Falsify Paper Workbench

Browser-local RED and BLUE scientific-document experiments, independent human review, and versioned user-defined harnesses. RED and BLUE may be used independently.

Falsify orchestrator planning reference: External AI RED BLUE North Star. It records the confirmed workflow and implementation plan for that coordinator; it is not required reading or a shared instruction for other agents. Implementation status is tracked separately.

This repository contains only the application code. Private project notes, research papers, answer keys, meeting recordings and execution records are not included. The original private research repository remains private.

The hosted frontend is static. Files are parsed in your browser; original local drafts and saved runs stay local until you explicitly publish a selected case to Confluence. Confluence is the only remote archive, and publication means sharing with the Falsify team. Fetched records open temporarily in native review, without automatic persistent browser copies; publish new review notes explicitly before closing or reloading. A thin OAuth/API relay stores no experiment database or files. Historical private records are not automatically published. Explicit execution sends the supplied paper text through your own local CLI or selected trusted harness to its model provider. GitHub is not an execution backend or shared experiment database.

Use the hosted platform and quick-first HTML guide. Follow local companion setup to connect your own Codex CLI or Claude Code subscription login, or use manual JSON exchange. Optional BLUE Evidence v0.3.3 setup uses each participant's own CLI login. Windows users can run start-companion.ps1 from an extracted source download, or once install the optional per-user falsify://connect click-to-launch link with install-connection.ps1 (no administrator rights; uninstall-connection.ps1 removes it). macOS and Linux users can prepare Python 3 and install the handler as described in the macOS connection guide. Click-pairing asks for local approval without copying the process token. Manual token paste remains available as a fallback; pairing tokens are never put in a link. No account credentials or API keys should be entered into a harness JSON or committed to this repository.

Deployment architecture · Harness protocol

Third-party components

The original Workbench application code is licensed under AGPL-3.0-only; see LICENSE and NOTICE. This grants no license to user-uploaded papers, private harnesses, recordings or research records. The complete application source is available in this repository, including the static frontend and local companion.

Browser parser license notices are retained in app/dist/vendor/. Python dependencies are installed by the user, not bundled. Short browser edits use MuPDF; paragraph edits reuse Falsify PDF Editor and BentoPDF/PDFium under AGPL-3.0-only. Since alpha.14, its pinned JS/WASM and corresponding source overlay are shipped to GitHub Pages and run in a browser worker. Editing needs no PC connection or model call. The optional local companion export adapter remains available. See PDF editor setup and limits.

Alpha.16 includes the features described below. Use a matching companion for new local execution features; sourceBuildVerified remains false. Alpha.15 adds bounded flow through compatible existing single-column pages before terminal pages, with original source/content page mappings and explicit table, figure and link-target barriers. The grid-icon 표 셀 편집 mode reuses selected Stirling MIT table modules to locate cells; a human edit is bound to exact source text and geometry. Cell sizes/borders do not automatically expand. Mixed table/body export records a table-first, then paragraph-stage hash/status chain; a readable failed stage stays marked failed or draft, and pending body edits remain explicit.

The optional local companion can replace one opaque raster placement inside supported simple nested Forms, isolating shared uses. This remains separate from browser text/table export; arbitrary figures, multi-column layouts and moving tables/figures together with captions are outside the accepted scope. In the real cold-shower eight-edit check, 15 original pages still yield 18 output pages and zero compacted flow groups because table and layout barriers remain. See implementation and reuse boundaries and figure replacement limits.

Model results and human reviews are not scientific gold, domain approval, or verified performance metrics. Custom harness code must be trusted: the companion is not an OS sandbox.

The guide covers bounded DOCX, HTML and ODT export, PDF visual input for built-in Codex, Web Locks admission and condition-matched Harness comparison. ODT visual rendering is unverified; server restart recovery is not implemented. See the guide for limits.