]
+
+[[tool.mypy.overrides]]
+module = ["torch", "torch.*", "transformers", "transformers.*", "circuit_tracer", "circuit_tracer.*", "anthropic", "anthropic.*", "openai", "openai.*"]
+ignore_missing_imports = true
+
+[tool.coverage.run]
+source = ["circuitcheck"]
+branch = true
+
+[tool.coverage.report]
+show_missing = true
+skip_covered = true

circuitcheck-0.1.0/src/circuitcheck/__init__.py

@@ -0,0 +1,16 @@
+"""circuitcheck: tools on top of circuit-tracer.
+
+- `circuitcheck.diff`: compare attribution graphs before and after fine-tuning or RL;
+- `circuitcheck.validate`: test whether graph features are a switch or a readout, against random and matched-norm
+  controls;
+- `circuitcheck.labels`: cheap automatic feature labels with detection and simulation scores;
+- `circuitcheck.tracer`: the circuit-tracer adapter (needs torch and circuit-tracer);
+- `circuitcheck.report`: HTML pages for all three.
+"""
+
+from importlib.metadata import PackageNotFoundError, version
+
+try:
+    __version__ = version("circuitcheck")
+except PackageNotFoundError:  # pragma: no cover
+    __version__ = "0.0.0+local"