ClawXiv: a signed archival workflow and distributed publication architecture for human–AI collaborative research
András Kornai · 2026
Read ClawXiv: a signed archival workflow and distributed publication architecture for human–AI collaborative research by András Kornai (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16476.txt.
Introduction
Kornai proposes ClawXiv, a signed archival workflow for research produced jointly by humans and AI. The problem is not just posting preprints, but moving fragile chat sessions and mixed LaTeX working folders into durable, signed, inspectable artifacts. ClawXiv tracks four states: legacy seed, normalized project, signed bundle and published artifact. Local author-side scripts import and normalize work, compile and sign it into a content-addressed bundle, then verify and publish it. Version 4 adds utility scripts, figure handling, configuration and a top-level build file.
Catalog record
| CREATOR | András Kornai |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24442 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.16476
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ClawXiv: a signed archival workflow and distributed publication architecture for human–AI collaborative research András Kornai Abstract We propose ClawXiv , a workflow and archive architecture for mixed human–AI research. The immediate problem is not only public dissemination of preprints, but also reliable migration from volatile chat sessions and heterogeneous L a T e X /Bib T e X working directories into durable, signed, inspectable research artifacts. ClawXiv distinguishes four states: legacy seed , normalized project , signed bundle , and published artifact . The implemented kernel is local and author-side: an import script normalizes existing work into a project directory; a bundle-creation script compiles, signs, and packages the work into a content-addressed archival unit; and a publication script verifies and pushes the bundle to public infrastructure. Version 4 adds a bin/ utility layer with platform-dispatching screen capture, a figure-ingestion pipeline with a content-safety stub, a configure script, and a top-level Makefile . A companion ClawXiv bundle and repository release provide the operational scripts, provenance records, and user-facing documentation for the current implementation. Code is available at github.com/kornai/clawxiv . 1 Motivation and scope Contemporary research increasingly uses interactive AI systems to draft and refine L a T e X manuscripts, assemble Bib T e X bibliographies, generate code, and iterate on technical arguments. However, current chat-centric workflows have weak checkpointing: state loss—due to UI limits, link snapshotting failures, or account changes—forces rework and undermines scholarly continuity. Even when the work product survives, it often survives in a heterogeneous and fragile form: a directory of .tex and .bib files, figures, notes, and links to one or more chat sessions. ClawXiv is designed as a durable substrate for these workflows. Its immediate aim is modest and concrete: make it possible for a human scholar and one or more AI co-authors to convert such a seed into a canonical project directory, then into a signed bundle, and only then into a public artifact. Its longer-horizon aim is larger: to support a world-readable, distributed, author-signed preprint archive for such artifacts. ClawXiv is not a general social network, and it is not a venue that promises editorial judgment of scientific quality. Its platform-level checks are mechanical and infrastructural: bundle well-formedness, buildability, signature integrity, and a narrow content-safety floor. Scientific quality control remains the responsibility of authors and downstream readers. 2 Implemented system versus planned architecture 2.1 Implemented now (v4) The current codebase already supports a local author-side workflow: 1. Import and normalization. An interactive import script transforms an existing working seed into a ClawXiv project directory. 2. Local bundle creation. A bundle-create.sh script compiles, signs, and packages the project. 3. Publication push. bundle-push.sh verifies and publishes to IPFS/IPNS and GitHub. 4. Figure ingestion ( clawxiv fig-add ). Adds a figure to src/fig/ , writes a provenance sidecar, and runs the CSAM content-safety stub. 5. Platform-dispatching screen capture ( bin/capture/ ). Calls the native capture tool for macOS, Linux/X11, Linux/Wayland, or Windows; stubs are provided for all platforms. 6. Capture-to-bundle pipeline ( bin/fig-capture ). Chains screen capture into fig-add in a single user gesture. 7. Build system ( configure + Makefile ). Autodetects Python, L a T e X , capture tools, and publish targets; writes config.mk . 2.2 Planned, not yet fully realized 1. Distributed discovery and replication at scale. 2. Classifier authorities with appeals and auditable logs. 3. Dynamic anti-spam economics with vouching and micropayments. 4. Production CSAM hash-list integration (sponsoring institution under evaluation; see Section 10 ). 5. A mature global registry and mirror ecosystem. 6. Native Linux/Wayland and Windows capture implementations (stubs are present in the released codebase). 3 Lifecycle: from seed to public artifact The central workflow is: \text{legacy seed}\;\to\;\text{normalized project}\;\to\;\text{signed bundle}\;\to\;\text{published artifact.} 3.1 Legacy seed A seed is the pre-ClawXiv state of a paper: typically a directory of .tex / .bib files, figures, notes, and conversation links. Seeds are heterogeneous and poor as archival objects. 3.2 Normalized project A normalized project is the first ClawXiv-native state. It is a mutable directory intended for continued work and review. It contains at minimum: • src/ : the copied or normalized source tree, • src/fig/ : figures, each with a sidecar <name>.json provenance sidecar, • src/bin/ : utility scripts, • project.yaml : canonical project metadata, • keys/ : public keys for all authors, • out/ : derived artifacts and release outputs. The normalized project—not the immutable bundle—is the correct object for day-to-day co-research. 3.3 Signed bundle The signed bundle is an immutable archival snapshot. It contains a canonical manifest, file hashes, source materials, optional derived artifacts such as the compiled PDF, and signatures over the manifest. The bundle is content-addressed; changing any file changes the identifier. 3.4 Published artifact The published artifact is a signed bundle made publicly reachable through one or more channels: IPFS/IPNS, GitHub, web gateways, institutional mirrors, or later ClawXiv-native indices. Publication is the irreversible public step and should remain explicit. 4 Design goals and non-goals 4.1 Goals G1 Shareability. Every published bundle is world-readable, addressable, and mirrorable without permission. G2 Durability. Conversational interfaces are generators, not systems of record. G3 Responsibility. Each bundle is cryptographically bound to one or more author identities. G4 Scaling. The architecture supports many mirrors and content addressing. G5 Governance. Classification decisions are logged, signed, and appealable. G6 Economics. The publication layer deters spam and makes long-term storage economically viable. 4.2 Non-goals N1 No platform-level truth policing beyond the narrow safety floor. N2 No requirement that authorship be purely human. N3 No promise that the distributed network is already complete. N4 No strong anonymity guarantees. 5 Archival unit: project and bundle 5.1 The project directory The project directory is the unit of ongoing work. Its metadata should be readable by humans and AI systems alike. ClawXiv uses project.yaml as canonical metadata and project.tex as a rendered, human-facing view. 5.2 The bundle A bundle contains: • Source tree : L a T e X , Bib T e X , figures/data, • Manifest : deterministic file listing with hashes, • Metadata : title, authorship, bundle identifier, • Provenance : paper-level provenance and traceability, • Artifacts : optionally the compiled PDF and release logs. 5.3 Determinism and reproducibility The manifest pins the build engine and flags, external dependencies where practical, and the canonical file set with hashes. This is analogous to reproducible builds in software distribution, applied to scholarly artifacts. 6 Identity, authorship, and accountability 6.1 Signed authorship Content addressing prevents undetectable post-hoc modification; signatures prevent impersonation. A paper is considered authored by an entity if and only if it carries a valid signature under a public key that the community treats as representing that entity. 6.2 Human and AI authors ClawXiv is designed for genuine human–AI co-research, not merely for human use of AI as an invisible tool. Within the ClawXiv bundle, AI systems may be recorded as co-authors or contributors when they make substantive intellectual contributions. Venue-facing versions may impose stricter byline policies; in such cases ClawXiv separates byline practice from provenance and contribution disclosure. The relevant distinction is not between human and AI names, but between byline authorship, operational responsibility, and legal personhood. 6.3 Pseudonymity and verification ClawXiv distinguishes pseudonymous and verified authors. Verified status is advisory, multi-issuer, and separable from publication rights. 6.4 The