inspect-drex 0.1.0__tar.gz

inspect_drex-0.1.0/.github/workflows/ci.yaml

@@ -0,0 +1,61 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ci-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  lint:
+    name: Lint
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
+          cache-dependency-glob: "uv.lock"
+
+      - name: Install dependencies
+        run: uv sync --all-groups --locked --python 3.12
+
+      - name: Ruff check
+        run: uv run ruff check .
+
+      - name: Ruff format
+        run: uv run ruff format --check .
+
+  test:
+    name: Offline tests (Python ${{ matrix.python-version }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
+          cache-dependency-glob: "uv.lock"
+
+      - name: Install dependencies
+        run: uv sync --all-groups --locked --python ${{ matrix.python-version }}
+
+      # The live tests need keys and make billed requests, so only the offline
+      # suite runs here: `tests/unit_tests` leaves them out of collection.
+      - name: Offline tests
+        run: uv run pytest tests/unit_tests --cov

inspect_drex-0.1.0/.github/workflows/release.yaml

@@ -0,0 +1,96 @@
+# Publishes inspect-drex to PyPI when an `inspect-drex-v*` tag is pushed.
+#
+# Releasing is deliberate: `git tag inspect-drex-v0.1.0 && git push origin inspect-drex-v0.1.0`.
+# A PyPI upload cannot be undone or replaced, so everything that can be checked runs
+# before the upload, and `workflow_dispatch` runs the same checks plus the build while
+# skipping the publish, so the pipeline can be rehearsed.
+name: Release
+
+on:
+  push:
+    tags: ["inspect-drex-v*"]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: release-${{ github.ref }}
+  # Never cancel: a half-finished release is worse than a queued one.
+  cancel-in-progress: false
+
+jobs:
+  verify:
+    name: Verify and build
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+        with:
+          enable-cache: true
+          cache-dependency-glob: "uv.lock"
+
+      - name: Verify lockfile and install dependencies
+        run: uv sync --all-groups --locked --python 3.12
+
+      # The tag names the version users will get, so a tag that disagrees with the
+      # package is caught here rather than after an unrepeatable upload.
+      - name: Tag matches package version
+        if: github.event_name == 'push'
+        run: |
+          tag_version="${GITHUB_REF_NAME#inspect-drex-v}"
+          pkg_version="$(uv run python -c "from inspect_drex import __version__; print(__version__)")"
+          if [ "$tag_version" != "$pkg_version" ]; then
+            echo "Tag ${GITHUB_REF_NAME} would publish $pkg_version, not $tag_version." >&2
+            echo "Bump inspect_drex/_version.py or retag before releasing." >&2
+            exit 1
+          fi
+          echo "Releasing inspect-drex $pkg_version"
+
+      - name: Ruff
+        run: uv run ruff check .
+
+      - name: Offline tests
+        run: uv run pytest tests/unit_tests --cov
+
+      - name: Build distributions
+        run: uv build --out-dir dist
+
+      # The README becomes the PyPI project page; twine rejects markup PyPI cannot
+      # render, which is otherwise only discovered at upload time.
+      - name: Check distribution metadata
+        run: uvx twine check dist/*
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: inspect-drex-dist
+          path: dist/
+          if-no-files-found: error
+
+  publish:
+    name: Publish to PyPI
+    needs: verify
+    # Only a tag publishes; workflow_dispatch stops after verify.
+    if: github.event_name == 'push'
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    # Half of the Trusted Publishing identity, and where release reviewers get
+    # added later without touching this file.
+    environment: pypi
+    permissions:
+      # Trusted Publishing mints a short-lived token from this; no PyPI
+      # credential is stored in the repository.
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: inspect-drex-dist
+          path: dist
+
+      # Pinned to a commit: anything running in this job can mint the PyPI token
+      # for inspect-drex. Update it deliberately when a new release is reviewed.
+      - name: Publish
+        uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33 # v1.14.2

inspect_drex-0.1.0/.gitignore

@@ -0,0 +1,11 @@
+.venv/
+dist/
+build/
+*.egg-info/
+.coverage
+.pytest_cache/
+.ruff_cache/
+__pycache__/
+.DS_Store
+.env
+logs/

inspect_drex-0.1.0/CHANGELOG.md

@@ -0,0 +1,15 @@
+# Changelog
+
+## 0.1.0
+
+First release.
+
+- `drex_scorer`: an Inspect scorer that asks Drex yes/no (`noul`), `choice` and `score`
+  questions about a sample's output and returns the pass probability as `Score.value`, with
+  Drex's full probabilities kept in `Score.metadata["drex"]`.
+- Registered with Inspect through the `inspect_ai` entry point, so
+  `inspect score --scorer inspect_drex/drex_scorer` works.
+- Image input for Drex Max, refused before sending for text-only models.
+- Drex through OpenRouter (`nace-ai/<model>` ids) with the key following the host.
+- Self-hosted Drex deployments through `NACE_BASE_URL` or `base_url=`.
+- Errors raise by default; `on_error="unscored"` leaves the sample out of the metrics.

inspect_drex-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Nace AI
+
+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.

inspect_drex-0.1.0/PKG-INFO

@@ -0,0 +1,206 @@
+Metadata-Version: 2.5
+Name: inspect-drex
+Version: 0.1.0
+Summary: An Inspect (UK AISI) scorer backed by Drex decision models from Nace
+Project-URL: Homepage, https://console.nace.ai/docs
+Project-URL: Repository, https://github.com/nace-ai/inspect-drex
+Project-URL: Issues, https://github.com/nace-ai/inspect-drex/issues
+Project-URL: Changelog, https://github.com/nace-ai/inspect-drex/blob/main/CHANGELOG.md
+Author-email: Nace AI <hello@nace.ai>
+License-Expression: MIT
+License-File: LICENSE
+Keywords: drex,evals,inspect,inspect-ai,llm-evaluation,nace,scorer
+Classifier: Development Status :: 4 - Beta
+Classifier: Framework :: Pytest
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: <4.0,>=3.11
+Requires-Dist: httpx<1.0.0,>=0.27.0
+Requires-Dist: inspect-ai>=0.3.245
+Requires-Dist: nace-sdk<1.0.0,>=0.4.0
+Description-Content-Type: text/markdown
+
+# inspect-drex
+
+A scorer for [Inspect](https://inspect.aisi.org.uk) (the UK AI Security Institute's evaluation
+framework) that asks [Drex](https://console.nace.ai/docs) decision models from Nace whether a
+sample's output meets your criteria. Drex answers typed questions and returns
+**probabilities**, so you can choose the pass threshold after the run instead of before it.
+
+- Drex Max reads text, JSON and images. `drex-v2.0` reads text and JSON only.
+- Any other model id that Drex serves is passed through unchanged; there is no closed list.
+- The default model is `drex-max`, which moves to the newest flagship with each Drex release.
+  Pin `model=` (or set `NACE_DEFAULT_DECISION_MODEL`) if you need results that never change.
+  See the Drex docs for each model's request limits.
+
+## Install
+
+```bash
+pip install inspect-drex
+export NACE_API_KEY=...   # an API key from the Drex console
+```
+
+Python 3.11 or newer (the floor of `inspect-ai`).
+
+`NACE_BASE_URL` is optional. Set it, or pass `base_url=`, for a self-hosted Drex deployment on
+your own domain. Any host that is not OpenRouter is treated as Drex itself: requests go to
+`<base_url>/v1/systemone` and use your `NACE_API_KEY`.
+
+## Use
+
+```python
+from inspect_ai import Task, task
+from inspect_ai.dataset import Sample
+from inspect_ai.solver import generate
+from inspect_drex import Criterion, drex_scorer
+
+
+@task
+def support():
+    return Task(
+        dataset=[Sample(input="How long do refunds take?", target="5 business days")],
+        solver=generate(),
+        scorer=drex_scorer(
+            [
+                Criterion.noul(
+                    "grounded", "Does the answer state the reference fact?", threshold=0.7
+                ),
+                Criterion.noul("filler", "Is the answer padded with filler?", want=False),
+            ],
+            model="drex-max",
+        ),
+    )
+```
+
+Save it as `support.py` and run it with any model Inspect supports for the solver:
+
+```bash
+inspect eval support.py --model openai/gpt-4o-mini
+```
+
+A rubric string is a single yes/no criterion: `drex_scorer("The answer is helpful.")`.
+The package registers itself with Inspect through an entry point, so a finished log can also be
+scored from the CLI:
+
+```bash
+inspect score logs/run.eval --scorer inspect_drex/drex_scorer -S criteria="The answer is helpful."
+```
+
+`examples/support_eval.py` is an offline example (a mock model and a fake Drex server).
+
+### What you get back
+
+- `Score.value` is the 0-1 **pass probability** (weighted mean of the criteria, or `min` with
+  `aggregate="min"`). `mean()` and `stderr()` report it. With `output="grade"` it is `CORRECT`
+  or `INCORRECT` by each criterion's own `threshold`, and `mean()` is the pass rate.
+- `Score.metadata["drex"]` keeps what Drex returned: `request_id`, `model`, `input_tokens`,
+  `image_tokens`, `evaluation_time_ms`, and per criterion its kind, pass score, threshold,
+  verdict and full probabilities (noul: `yes` / `no`; choice and score: per label).
+  Input tokens are counted once per request, not once per question.
+- For an unscored sample, `Score.metadata["drex"]` holds only `error` (the exception class
+  name); `Score.explanation` has the message, including the status and `request_id`.
+- `Score.explanation` lists the criteria that failed, with their probabilities.
+
+### Criteria and score direction
+
+`Criterion.noul` asks a yes/no question, `Criterion.choice` picks among labelled options and
+`Criterion.score` places the output on an ordered scale. A criterion's pass score always means
+"how well it holds": higher is better. A *defect* question ("is the reply rude?") needs
+`want=False`, which scores `1 - P(yes)`. Choice criteria sum the probability of the labels you
+mark `good`. Score criteria use the expected level over `levels - 1` (or its complement with
+`invert=True`); an expected level is a fraction that can fall between two levels, and the raw
+level is kept in the detail text. See Drex's guides on
+[questions](https://console.nace.ai/docs/drex/guides/questions) and
+[reading answers](https://console.nace.ai/docs/drex/guides/reading-answers).
+
+### What Drex sees
+
+By default `{"input", "output", "reference"}`: the user messages, the model output and the
+sample target. Give Drex only what the criteria need, or pass your own
+`state=lambda state, target: ...` (and `images=` for Drex Max).
+
+Images in the user's messages are sent in order, and the input text names each one
+`[Image 1]`, `[Image 2]`, so a criterion can say "Image 1". Images are never dropped: with a
+text-only model (`drex-v2.0`, `drex-v1.5`, `drex-latest`) the scorer raises before sending
+(with `on_error="unscored"` the sample is left unscored instead). Drex never fetches URLs, so
+remote image URLs are refused; use data URLs or local files. Keep each request within Drex's
+size limit (inline images count, base64 included). For larger images, upload them with
+`nace-sdk` and pass the uploaded file through `images=`.
+
+### Errors are not failed criteria
+
+A 401 (bad key), 402 (no funds; the message links the billing page), 422 (each issue's
+`path: message` is shown) or an outage raises, and Inspect marks the sample as errored. With
+`on_error="unscored"` the sample stays in the log as `Score.unscored(reason="scoring_failed")`
+and is left out of the metrics. 429 and 529 are retried by `nace-sdk`, honouring
+`retry-after-ms`. The client timeout is at least 60 s, and the client caps concurrent requests
+(`max_in_flight=`, default 8).
+
+### OpenRouter
+
+Drex is also served through OpenRouter. Pass `base_url="https://openrouter.ai/api/v1"` and use
+model ids such as `nace-ai/<model>`; a bare id (`drex-max`) is sent as `nace-ai/drex-max`, and ids
+with a provider prefix are sent as given. `drex-latest` is refused because OpenRouter has no id
+for it.
+
+```bash
+export OPENROUTER_API_KEY=...
+```
+
+```python
+drex_scorer(
+    "The answer is helpful.",
+    base_url="https://openrouter.ai/api/v1",
+    model="nace-ai/drex-max",
+)
+```
+
+The API key follows the host: for OpenRouter it is read from `OPENROUTER_API_KEY`, never from
+`NACE_API_KEY`, so a Nace key is not sent to OpenRouter and an OpenRouter key is not sent to
+Nace. Through OpenRouter, images work only inline (in `state` content parts); uploaded and
+workspace-file images are refused.
+
+### Secrets and the eval log
+
+Inspect writes scorer arguments to the eval log in plain text, so `drex_scorer` refuses
+`api_key=` and `extra_headers=`. Put the key in the environment, or build the client yourself
+and pass `client=AsyncDrexClient(base_url=..., api_key=..., model=...)`.
+
+### Attribution
+
+Every request sends `User-Agent` and `X-Nace-Client` as `inspect-drex/<version>`, so Nace can
+count usage per integration.
+
+## What is verified
+
+- The offline tests (`tests/unit_tests`) run against a fake Drex server and are the CI gate
+  on Python 3.11, 3.12 and 3.13.
+- The live tests passed against production Drex and against OpenRouter on 2026-10-29. They are
+  not run in CI, because they need keys and make billed requests.
+
+```bash
+uv sync --all-groups
+uv run ruff check . && uv run ruff format --check .
+uv run pytest tests/unit_tests --cov                          # offline
+NACE_API_KEY=... uv run pytest tests/integration_tests -m live -s -rs
+OPENROUTER_API_KEY=... uv run pytest tests/integration_tests -m openrouter -s -rs
+```
+
+## Limits
+
+- Images need a model that reads them (Drex Max); text-only models raise before sending.
+- Each request must fit Drex's size and question limits; one request is made per sample, however
+  many criteria it holds.
+- The scorer judges what you give it. It does not fetch URLs or files for the judge.
+
+## License
+
+MIT. See [LICENSE](LICENSE). Issues: <https://github.com/nace-ai/inspect-drex/issues>.

inspect_drex-0.1.0/README.md

@@ -0,0 +1,176 @@
+# inspect-drex
+
+A scorer for [Inspect](https://inspect.aisi.org.uk) (the UK AI Security Institute's evaluation
+framework) that asks [Drex](https://console.nace.ai/docs) decision models from Nace whether a
+sample's output meets your criteria. Drex answers typed questions and returns
+**probabilities**, so you can choose the pass threshold after the run instead of before it.
+
+- Drex Max reads text, JSON and images. `drex-v2.0` reads text and JSON only.
+- Any other model id that Drex serves is passed through unchanged; there is no closed list.
+- The default model is `drex-max`, which moves to the newest flagship with each Drex release.
+  Pin `model=` (or set `NACE_DEFAULT_DECISION_MODEL`) if you need results that never change.
+  See the Drex docs for each model's request limits.
+
+## Install
+
+```bash
+pip install inspect-drex
+export NACE_API_KEY=...   # an API key from the Drex console
+```
+
+Python 3.11 or newer (the floor of `inspect-ai`).
+
+`NACE_BASE_URL` is optional. Set it, or pass `base_url=`, for a self-hosted Drex deployment on
+your own domain. Any host that is not OpenRouter is treated as Drex itself: requests go to
+`<base_url>/v1/systemone` and use your `NACE_API_KEY`.
+
+## Use
+
+```python
+from inspect_ai import Task, task
+from inspect_ai.dataset import Sample
+from inspect_ai.solver import generate
+from inspect_drex import Criterion, drex_scorer
+
+
+@task
+def support():
+    return Task(
+        dataset=[Sample(input="How long do refunds take?", target="5 business days")],
+        solver=generate(),
+        scorer=drex_scorer(
+            [
+                Criterion.noul(
+                    "grounded", "Does the answer state the reference fact?", threshold=0.7
+                ),
+                Criterion.noul("filler", "Is the answer padded with filler?", want=False),
+            ],
+            model="drex-max",
+        ),
+    )
+```
+
+Save it as `support.py` and run it with any model Inspect supports for the solver:
+
+```bash
+inspect eval support.py --model openai/gpt-4o-mini
+```
+
+A rubric string is a single yes/no criterion: `drex_scorer("The answer is helpful.")`.
+The package registers itself with Inspect through an entry point, so a finished log can also be
+scored from the CLI:
+
+```bash
+inspect score logs/run.eval --scorer inspect_drex/drex_scorer -S criteria="The answer is helpful."
+```
+
+`examples/support_eval.py` is an offline example (a mock model and a fake Drex server).
+
+### What you get back
+
+- `Score.value` is the 0-1 **pass probability** (weighted mean of the criteria, or `min` with
+  `aggregate="min"`). `mean()` and `stderr()` report it. With `output="grade"` it is `CORRECT`
+  or `INCORRECT` by each criterion's own `threshold`, and `mean()` is the pass rate.
+- `Score.metadata["drex"]` keeps what Drex returned: `request_id`, `model`, `input_tokens`,
+  `image_tokens`, `evaluation_time_ms`, and per criterion its kind, pass score, threshold,
+  verdict and full probabilities (noul: `yes` / `no`; choice and score: per label).
+  Input tokens are counted once per request, not once per question.
+- For an unscored sample, `Score.metadata["drex"]` holds only `error` (the exception class
+  name); `Score.explanation` has the message, including the status and `request_id`.
+- `Score.explanation` lists the criteria that failed, with their probabilities.
+
+### Criteria and score direction
+
+`Criterion.noul` asks a yes/no question, `Criterion.choice` picks among labelled options and
+`Criterion.score` places the output on an ordered scale. A criterion's pass score always means
+"how well it holds": higher is better. A *defect* question ("is the reply rude?") needs
+`want=False`, which scores `1 - P(yes)`. Choice criteria sum the probability of the labels you
+mark `good`. Score criteria use the expected level over `levels - 1` (or its complement with
+`invert=True`); an expected level is a fraction that can fall between two levels, and the raw
+level is kept in the detail text. See Drex's guides on
+[questions](https://console.nace.ai/docs/drex/guides/questions) and
+[reading answers](https://console.nace.ai/docs/drex/guides/reading-answers).
+
+### What Drex sees
+
+By default `{"input", "output", "reference"}`: the user messages, the model output and the
+sample target. Give Drex only what the criteria need, or pass your own
+`state=lambda state, target: ...` (and `images=` for Drex Max).
+
+Images in the user's messages are sent in order, and the input text names each one
+`[Image 1]`, `[Image 2]`, so a criterion can say "Image 1". Images are never dropped: with a
+text-only model (`drex-v2.0`, `drex-v1.5`, `drex-latest`) the scorer raises before sending
+(with `on_error="unscored"` the sample is left unscored instead). Drex never fetches URLs, so
+remote image URLs are refused; use data URLs or local files. Keep each request within Drex's
+size limit (inline images count, base64 included). For larger images, upload them with
+`nace-sdk` and pass the uploaded file through `images=`.
+
+### Errors are not failed criteria
+
+A 401 (bad key), 402 (no funds; the message links the billing page), 422 (each issue's
+`path: message` is shown) or an outage raises, and Inspect marks the sample as errored. With
+`on_error="unscored"` the sample stays in the log as `Score.unscored(reason="scoring_failed")`
+and is left out of the metrics. 429 and 529 are retried by `nace-sdk`, honouring
+`retry-after-ms`. The client timeout is at least 60 s, and the client caps concurrent requests
+(`max_in_flight=`, default 8).
+
+### OpenRouter
+
+Drex is also served through OpenRouter. Pass `base_url="https://openrouter.ai/api/v1"` and use
+model ids such as `nace-ai/<model>`; a bare id (`drex-max`) is sent as `nace-ai/drex-max`, and ids
+with a provider prefix are sent as given. `drex-latest` is refused because OpenRouter has no id
+for it.
+
+```bash
+export OPENROUTER_API_KEY=...
+```
+
+```python
+drex_scorer(
+    "The answer is helpful.",
+    base_url="https://openrouter.ai/api/v1",
+    model="nace-ai/drex-max",
+)
+```
+
+The API key follows the host: for OpenRouter it is read from `OPENROUTER_API_KEY`, never from
+`NACE_API_KEY`, so a Nace key is not sent to OpenRouter and an OpenRouter key is not sent to
+Nace. Through OpenRouter, images work only inline (in `state` content parts); uploaded and
+workspace-file images are refused.
+
+### Secrets and the eval log
+
+Inspect writes scorer arguments to the eval log in plain text, so `drex_scorer` refuses
+`api_key=` and `extra_headers=`. Put the key in the environment, or build the client yourself
+and pass `client=AsyncDrexClient(base_url=..., api_key=..., model=...)`.
+
+### Attribution
+
+Every request sends `User-Agent` and `X-Nace-Client` as `inspect-drex/<version>`, so Nace can
+count usage per integration.
+
+## What is verified
+
+- The offline tests (`tests/unit_tests`) run against a fake Drex server and are the CI gate
+  on Python 3.11, 3.12 and 3.13.
+- The live tests passed against production Drex and against OpenRouter on 2026-10-29. They are
+  not run in CI, because they need keys and make billed requests.
+
+```bash
+uv sync --all-groups
+uv run ruff check . && uv run ruff format --check .
+uv run pytest tests/unit_tests --cov                          # offline
+NACE_API_KEY=... uv run pytest tests/integration_tests -m live -s -rs
+OPENROUTER_API_KEY=... uv run pytest tests/integration_tests -m openrouter -s -rs
+```
+
+## Limits
+
+- Images need a model that reads them (Drex Max); text-only models raise before sending.
+- Each request must fit Drex's size and question limits; one request is made per sample, however
+  many criteria it holds.
+- The scorer judges what you give it. It does not fetch URLs or files for the judge.
+
+## License
+
+MIT. See [LICENSE](LICENSE). Issues: <https://github.com/nace-ai/inspect-drex/issues>.