Causal Capybara v0.2.0 — Windows preview

Causal Capybara v0.2.0 — Windows preview

Windows 64-bit installer, with Python and analysis packages included:
Causal-Capybara-0.2.0-windows-x64-setup.exe under Assets.

Download the Windows installer (.exe)

Windows preview release — 26 October 2026. Use the installer asset for the desktop
app; GitHub's Source code archives require a development setup.

New since v0.1.0

• Offline research tree. Explore 933 linked records and 5,028 relationships,
  including 79 methods, 208 concepts and 119 diagnostics. Follow assumptions,
  estimands, source metadata, evidence notes and explicit coverage gaps.
• Conversational research assistant. Keep a conversation beside your data,
  design or results. Conversations and unsent drafts are saved locally per study.
  Ask follow-up questions, inspect the selected context, and copy or save answers
  as Markdown. Sources-only retrieval works without a model provider.
• Optional AI integrations. Use configured native Codex, Claude Code or Grok
  CLI clients, Ollama or supported cloud APIs. External providers require separate
  setup and may receive the context you choose to send. Study changes are shown
  as a validated proposal and require approval; running an analysis is separate.
• Survival and heterogeneity analyses. Added right-censored Kaplan–Meier,
  RMST and log-normal/Weibull AFT workflows, plus direct heterogeneity tests across
  continuous, binary, count/rate, categorical and survival outcomes. The app
  labels effect scales, assumptions and the scope of each comparison.
• Workflow improvements. Better browser imports, reversible column selection,
  question setup, package-install progress, result comparisons, keyboard access
  and handling of assistant requests when switching views or studies.
• Reproducibility and knowledge tools. Direct run provenance, numerical
  reproduction checks, an independently installable read-only knowledge MCP and
  a companion research skill. The knowledge pack keeps its separate version 0.3.0.

Install

1. Download the .exe under Assets and open it.
2. Follow the installer, then launch Causal Capybara and try a worked example.
3. Python and its analysis dependencies are bundled. R and AI clients are optional
   separate installations. Setup may need internet access for Microsoft WebView2.

The installer is unsigned; Windows may display an unknown-publisher warning.
An SHA-256 checksum accompanies the installer in SHA256SUMS.txt.

See installation help,
the assistant guide
and knowledge client setup.

Verification

The installer was built from tag v0.2.0, commit
b9f2a8df0d1946987c741a67cb2730420be27332, in the
tagged Windows release workflow.

• 95 frontend tests, TypeScript checking, the production UI build and the Rust
  bootstrap test passed in the tagged build. The frontend tests and production
  build also passed locally on 26 October 2026.
• The bundled CPython 3.13.15 suite passed 1,279 tests, with 6 skipped and 7
  warnings. Optional MCP protocol/client tests and Python/R concordance are
  outside that bundled-runtime suite.
• The staged and installed runtime passed knowledge validation, cited assistant
  context retrieval, a worked example with two estimators and result validation.
• The exact installer passed isolated silent installation, desktop startup and
  health, bundled-engine ownership, graceful shutdown, child-process cleanup,
  resource-tree checks and uninstall on the Windows CI runner.
• License notices were collected for the exact bundled dependencies.
• The source CI run
  passed frontend, Python on Windows/Linux/macOS, Windows desktop build and
  Python/R concordance. Its combined browser job previously reached its 20-minute
  limit; that hosted result does not establish complete browser acceptance.
• On 26 October, the unchanged release source passed both live-engine and
  synthetic ontology browser suites locally on Windows, including accessibility,
  keyboard navigation, cited context, redaction, provider consent and failure,
  cancellation, approval boundaries, stale proposals and survival results.
• The same local run passed the general workflow/accessibility, import, policy,
  heterogeneity and CLI browser suites. The 13 CLI checks use real sidecar context
  with mocked native clients/providers; they do not certify external sign-in.

Installer: Causal-Capybara-0.2.0-windows-x64-setup.exe — 135,140,903 bytes
(about 135 MB).

SHA-256: b1105aca3f4f83e152c014b199b09721309603da8d674dcdaf632b41f8d328aa

A fresh download on 26 October matched both SHA256SUMS.txt and GitHub's asset
digest. Publication documentation is updated separately; the release tag and
installer remain tied to the build commit above. The automated evidence does not
certify fresh-account interactive use or display scaling.

Known limits

• Windows x64 is the desktop package for this preview. Fresh-account installation,
  a separate Windows 11 check and 150%/200% scaling verification remain pending.
  macOS, Linux and Windows ARM64 desktop packages are not included.
• AI providers, native CLI sign-in, R and a TeX toolchain for PDF export are not
  bundled. Sources-only guidance and synthetic examples are available offline.
• Knowledge coverage is uneven: 44 of 93 method/probe entries link to scoped
  evidence claims, 66 source references remain unresolved, and independent human
  scientific review is not recorded. Source records often contain metadata rather
  than full publication text. A graph link does not establish causal identification.
• Review the identifying assumptions, diagnostics and implementation limits for
  the analysis you run. The knowledge graph does not assess your study for you.

Project code is Apache-2.0 licensed.
Bundled dependencies retain their own licenses and notices. Report bugs through
the bug form
using redacted examples; use SECURITY.md
for the private vulnerability-reporting route.

Causal Capybara v0.1.0 — Windows preview

Causal Capybara v0.1.0

Download the Windows 64-bit installer (.exe)

Windows preview — 5 October 2026

The installer is also listed under Assets below as
Causal-Capybara-0.1.0-windows-x64-setup.exe (about 132 MB).
The download is public and does not require a GitHub account.
Choose the .exe to install the desktop app; the automatically generated
Source code archives are for source installation.

Python and the analysis packages are included. No programming setup required.

1. Download the .exe using the link above, then double-click it.
2. Choose Next, review the Apache license, and choose I Agree.
3. Keep the default folder for your Windows user and choose Next. When setup completes, choose Next again.
4. Leave Run Causal Capybara checked and choose Finish. Open a worked example from Home.

The app installs for your Windows user. If Microsoft WebView2 is missing, setup
downloads that component, so keep an internet connection available during installation.
The preview is unsigned; Windows may show an unknown-publisher warning.

Prefer a source installation? The command-line alternative
clones the v0.1.0 source and runs locally in a browser, using Git, Python 3.12,
and Node.js 22.12 or newer. An analysis-only option needs Git and Python. Managed devices
must permit these tools and local servers.

See installation help
for Windows warnings, troubleshooting, optional checksum verification, and developer setup.
The SHA-256 checksum is provided in SHA256SUMS.txt under Assets.

Causal Capybara is a local desktop studio for applied causal inference. Build a
causal question, choose a research design, inspect diagnostics, compare estimators,
and export the analysis and its methodological references.

Included

• Causal-question workspace with design cards, diagnostics, an assumption ledger,
  multiple estimation methods, comparisons, reproducible code, and reports.
• Searchable method and literature catalogues available even without an analysis engine.
• Synthetic worked examples that run offline. Published study data is fetched from
  its publisher only when requested and retains that publisher's reuse terms.
• Bundled Python engine and packages. R methods are optional and require a separate installation.

Local validation

The local Windows x64 installer built from commit 90cb5e4 bundles CPython 3.13.15.
Its SHA-256 is f1773c3888d0e4b90e9fb96c067d888af872a186b67d3da0fffce52c4935d313.
Local checks installed it without system Python on PATH, ran an example analysis,
started the desktop app with its bundled engine, confirmed clean shutdown, and
uninstalled it. The current bundled Python run passed 307 checks; a cache-location
assertion caused by the audit harness passed after correcting its cache directory
(308 checks across the run and rerun). Frontend tests (38), TypeScript checking,
the Rust test, and production builds passed. The earlier local suite with R passed
311 tests. Automatic launch from Finish was slow on this host; a fresh-machine
timing check remains pending.

At the September 30 audit, the hosted build could not start because of a
GitHub Actions account restriction. Hosted CI for current main at a1433ab
passed all seven jobs on 5 October, including Windows desktop, browser workflow,
Python engines, UI, and cross-engine concordance. This does not replace the
installer-specific checks above: the download is the locally audited installer,
and its checksum was checked again before publication. A fresh-account
interactive test remains pending in the
release checklist.

Known limits

• Windows x64 is the only desktop download. Windows 10 x64 has been tested;
  a separate Windows 11 clean-machine test is pending. macOS, Linux, and Windows
  ARM64 desktop packages are not included.
• The installer is unsigned. A checksum confirms file integrity; it does not
  establish publisher identity or remove Windows warnings.
• R is optional and not bundled. PDF export needs a separate TeX toolchain;
  without it, export can provide LaTeX source. Word, HTML, and Markdown exports
  use the included dependencies.
• Fresh-account installation, the remaining native GUI checks, and Windows display
  scaling remain pending. This is an unsigned preview,
  and publication does not certify those checks. See the
  final audit.

The original project is Apache-2.0 licensed, with
Copyright 2026 Yusufhan Balci; VCESR. Third-party components keep their own
licenses and notices. For bugs, use the bug-report form
without attaching private data. Report vulnerabilities through the private route in
SECURITY.md.