scholarmend 0.1.0__tar.gz

scholarmend-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,29 @@
+name: ci
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
+        with:
+          python-version: ${{ matrix.python-version }}
+      - run: pip install -e ".[dev]"
+      # The suite is offline by design: an autouse fixture fails any real urlopen,
+      # and the acceptance tests skip without the sibling Trust-Evals-LitReview corpus.
+      - run: pytest -q
+      - run: ruff check src tests
+      - run: mypy src

scholarmend-0.1.0/.github/workflows/release.yml

@@ -0,0 +1,39 @@
+name: release
+
+# Publishes to PyPI via Trusted Publishing (OIDC). There is deliberately no API
+# token stored in this repository: PyPI verifies the workflow's identity
+# directly, so there is no long-lived credential to leak or rotate.
+on:
+  release:
+    types: [published]
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
+        with:
+          python-version: "3.12"
+      - run: pip install build
+      - run: python -m build
+      - uses: actions/upload-artifact@v7
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    needs: build
+    runs-on: ubuntu-latest
+    environment: pypi          # add required reviewers here to gate releases
+    permissions:
+      id-token: write          # required for Trusted Publishing
+    steps:
+      - uses: actions/download-artifact@v8
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1

scholarmend-0.1.0/.gitignore

@@ -0,0 +1,15 @@
+__pycache__/
+*.py[cod]
+.venv/
+venv/
+dist/
+build/
+*.egg-info/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.env
+.superpowers/
+out/
+out-*/
+*.log

scholarmend-0.1.0/CHANGELOG.md

@@ -0,0 +1,25 @@
+# Changelog
+
+## 0.1.0
+
+First release on PyPI.
+
+- Reads Google Scholar RIS exports and recovers venue, year, and track, mostly
+  from the proceedings URL Scholar already put in each record. OpenReview,
+  PMLR, PMC, and Semantic Scholar fill in what the URL alone can't.
+- OpenReview forums that were never moved to API v2, such as some 2022-2023
+  workshops, are read from API v1 instead.
+- A withdrawn or non-public OpenReview forum is cached as hidden, so a rerun
+  doesn't ask again. Delete its cache entry to check whether it has since
+  been made public.
+- Writes `resolved.json` (every claim behind each field), `mended.ris` (the
+  corrected RIS, ready for Covidence), and `report.txt` (what stayed
+  unresolved, and why).
+- Never drops a record and never guesses. A venue it can't determine keeps
+  Scholar's value and is marked unresolved for a human to check.
+- `--offline` replays a committed response cache and never reaches the
+  network. A record whose lookup is missing from the cache falls back to
+  URL-only resolution, and the run exits 1 so you know it was partial.
+- `--abstracts` replaces Scholar's snippet with the full abstract, but only
+  when the source shows the record's own title.
+- No runtime dependencies. Python 3.10 to 3.13.

scholarmend-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 SHARE Lab, University of Waterloo
+
+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.

scholarmend-0.1.0/PKG-INFO

@@ -0,0 +1,186 @@
+Metadata-Version: 2.5
+Name: scholarmend
+Version: 0.1.0
+Summary: Recover true venue, year and track for Google Scholar exports, so screening reasons over facts rather than truncations.
+Project-URL: Homepage, https://github.com/uw-share-lab/scholarmend
+Project-URL: Issues, https://github.com/uw-share-lab/scholarmend/issues
+License: MIT
+License-File: LICENSE
+Keywords: bibliometrics,google-scholar,openreview,prisma,ris,systematic-review
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+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 :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Provides-Extra: dev
+Requires-Dist: mypy>=1.5; extra == 'dev'
+Requires-Dist: pytest>=7; extra == 'dev'
+Requires-Dist: ruff>=0.4; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# scholarmend
+
+Recover true bibliographic metadata for Google Scholar exports, so that
+screening tools reason over facts rather than over Scholar's truncations.
+
+## Usage
+
+    pip install scholarmend
+    scholarmend --input path/to/scholar-exports --out out
+
+`--input` takes a directory of RIS files exported from Google Scholar, or a
+single `.ris`. For development, clone the repo and `pip install -e ".[dev]"`.
+
+Tier 1 needs no configuration and resolves venue, year and track for 77% of a
+Scholar corpus. Tier 2 needs an OpenReview account:
+
+    export SCHOLARMEND_OPENREVIEW_USER='you@example.edu'
+    export SCHOLARMEND_OPENREVIEW_PASSWORD='...'
+
+Every response is written to `--cache` (default `.scholarmend-cache`). Commit
+it: a rerun then reproduces byte-identically and makes no API calls.
+
+    scholarmend --input path/to/scholar-exports --out out --offline
+
+`--offline` never reaches the network. A record whose lookup is missing from
+the cache falls back to tier 1, and the run exits 1 to say it was partial.
+
+### Outputs
+
+| File | Contents |
+|------|----------|
+| `resolved.json` | canonical records: winning value per field, plus every claim behind it |
+| `mended.ris` | the RIS projection, for Covidence and venuetriage |
+| `report.txt` | what stayed unresolved, and why |
+
+## The problem
+
+Google Scholar's `.ris` export is the standard input to a systematic review, and
+it is lossy in ways that stay invisible until they corrupt a result. Measured on
+a 2,413-record corpus of NeurIPS, ICLR and ICML search results:
+
+| Defect | Rate |
+|--------|------|
+| Venue string ellipsized (`… Neural Information …`) | 98% |
+| Author list truncated to five plus `...` | 71% |
+| Records carrying a DOI | 2 of 2,413 |
+| Scholar's `PY` disagrees with the year in the record's own URL | 68% |
+
+One record shows the whole problem, and its own answer:
+
+    JF  - … Neural Information …
+    AU  - ...
+    PY  - 2026///
+    UR  - https://proceedings.neurips.cc/paper_files/paper/2025/hash/4da4f3c0…-Abstract-Datasets_and_Benchmarks_Track.html
+
+Scholar reports an ellipsis for the venue, the wrong year, and no track. The URL
+in the same record states NeurIPS, 2025, Datasets and Benchmarks Track.
+
+## The approach
+
+The identifier in this corpus is not a DOI — there are two — it is the **URL
+path**. Offline URL mining yields a key for **99%** of records and fully resolves
+venue, year and track for **77%**, with no network access at all. Records are
+then escalated per field through cached, credentialed lookups only where mining
+leaves a gap.
+
+Every claim is retained with its source, tier, confidence and evidence, so the
+losing values stay auditable rather than being overwritten silently.
+
+**Abstracts, on request.** Scholar's `AB` is a search snippet on every record —
+fragments joined by `…` around the query terms — and that is what a screening
+tool shows reviewers. `--abstracts` replaces it with the paper's abstract, from
+the proceedings page, the OpenReview submission note, or Semantic Scholar, in
+that order, admitting each only when the source names the record's own title.
+Run it on the deduplicated, triaged upload rather than the whole corpus:
+
+    scholarmend --input ../Trust-Evals-LitReview/out/clean.ris --out out-covidence --abstracts
+
+`report.txt` lists every record whose abstract is still Scholar's snippet.
+
+**What it does not recover: authors.** Scholar truncates 71% of author lists,
+and scholarmend leaves them as it found them: rewriting `AU` would restructure a
+repeated field, which the projection never does. See `BACKLOG.md` §5.
+
+Querying a structured database instead does not work here, and the spec records
+the measurement: on a 30-title sample, OpenAlex matched 63% of titles and
+returned a correct conference venue for **none** of them, describing the arXiv
+preprint instead. The preprint carries a DOI and the proceedings version does
+not, so a DOI-anchored index indexes the preprint.
+
+## Validation
+
+scholarmend is tested against hand-verified ground truth rather than against
+its own output. The Trust-Evals-LitReview review produced labels for 112
+records that escaped an automated rule table and were resolved individually,
+with the evidence for each recorded.
+
+| Check | Bar |
+|-------|-----|
+| Records with an automated resolution path, of those 112 | at least 103 (measured: 104) |
+| …of which, venue determined **and verified against the reviewers' labels** | **103**, zero disagreements |
+| …the remaining 1 | a sibling proceedings URL, reached but not separately asserted |
+| Workshop status against reviewer labels (the 90 OpenReview records) | zero per-record disagreements |
+| PMLR volumes against the reviewers' `venue_true` | 10, all out of scope, all correctly declined |
+| PMC bridge against the reviewers' override decisions | 3, v267 named ICML 2025, v287 and v297 declined |
+| Verdict flips in `decisions.csv` without a recorded reason | 112 flips, **0** unreasoned |
+| Per-record overrides | 5 of 10 reproduced; the other 5 asserted unreachable |
+| Scholar's year losing every disagreement | all 1,264 |
+| Proceedings mining coverage | exactly 1,854 of 2,413 |
+| Records with no miner at all | exactly 17 |
+| Hand-maintained merge list | retired; 4 collapse at tier 1, 6 at tier 2 |
+
+### What "settled" does and does not mean
+
+The 104 figure counts records for which the miners extract something a live
+resolver consumes — a venue directly, or an OpenReview forum id, a PMLR volume
+or a PMC id. That is a claim about *reach*, not about correctness, and the two
+halves of it are verified to different depths:
+
+- **90 records** reached through an OpenReview `venueid`: venue, year and a
+  workshop verdict, every one agreeing with the reviewers' hand-verified label.
+  Zero disagreements, compared record by record.
+- **10 PMLR volumes**, all out of scope for this review — the Canadian
+  Conference on AI, a workshop at ACML, an AAAI bridge programme, a parsimony
+  conference. The resolver retrieves each proceedings title and **deliberately
+  declines to name a venue**, rather than coerce an unfamiliar conference onto a
+  known one. Verified against the volume the reviewers recorded in `venue_true`.
+- **3 records via the PMC bridge**, reproducing the reviewers' override
+  decisions exactly: PMC volume 267 → PMLR v267 → ICML 2025, named; v287 (CHIL)
+  and v297 (ML4H) retrieved and declined.
+- **1 record** reached by a sibling proceedings URL, not separately asserted.
+- **8 records** have no automated route at all: NSF landing pages, Google Books
+  chapters, an SPIE paper, a PDF on a personal page. These needed human
+  judgement before and still do — and the suite asserts they resolve to nothing,
+  so the day a miner starts covering one, a test says so.
+
+So the honest summary is that scholarmend **determines and verifies** the venue
+for **103** of the 112, reaches 1 more without a separate assertion, and leaves
+8 untouched.
+
+Run them with the review repository checked out alongside this one:
+
+    pytest tests/test_acceptance.py -v
+
+They skip cleanly when it is not.
+
+## Relationship to other tools
+
+- [`venuetriage`](../Trust-Evals-LitReview) — consumes scholarmend's output to
+  separate workshop from main-track papers before Covidence.
+- [`refaudit`](https://github.com/uw-share-lab/refaudit) — verifies a finished
+  bibliography against Crossref, OpenAlex and arXiv. Different job, different
+  input; its DOI-centric resolvers reach only ~8% coverage on this corpus.
+
+## Licence
+
+MIT.

scholarmend-0.1.0/README.md

@@ -0,0 +1,157 @@
+# scholarmend
+
+Recover true bibliographic metadata for Google Scholar exports, so that
+screening tools reason over facts rather than over Scholar's truncations.
+
+## Usage
+
+    pip install scholarmend
+    scholarmend --input path/to/scholar-exports --out out
+
+`--input` takes a directory of RIS files exported from Google Scholar, or a
+single `.ris`. For development, clone the repo and `pip install -e ".[dev]"`.
+
+Tier 1 needs no configuration and resolves venue, year and track for 77% of a
+Scholar corpus. Tier 2 needs an OpenReview account:
+
+    export SCHOLARMEND_OPENREVIEW_USER='you@example.edu'
+    export SCHOLARMEND_OPENREVIEW_PASSWORD='...'
+
+Every response is written to `--cache` (default `.scholarmend-cache`). Commit
+it: a rerun then reproduces byte-identically and makes no API calls.
+
+    scholarmend --input path/to/scholar-exports --out out --offline
+
+`--offline` never reaches the network. A record whose lookup is missing from
+the cache falls back to tier 1, and the run exits 1 to say it was partial.
+
+### Outputs
+
+| File | Contents |
+|------|----------|
+| `resolved.json` | canonical records: winning value per field, plus every claim behind it |
+| `mended.ris` | the RIS projection, for Covidence and venuetriage |
+| `report.txt` | what stayed unresolved, and why |
+
+## The problem
+
+Google Scholar's `.ris` export is the standard input to a systematic review, and
+it is lossy in ways that stay invisible until they corrupt a result. Measured on
+a 2,413-record corpus of NeurIPS, ICLR and ICML search results:
+
+| Defect | Rate |
+|--------|------|
+| Venue string ellipsized (`… Neural Information …`) | 98% |
+| Author list truncated to five plus `...` | 71% |
+| Records carrying a DOI | 2 of 2,413 |
+| Scholar's `PY` disagrees with the year in the record's own URL | 68% |
+
+One record shows the whole problem, and its own answer:
+
+    JF  - … Neural Information …
+    AU  - ...
+    PY  - 2026///
+    UR  - https://proceedings.neurips.cc/paper_files/paper/2025/hash/4da4f3c0…-Abstract-Datasets_and_Benchmarks_Track.html
+
+Scholar reports an ellipsis for the venue, the wrong year, and no track. The URL
+in the same record states NeurIPS, 2025, Datasets and Benchmarks Track.
+
+## The approach
+
+The identifier in this corpus is not a DOI — there are two — it is the **URL
+path**. Offline URL mining yields a key for **99%** of records and fully resolves
+venue, year and track for **77%**, with no network access at all. Records are
+then escalated per field through cached, credentialed lookups only where mining
+leaves a gap.
+
+Every claim is retained with its source, tier, confidence and evidence, so the
+losing values stay auditable rather than being overwritten silently.
+
+**Abstracts, on request.** Scholar's `AB` is a search snippet on every record —
+fragments joined by `…` around the query terms — and that is what a screening
+tool shows reviewers. `--abstracts` replaces it with the paper's abstract, from
+the proceedings page, the OpenReview submission note, or Semantic Scholar, in
+that order, admitting each only when the source names the record's own title.
+Run it on the deduplicated, triaged upload rather than the whole corpus:
+
+    scholarmend --input ../Trust-Evals-LitReview/out/clean.ris --out out-covidence --abstracts
+
+`report.txt` lists every record whose abstract is still Scholar's snippet.
+
+**What it does not recover: authors.** Scholar truncates 71% of author lists,
+and scholarmend leaves them as it found them: rewriting `AU` would restructure a
+repeated field, which the projection never does. See `BACKLOG.md` §5.
+
+Querying a structured database instead does not work here, and the spec records
+the measurement: on a 30-title sample, OpenAlex matched 63% of titles and
+returned a correct conference venue for **none** of them, describing the arXiv
+preprint instead. The preprint carries a DOI and the proceedings version does
+not, so a DOI-anchored index indexes the preprint.
+
+## Validation
+
+scholarmend is tested against hand-verified ground truth rather than against
+its own output. The Trust-Evals-LitReview review produced labels for 112
+records that escaped an automated rule table and were resolved individually,
+with the evidence for each recorded.
+
+| Check | Bar |
+|-------|-----|
+| Records with an automated resolution path, of those 112 | at least 103 (measured: 104) |
+| …of which, venue determined **and verified against the reviewers' labels** | **103**, zero disagreements |
+| …the remaining 1 | a sibling proceedings URL, reached but not separately asserted |
+| Workshop status against reviewer labels (the 90 OpenReview records) | zero per-record disagreements |
+| PMLR volumes against the reviewers' `venue_true` | 10, all out of scope, all correctly declined |
+| PMC bridge against the reviewers' override decisions | 3, v267 named ICML 2025, v287 and v297 declined |
+| Verdict flips in `decisions.csv` without a recorded reason | 112 flips, **0** unreasoned |
+| Per-record overrides | 5 of 10 reproduced; the other 5 asserted unreachable |
+| Scholar's year losing every disagreement | all 1,264 |
+| Proceedings mining coverage | exactly 1,854 of 2,413 |
+| Records with no miner at all | exactly 17 |
+| Hand-maintained merge list | retired; 4 collapse at tier 1, 6 at tier 2 |
+
+### What "settled" does and does not mean
+
+The 104 figure counts records for which the miners extract something a live
+resolver consumes — a venue directly, or an OpenReview forum id, a PMLR volume
+or a PMC id. That is a claim about *reach*, not about correctness, and the two
+halves of it are verified to different depths:
+
+- **90 records** reached through an OpenReview `venueid`: venue, year and a
+  workshop verdict, every one agreeing with the reviewers' hand-verified label.
+  Zero disagreements, compared record by record.
+- **10 PMLR volumes**, all out of scope for this review — the Canadian
+  Conference on AI, a workshop at ACML, an AAAI bridge programme, a parsimony
+  conference. The resolver retrieves each proceedings title and **deliberately
+  declines to name a venue**, rather than coerce an unfamiliar conference onto a
+  known one. Verified against the volume the reviewers recorded in `venue_true`.
+- **3 records via the PMC bridge**, reproducing the reviewers' override
+  decisions exactly: PMC volume 267 → PMLR v267 → ICML 2025, named; v287 (CHIL)
+  and v297 (ML4H) retrieved and declined.
+- **1 record** reached by a sibling proceedings URL, not separately asserted.
+- **8 records** have no automated route at all: NSF landing pages, Google Books
+  chapters, an SPIE paper, a PDF on a personal page. These needed human
+  judgement before and still do — and the suite asserts they resolve to nothing,
+  so the day a miner starts covering one, a test says so.
+
+So the honest summary is that scholarmend **determines and verifies** the venue
+for **103** of the 112, reaches 1 more without a separate assertion, and leaves
+8 untouched.
+
+Run them with the review repository checked out alongside this one:
+
+    pytest tests/test_acceptance.py -v
+
+They skip cleanly when it is not.
+
+## Relationship to other tools
+
+- [`venuetriage`](../Trust-Evals-LitReview) — consumes scholarmend's output to
+  separate workshop from main-track papers before Covidence.
+- [`refaudit`](https://github.com/uw-share-lab/refaudit) — verifies a finished
+  bibliography against Crossref, OpenAlex and arXiv. Different job, different
+  input; its DOI-centric resolvers reach only ~8% coverage on this corpus.
+
+## Licence
+
+MIT.

scholarmend-0.1.0/docs/openreview-auth.md

@@ -0,0 +1,82 @@
+# OpenReview authentication spike
+
+**Question:** does an OpenReview login token clear the CAPTCHA/challenge that
+blocks anonymous reads of `api2.openreview.net/notes`?
+
+**Answer: yes.**
+
+## Steps and results
+
+1. **Anonymous baseline** — `GET /notes?forum=0GgFeojE4a&limit=1` with no
+   auth header still returns `403`. The anonymous challenge is unchanged.
+
+2. **Obtaining a token** — `POST /login` with `{"id": <email>, "password":
+   <password>}` (`Content-Type: application/json`) returns a JSON body with
+   two top-level keys: `token` and `user`. `token` is a JWT.
+
+3. **Using the token** — a request to the endpoint that matters, with header
+
+   ```
+   Authorization: Bearer <token>
+   ```
+
+   returns `200` (not `403`) for forum id `0GgFeojE4a`, which is the same
+   forum blocked in step 1. This confirms the token clears the challenge.
+
+## JSON path to venueid — confirmed, no correction needed
+
+The brief's assumed path is correct:
+
+```
+notes[0].content.venueid.value
+```
+
+For forum `0GgFeojE4a` this returned `ICML.cc/2026/Workshop/AI4GOOD`, which
+matches the gold value in
+`../Trust-Evals-LitReview/verification/openreview-venues.json` for that
+forum id.
+
+Other fields present alongside `venueid` in `content` (for reference, not
+needed by Task 7): `title`, `authors`, `authorids`, `keywords`, `TLDR`,
+`abstract`, `pdf`, `email_sharing`, `data_release`, `venue`, `_bibtex`,
+`paperhash`.
+
+## Rate-limit behaviour (10 calls, 1s delay between each)
+
+All 10 calls against distinct forum ids returned `200`. Response headers
+exposed both a legacy and a standard rate-limit header set (values agreed
+in every response):
+
+- `ratelimit-policy: 500;w=3600` — **500 requests per 3600-second (1 hour)
+  rolling window**, per authenticated token/user.
+- `ratelimit-remaining` decremented by exactly 1 per call (497 → 488 across
+  the 10 calls), consistent with the stated policy.
+- No throttling, no `429`, and no slowdown in response latency (each call
+  completed in ~0.09–0.13s) across the 10-call sample.
+
+**Implication for tier 2:** the 527 calls tier 2 needs exceed the 500/hour
+budget by 27 calls. A single unbroken run will hit the limit near the end.
+Tier 2 should either (a) pace calls so the run spans slightly over an hour
+(e.g. resume after the window resets, using `ratelimit-reset`, which the
+response gives as seconds until the window rolls over), or (b) split the
+527 calls into two batches separated by a wait for the window to reset.
+Either is a small addition to the existing 1-second-delay loop in the
+brief's Step 4 script.
+
+## Token lifetime
+
+The JWT's decoded claims (`user`, `iat`, `exp`, `iss` — decoded locally to
+inspect timing only, never printed in full or committed) show:
+
+- **`exp - iat` = 86400 seconds = 24 hours.**
+
+A token obtained once is valid for the full 527-call run and does not need
+mid-run refresh, as long as the run completes within 24 hours of login.
+
+## Conclusion
+
+Authentication via `POST /login` + `Authorization: Bearer <token>` clears
+the challenge that blocks anonymous reads. Task 7 does **not** need the
+one-time authenticated-browser-capture fallback: it can call `/login` once
+per run (or once per 24h), reuse the token for all reads, and must budget
+calls to the 500-requests/hour limit (pace or split runs of >500 calls).

scholarmend-0.1.0/pyproject.toml

@@ -0,0 +1,72 @@
+[build-system]
+requires = ["hatchling>=1.20"]
+build-backend = "hatchling.build"
+
+[tool.hatch.version]
+path = "src/scholarmend/_version.py"
+
+[project]
+name = "scholarmend"
+dynamic = ["version"]
+description = "Recover true venue, year and track for Google Scholar exports, so screening reasons over facts rather than truncations."
+readme = "README.md"
+requires-python = ">=3.10"
+license = { text = "MIT" }
+keywords = ["systematic-review", "google-scholar", "ris", "openreview", "prisma", "bibliometrics"]
+classifiers = [
+  "Development Status :: 3 - Alpha",
+  "Environment :: Console",
+  "Intended Audience :: Science/Research",
+  "License :: OSI Approved :: MIT License",
+  "Operating System :: OS Independent",
+  "Programming Language :: Python :: 3",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.11",
+  "Programming Language :: Python :: 3.12",
+  "Programming Language :: Python :: 3.13",
+  "Programming Language :: Python :: 3 :: Only",
+  "Topic :: Scientific/Engineering",
+  "Typing :: Typed",
+]
+# No runtime dependencies, matching refaudit. Tier-1 mining is pure stdlib
+# string work, and the network tiers use urllib rather than pulling in requests.
+dependencies = []
+
+[project.optional-dependencies]
+dev = ["pytest>=7", "mypy>=1.5", "ruff>=0.4"]
+
+[project.scripts]
+scholarmend = "scholarmend.cli:main"
+
+[project.urls]
+Homepage = "https://github.com/uw-share-lab/scholarmend"
+Issues = "https://github.com/uw-share-lab/scholarmend/issues"
+
+# The committed cache backs the repo's reproducibility claim, but it is 8.8 MB
+# of API responses no installed user needs. Keep it, and the agent tooling, in
+# git only.
+[tool.hatch.build.targets.sdist]
+exclude = [".scholarmend-cache", ".claude", "docs/superpowers", "BACKLOG.md", "CLAUDE.md"]
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/scholarmend"]
+artifacts = ["src/scholarmend/py.typed"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-q"
+markers = [
+  # The suite is hermetic: tier-1 miners are pure functions and the network
+  # tiers are served from the committed cache. The live check is the deliberate
+  # exception and needs real OpenReview credentials.
+  "live: talks to the real APIs; needs SCHOLARMEND_OPENREVIEW_USER and _PASSWORD",
+]
+
+[tool.ruff]
+line-length = 100
+target-version = "py310"
+
+[tool.mypy]
+python_version = "3.10"
+warn_unused_ignores = true
+disallow_untyped_defs = false