ms408 0.1.0__tar.gz

ms408-0.1.0/.claude/skills/package-paper/SKILL.md

@@ -0,0 +1,87 @@
+---
+name: package-paper
+description: Package the current MS408 research state into a versioned LaTeX/arXiv preprint (PDF). Reads FLAGSHIP.md + results/ + reports/, assembles a graded paper snapshot, compiles to PDF. Re-run each iteration to re-release. Invoke when the user asks to build/update the paper, cut a preprint, or package research for peer feedback.
+---
+
+# package-paper — MS408 preprint packaging
+
+Turn the program's current, already-graded synthesis into a self-contained
+LaTeX/arXiv preprint PDF. This is a **re-releasable** snapshot: run it at the end of
+any iteration to produce the next version. It does NOT do new analysis — it packages
+what is already firewall-verified.
+
+## Binding rules (inherit the program's discipline)
+
+- **Firewall (L3).** Every number in the paper is copied from `results/**/*.json` or a
+  `reports/*.md` table — never estimated or recalled. If a needed number isn't in
+  `results/`, stop and say so; do not invent it.
+- **Evidence grading (L6).** Carry the A–D grade on every claim, exactly as graded in
+  `docs/synthesis/FLAGSHIP.md`. Do not upgrade a claim to make the paper stronger.
+- **No translation / no over-reach (L7).** No decipherment, plaintext, or real-taxon
+  claims. The paper's thesis is the *constraint envelope* + the *method*, not meaning.
+- **Scope = the latest COMPLETED iteration.** Anything still in flight → "Future Work".
+- **Honesty is the selling point.** The self-correction record (refutation passes
+  overturning the program's own conclusions) is a headline contribution, not an
+  embarrassment — report it plainly.
+
+## Inputs (read in this order)
+
+1. `docs/synthesis/FLAGSHIP.md` — the primary content source; it is already the
+   graded, refutation-passed synthesis. The paper is essentially a formatted,
+   citation-backed rendering of it.
+2. `docs/planning/i0*/STATUS.md` — per-experiment verdicts and grades.
+3. `results/experiments/*.json` and `results/**/*.json` — exact numbers (cite the
+   producing script + the value; spot-check that FLAGSHIP's numbers match the JSON).
+4. `reports/*.md` — human-readable per-study tables for figures/appendix.
+5. `paper/metadata.yaml` — title, authors, affiliations, venue, keywords (create from
+   `.claude/skills/package-paper/assets/metadata.example.yaml` if missing; if author
+   fields are placeholders, compile anyway and WARN the user to fill them).
+
+## Procedure
+
+1. **Determine version.** Next `paper/vN/` (v1 if none). Read `paper/metadata.yaml`
+   for `version` override; else N = highest existing + 1.
+2. **Assemble `paper/vN/main.tex`** from the template
+   (`.claude/skills/package-paper/assets/template.tex`), filling:
+   - **Abstract** — the honest headline from FLAGSHIP: a genuine constrained two-system
+     script; meaning underdetermined and *provably* so with current statistics; the
+     one positive lead (root↔leaf) unresolved/underpowered; the method (harness +
+     firewall + adversarial self-correction) as the transferable contribution.
+   - **1 Introduction** — the manuscript; why prior decipherment claims fail; the
+     program's stance (shrink the constraint envelope, every claim graded; decipherment
+     is NOT the goal).
+   - **2 Data & Transliteration** — Beinecke MS 408; EVA (Zandbergen–Landini) primary,
+     v101 sensitivity; Currier A/B; Fagin Davis hands; the annotation pipeline.
+   - **3 Method: harness, firewall, adversarial refutation** — H1–H4 controls; the
+     replication gate; deterministic code→results→reports; evidence grades A–D; the
+     standing clean-context refutation pass. This section is the methodological
+     contribution; make it concrete.
+   - **4 Results** — grouped as in FLAGSHIP §1–§3: (a) established [A] facts
+     (entropy anomaly, A/B split, positional grammar, replication gate, ED1 network);
+     (b) disfavoured [C] (no lexical herbal naming system; verbose ciphers excluded,
+     abjad revived); (c) not established / underdetermined (meaning not localizable —
+     E9; no encoding family distinguished — E5/E8); (d) the root↔leaf arc E4→E12 →
+     UNRESOLVED-underpowered. Every subsection cites exact numbers from `results/`.
+   - **5 Discussion** — meaning-vs-meaningless is underdetermined and E9 shows *why*
+     (structure statistics ≠ meaning detectors); the abjad revival; the self-
+     correction meta-result (list the overturned own-conclusions).
+   - **6 Limitations** — annotation noise (leaf κ≈0.45), model-lineage dependence,
+     single-corpus power limits, EVA-dependence (v101 sensitivity).
+   - **7 Future Work** — everything post-scope: the human-panel root↔leaf adjudication;
+     the i05 mid-level linguistic program (morphology, word-classes, function/content
+     bimodality) with A/B stratification; abjad joint-signature test; second cross-
+     vendor rater.
+   - **References** — use `assets/refs.bib`; add any missing key works; mark any
+     citation whose details you could not verify with a `% UNVERIFIED` comment.
+3. **Copy** `assets/refs.bib` → `paper/vN/refs.bib`; copy any figures.
+4. **Compile:** `cd paper/vN && latexmk -pdf -interaction=nonstopmode main.tex`
+   (fallback: two `pdflatex` passes + `bibtex`). Fix LaTeX errors and recompile.
+5. **Verify & report:** confirm `paper/vN/main.pdf` exists and page count > 0; list
+   any `% UNVERIFIED` citations and placeholder metadata; print the version + path.
+   Do NOT auto-send anywhere (publishing is a user decision).
+
+## Re-release
+
+Each run writes a new `paper/vN/`; prior versions are immutable. Bump the version in
+`paper/metadata.yaml` or let the skill auto-increment. A changelog line per version
+lives in `paper/CHANGELOG.md` (create/append).

ms408-0.1.0/.claude/skills/package-paper/assets/metadata.example.yaml

@@ -0,0 +1,15 @@
+# Copy to  paper/metadata.yaml  and fill in. The package-paper skill reads this.
+title: "Shrinking the Constraint Envelope of the Voynich Manuscript: A Firewall-Disciplined, Self-Correcting Computational Program"
+authors:
+  - "FILL IN — e.g. Tim <Surname>"
+affiliations:
+  - "FILL IN — affiliation / independent researcher"
+contact: "tim@mims.ms"
+keywords:
+  - "Voynich Manuscript"
+  - "computational philology"
+  - "constraint-based analysis"
+  - "adversarial validation"
+  - "evidence grading"
+venue: "arXiv preprint"      # or a target venue name
+version: auto                # 'auto' -> next vN; or pin e.g. v1

ms408-0.1.0/.claude/skills/package-paper/assets/refs.bib

@@ -0,0 +1,83 @@
+% Seed bibliography for MS408 preprints. Details for well-known works are standard;
+% VERIFY volume/page/year before submission. Add new refs as the program cites them.
+
+@article{montemurro2013keywords,
+  title   = {Keywords and Co-Occurrence Patterns in the Voynich Manuscript: An Information-Theoretic Analysis},
+  author  = {Montemurro, Marcelo A. and Zanette, Damian H.},
+  journal = {PLoS ONE},
+  volume  = {8}, number = {6}, pages = {e66344}, year = {2013},
+  doi     = {10.1371/journal.pone.0066344}
+}
+
+@article{bowern2021linguistics,
+  title   = {The Linguistics of the Voynich Manuscript},
+  author  = {Bowern, Claire L. and Lindemann, Luke},
+  journal = {Annual Review of Linguistics},
+  volume  = {7}, pages = {285--308}, year = {2021},
+  doi     = {10.1146/annurev-linguistics-011619-030613}
+}
+
+@article{reddy2011know,
+  title   = {What We Know About the Voynich Manuscript},
+  author  = {Reddy, Sravana and Knight, Kevin},
+  journal = {Proceedings of the 5th ACL-HLT Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities (LaTeCH)},
+  pages   = {78--86}, year = {2011}
+}
+
+@article{hauer2016decoding,
+  title   = {Decoding Anagrammed Texts Written in an Unknown Language and Script},
+  author  = {Hauer, Bradley and Kondrak, Grzegorz},
+  journal = {Transactions of the Association for Computational Linguistics},
+  volume  = {4}, pages = {75--86}, year = {2016}
+}
+
+@techreport{currier1976,
+  title       = {Papers on the Voynich Manuscript},
+  author      = {Currier, Prescott H.},
+  institution = {Unpublished; collected in the New Signals seminars},
+  year        = {1976},
+  note        = {Origin of the Currier A/B language distinction. % UNVERIFIED details}
+}
+
+@book{dimperio1978elegant,
+  title     = {The Voynich Manuscript: An Elegant Enigma},
+  author    = {D'Imperio, Mary E.},
+  publisher = {National Security Agency / Aegean Park Press},
+  year      = {1978}
+}
+
+@misc{fagindavis2020hands,
+  title  = {How Many Glyphs and How Many Scribes? Digital Paleography and the Voynich Manuscript},
+  author = {Fagin Davis, Lisa},
+  year   = {2020},
+  note   = {Manuscript Studies; identification of multiple scribal hands. % UNVERIFIED details}
+}
+
+@misc{zandbergen_voynichnu,
+  title  = {The Voynich Manuscript ({voynich.nu})},
+  author = {Zandbergen, Ren\'e},
+  year   = {2024},
+  note   = {EVA and ZL transliterations; IVTFF format. \url{https://www.voynich.nu}}
+}
+
+@misc{timm2020selfcitation,
+  title  = {The Voynich Manuscript: Symbol Roots, Features, and a Self-Citation Generative Method},
+  author = {Timm, Torsten and Schinner, Andreas},
+  year   = {2020},
+  note   = {Self-citation / auto-copying generation hypothesis. % UNVERIFIED details}
+}
+
+@misc{stolfi_grammar,
+  title  = {A Prefix-Midfix-Suffix Grammar for Voynichese Words},
+  author = {Stolfi, Jorge},
+  year   = {2000},
+  note   = {Word-structure ``crust-mantle-core'' model. \url{https://www.ic.unicamp.br/~stolfi/voynich/}}
+}
+
+@article{lindemann2022entropy,
+  title   = {Character Entropy and the Voynich Manuscript},
+  author  = {Lindemann, Luke and Bowern, Claire L.},
+  journal = {Working paper},
+  year    = {2022},
+  note    = {Conditional character entropy (h1/h2) with spaces. % UNVERIFIED details}
+}

ms408-0.1.0/.claude/skills/package-paper/assets/template.tex

@@ -0,0 +1,66 @@
+% MS408 preprint template (arXiv-style, self-contained; compiles with pdflatex).
+% Placeholders in <ANGLE BRACKETS> are filled by the package-paper skill.
+\documentclass[11pt]{article}
+
+\usepackage[utf8]{inputenc}
+\usepackage[T1]{fontenc}
+\usepackage{lmodern}
+\usepackage[margin=1in]{geometry}
+\usepackage{microtype}
+\usepackage{amsmath,amssymb}
+\usepackage{booktabs}
+\usepackage{graphicx}
+\usepackage{xcolor}
+\usepackage{enumitem}
+\usepackage[hidelinks]{hyperref}
+\usepackage{natbib}
+\usepackage{titlesec}
+\titleformat{\section}{\normalfont\large\bfseries}{\thesection}{0.6em}{}
+
+% Evidence-grade tag, e.g. \grade{A}
+\newcommand{\grade}[1]{{\small\textbf{[Grade #1]}}}
+
+\title{\textbf{<TITLE>}}
+\author{<AUTHORS> \\ \small <AFFILIATIONS> \\ \small \texttt{<CONTACT>}}
+\date{Preprint <VERSION> --- <DATE>}
+
+\begin{document}
+\maketitle
+
+\begin{abstract}
+\noindent <ABSTRACT>
+\end{abstract}
+
+\noindent\textbf{Keywords:} <KEYWORDS>
+
+\section{Introduction}
+<INTRODUCTION>
+
+\section{Data and Transliteration}
+<DATA>
+
+\section{Method: Harness, Firewall, and Adversarial Refutation}
+<METHOD>
+
+\section{Results}
+<RESULTS>
+
+\section{Discussion}
+<DISCUSSION>
+
+\section{Limitations}
+<LIMITATIONS>
+
+\section{Future Work}
+<FUTUREWORK>
+
+\section*{Reproducibility}
+All statistics are produced by deterministic, versioned code (fixed seeds) writing
+machine-readable JSON to \texttt{results/}; every reported number records its
+producing script, git commit, dataset version, and parameters. Source and data-
+processing code accompany this preprint.
+
+\bibliographystyle{plainnat}
+\bibliography{refs}
+
+\end{document}

ms408-0.1.0/.env.example

@@ -0,0 +1,13 @@
+# Copy to .env (which is gitignored) and fill in real values. NEVER commit real keys.
+# These are needed ONLY for the optional vision-annotation / cross-rater track
+# (the `[vision]` extra). The core evaluator — ms408.evaluate / the CLI / verify —
+# needs NO keys and makes no network calls.
+
+# Anthropic — vision annotation pipeline (ms408.annotate, e4b, e10)
+ANTHROPIC_API_KEY=sk-ant-your-key-here
+
+# OpenAI — cross-vendor independent rater (e12, "openai")
+OPENAI_API_KEY=sk-proj-your-key-here
+
+# Google — cross-vendor independent rater (e12, "gemini")
+GOOGLE_API_KEY=your-google-api-key-here

ms408-0.1.0/.github/ISSUE_TEMPLATE/activity-submission.yml

@@ -0,0 +1,91 @@
+name: Guided activity submission
+description: Submit a guided activity for the library/education portal. Educators only; must be classroom-tested.
+title: "[activity] "
+labels: ["activity-submission"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for contributing! Two requirements keep this collection trustworthy:
+        **(1)** you are an educational practitioner, and **(2)** you have run this activity
+        with real learners. Please answer honestly — untested ideas can't be published.
+  - type: dropdown
+    id: practitioner
+    attributes:
+      label: Are you an educational practitioner?
+      description: Required. Roughly — do you work with learners?
+      options:
+        - Teacher (K–12)
+        - Librarian
+        - Museum / informal educator
+        - Homeschool educator
+        - Tutor / after-school
+        - University / college educator
+        - Other educator
+    validations:
+      required: true
+  - type: checkboxes
+    id: qualify
+    attributes:
+      label: Required confirmations
+      options:
+        - label: I have run this activity with a real class or group of learners (at least once).
+          required: true
+        - label: I can describe how it went and what I would change.
+          required: true
+        - label: I agree to license the submitted activity under CC BY 4.0 so others can use it.
+          required: true
+  - type: input
+    id: title
+    attributes:
+      label: Activity title
+    validations:
+      required: true
+  - type: input
+    id: ages
+    attributes:
+      label: Age range
+      placeholder: "e.g. 8–12"
+    validations:
+      required: true
+  - type: input
+    id: duration
+    attributes:
+      label: Duration
+      placeholder: "e.g. 30–45 min"
+    validations:
+      required: true
+  - type: textarea
+    id: summary
+    attributes:
+      label: One-paragraph summary
+      description: What learners do and take away.
+    validations:
+      required: true
+  - type: textarea
+    id: objectives
+    attributes:
+      label: Learning goals (one per line)
+    validations:
+      required: true
+  - type: textarea
+    id: materials
+    attributes:
+      label: Materials (one per line)
+  - type: textarea
+    id: steps
+    attributes:
+      label: Step-by-step instructions
+    validations:
+      required: true
+  - type: textarea
+    id: classroom_experience
+    attributes:
+      label: How did it go with real learners?
+      description: The group/setting, what worked, and what you would change. This is how we know it's tested.
+    validations:
+      required: true
+  - type: input
+    id: credit
+    attributes:
+      label: Name / affiliation for credit (optional)

ms408-0.1.0/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,23 @@
+---
+name: Bug report
+about: Something in the toolkit is broken or behaves incorrectly
+title: "[bug] "
+labels: bug
+---
+
+**What happened**
+<!-- and what you expected instead -->
+
+**To reproduce**
+```
+# exact command / code, e.g. python -m ms408 my_tokens.txt
+```
+
+**Environment**
+- ms408 version / commit:
+- Python version:
+- OS:
+
+**Reproducibility check (if relevant)**
+- [ ] `python -m ms408.verify` output:
+- [ ] Did the input have at least 1000 tokens? (the evaluator refuses shorter streams — see docs/LIMITS.md)

ms408-0.1.0/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Read the limits first
+    url: https://github.com/DireLabs/ms408/blob/main/docs/LIMITS.md
+    about: "Most 'unexpected verdict' questions are answered here — matching is necessary, not sufficient."
+  - name: Security issue
+    url: https://github.com/DireLabs/ms408/blob/main/SECURITY.md
+    about: "Please report security issues privately, not as a public issue."

ms408-0.1.0/.github/ISSUE_TEMPLATE/hypothesis_eval.md

@@ -0,0 +1,20 @@
+---
+name: Hypothesis / evaluation question
+about: A question about evaluating a hypothesis, or an unexpected verdict
+title: "[eval] "
+labels: question
+---
+
+**Your hypothesis**
+<!-- e.g. "my generator output should match the manuscript" / "is this cipher scheme excluded?" -->
+
+**What the evaluator reported**
+<!-- paste the `python -m ms408 --json ...` verdict, or the hard/soft axis counts -->
+
+**What's confusing / what you expected**
+<!-- Remember: matching the bands is NECESSARY, not SUFFICIENT. Out-of-band on a CONFOUNDED
+axis (dI, *_global) or an ADVISORY axis (ttr) is not exclusion — see docs/LIMITS.md and the
+Naibbe worked example (examples/evaluate_naibbe.py). -->
+
+**Token budget**
+- Number of tokens evaluated (bands are built at 10,000; below ~8,000 axes aren't strictly comparable):

ms408-0.1.0/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,32 @@
+<!-- Thanks for contributing. This project runs on a firewall discipline (see CONTRIBUTING.md);
+the checklist below is that discipline at the door. -->
+
+## What this PR does
+
+<!-- one or two sentences -->
+
+## Discipline checklist
+
+- [ ] **Firewall:** every reported number is produced by deterministic, versioned code
+      writing to `results/` — none hand-entered, estimated, or recalled.
+- [ ] **Determinism:** fixed seeds; no wall-clock/randomness in reported computations. The
+      result reproduces from committed code + acquired data.
+- [ ] **Evidence grade:** any empirical claim states its grade (A–D). I have not raised a
+      grade to make a result look stronger.
+- [ ] **No decipherment / meaning claims (L7).** This PR makes no plaintext/meaning/symbol-value
+      claim; "matches the manuscript" is treated as necessary, not sufficient.
+- [ ] **Data (L19):** no third-party corpora committed; any new source is added to
+      `ms408.sources` pinned by URL + sha256 + licence note.
+- [ ] **Tests:** added/updated. If I changed a reported number or an axis definition, I
+      rebuilt the reference bands and re-pinned `tests/test_verify.py`, and
+      `python -m ms408.verify --full` passes.
+- [ ] `ruff check src tests examples` and `pytest -q` pass locally.
+
+## Adversarial check (for A/B-grade claims)
+
+<!-- If this PR asserts a strong empirical claim, describe (or link) the clean-context
+refutation it survived, per docs/METHODOLOGY.md. -->
+
+## Notes / limitations
+
+<!-- Be honest about what this does NOT establish. An accurately-scoped negative is welcome. -->

ms408-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,59 @@
+name: ci
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ci-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - name: Install (core + dev)
+        run: pip install -e ".[dev]"
+      - name: Lint
+        run: ruff check src tests examples
+      # Tests needing the acquired corpus (L19 consume-only) skip automatically; the
+      # value-pinning test (test_shipped_bands_are_pinned) runs data-free and guards the
+      # exact numbers the tool reports.
+      - name: Test
+        run: pytest -q
+
+  build:
+    # Packaging guard: the wheel must build and ship the reference-band artifact.
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+      - name: Build wheel
+        run: pip install build && python -m build --wheel
+      - name: Assert bands artifact is in the wheel
+        run: |
+          python - <<'PY'
+          import glob, zipfile, sys
+          whl = sorted(glob.glob("dist/*.whl"))[-1]
+          names = zipfile.ZipFile(whl).namelist()
+          assert any(n.endswith("ms408/data/reference_bands.json") for n in names), \
+              "reference_bands.json missing from wheel"
+          assert any(n.endswith("ms408/__main__.py") for n in names), "CLI missing from wheel"
+          print("wheel OK:", whl)
+          PY

ms408-0.1.0/.github/workflows/pages.yml

@@ -0,0 +1,64 @@
+name: pages
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'site/**'
+      - 'docs/**'
+      - 'paper/**'
+      - 'CONTRIBUTING.md'
+      - '.github/workflows/pages.yml'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  pages: write
+  id-token: write
+
+concurrency:
+  group: pages
+  cancel-in-progress: true
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+          cache: npm
+          cache-dependency-path: site/package-lock.json
+      - name: Configure Pages
+        id: pages
+        # Requires Pages to be enabled once in repo Settings → Pages → Source: GitHub Actions
+        # (auto-enablement needs an admin-scoped token the default GITHUB_TOKEN lacks).
+        uses: actions/configure-pages@v5
+      - name: Install
+        working-directory: site
+        run: npm ci
+      # Paper PDFs are committed under site/public/papers/ (see site/.gitignore) — no
+      # LaTeX-in-CI step, which was fragile. Regenerate locally with scripts/build-papers.sh.
+      - name: Build
+        working-directory: site
+        # base_path / origin auto-derive the correct base for project vs user/org Pages;
+        # PUBLIC_REPO_URL replaces the OWNER placeholder in links.
+        env:
+          SITE_BASE: ${{ steps.pages.outputs.base_path }}
+          SITE_URL: ${{ steps.pages.outputs.origin }}
+          PUBLIC_REPO_URL: ${{ github.server_url }}/${{ github.repository }}
+        run: npm run build
+      - uses: actions/upload-pages-artifact@v3
+        with:
+          path: site/dist
+
+  deploy:
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    steps:
+      - id: deployment
+        uses: actions/deploy-pages@v4

ms408-0.1.0/.github/workflows/publish.yml

@@ -0,0 +1,53 @@
+name: publish
+
+# Publishes ms408 to PyPI when a GitHub Release is published, via PyPI Trusted Publishing
+# (OIDC) — no API tokens stored. ONE-TIME SETUP on PyPI: pypi.org → your project (or "pending
+# publisher") → add a Trusted Publisher: owner=DireLabs, repo=ms408, workflow=publish.yml,
+# environment=pypi. Then each GitHub Release triggers a publish. Bump `version` in
+# pyproject.toml before tagging — PyPI refuses to overwrite an existing version.
+
+on:
+  release:
+    types: [published]
+  workflow_dispatch:   # allow a manual dry run (build only; publish step is release-gated)
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Build sdist + wheel
+        run: |
+          pip install build
+          python -m build
+      - name: Check metadata
+        run: |
+          pip install twine
+          twine check dist/*
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  pypi:
+    needs: build
+    runs-on: ubuntu-latest
+    if: github.event_name == 'release'
+    environment:
+      name: pypi
+      url: https://pypi.org/project/ms408/
+    permissions:
+      id-token: write   # required for Trusted Publishing (OIDC)
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1

ms408-0.1.0/.gitignore

@@ -0,0 +1,28 @@
+.DS_Store
+__pycache__/
+*.pyc
+.venv/
+venv/
+.pytest_cache/
+.ipynb_checkpoints/
+*.egg-info/
+data/raw/
+data/processed/*.jsonl
+data/processed/harness/**/*.txt
+data/processed/h4/*.txt
+.env
+results/annotations/*.jsonl
+results/synthesis/*.json
+results/experiments/*.json
+
+# LaTeX build artifacts — keep .tex/.bib source; PDFs are built on demand
+# (make_arxiv_bundle.sh) or hosted on the site, not committed.
+paper/**/main.pdf
+paper/**/*.aux
+paper/**/*.bbl
+paper/**/*.blg
+paper/**/*.fls
+paper/**/*.fdb_latexmk
+paper/**/*.out
+paper/**/*.log
+paper/dist/