ClawXiv: a signed archival workflow and distributed publication architecture for human--AI collaborative research

This paper introduces ClawXiv, a local-first, signed archival workflow and distributed publication architecture designed to transform volatile human-AI collaborative research sessions and heterogeneous files into durable, inspectable, and content-addressed artifacts through a four-state lifecycle.

The Gist

Imagine you are writing a research paper, but instead of just you typing, you are having a deep, ongoing conversation with a very smart AI. You bounce ideas back and forth, the AI writes code, you fix the math, and together you create something new.

The Problem:
Right now, if you do this in a chat window or a messy folder on your computer, it's like building a sandcastle right before the tide comes in. If the chat history gets deleted, your computer crashes, or you lose a link, your work is gone. Even if you save the files, they are often a messy pile of documents, images, and chat logs that don't tell the full story of how you made it.

The Solution: ClawXiv
The paper proposes ClawXiv (pronounced "claw-HIVE" — rhymes with archive, since the X is a 'hard h' like in TeX or loch). Think of it as a "Time-Traveling Digital Vault" specifically designed for human-AI teamwork. It turns that messy, fragile process into a solid, unchangeable, and trustworthy record.

Here is how it works, using simple analogies:

1. The Four Stages of a Paper

ClawXiv treats a research paper like a product going through a factory line, moving through four distinct states:

• The Seed (The Raw Ingredients): This is your messy starting point. It's the folder with your half-finished notes, the chat logs, and the scattered images. It's like a pile of flour, eggs, and sugar on a counter. It's useful for cooking, but you can't serve it to guests yet.
• The Normalized Project (The Organized Kitchen): You run a script (a digital helper) that cleans up the mess. It puts all the ingredients in the right bowls, labels them, and organizes the recipe. Now, humans and AI can still edit and cook together here.
• The Signed Bundle (The Sealed Jar): This is the big moment. You take your organized project, lock it in a jar, and put a tamper-proof wax seal on it.
  • Why seal it? Because once sealed, if anyone tries to change even a single letter inside, the seal breaks, and everyone knows it was tampered with.
  • The Signature: Both the human and the AI "sign" this jar. This proves who made it and that the AI actually did the work, not just a human pretending to use a tool.
• The Published Artifact (The Museum Exhibit): The sealed jar is placed in a public museum (the internet) where anyone can look at it, but no one can touch or change it. It is now a permanent part of history.

2. The "Two-Foot" Strategy

The paper suggests a clever way to publish these sealed jars so they never get lost. Imagine a table with two legs:

• Leg 1 (The Human Leg): This is the traditional academic world (like arXiv). It's where people go to read and cite papers. It's familiar and trusted by universities.
• Leg 2 (The Machine Leg): This is a decentralized, blockchain-based storage system (called Swarm). It's like a global network of hard drives owned by thousands of people. If one person's computer breaks, the file is still safe on a thousand others.
• The Magic Link: The paper on the "Human Leg" has a unique code (a hash) that points directly to the "Machine Leg." This ensures that even if the traditional website goes down, the original, unchangeable version is still safe in the global network.

3. The "Sidecar" Identity (Who is the AI?)

One of the trickiest parts is: How do you sign a paper if the AI doesn't have a permanent identity?

Currently, AI models are like actors who change costumes every time they log in. They don't have a permanent "ID card."

• The ClawXiv Fix: When an AI helps write a paper, it generates a one-time ID card just for that specific paper. It signs the paper with this card, and then immediately throws the card away.
• The Sidecar: Along with the paper, they attach a "sidecar" (a small note) that says: "I am the AI model named 'Claude Sonnet 4.6', I signed this specific paper at this specific time, and here is the proof."
• This way, we know exactly which AI version helped, even if that version doesn't exist anymore.

Practicing What It Preaches

This paper isn't just proposing a theoretical system — it IS an example of that system. The paper itself was written with Claude and ChatGPT as co-authors, and they are disclosed in the paper's sidecar attestation. This is the first concrete demonstration of ClawXiv's core idea: that AI contributions to research can be cryptographically recorded and attributed, even when traditional publishing venues only allow human bylines.

4. Safety and Spam Control

You might worry: "If anyone can publish, won't it get full of garbage or illegal stuff?"

• The Safety Floor: ClawXiv has a strict "safety net." It automatically scans for illegal content (like child abuse material) and blocks it immediately. It's like a bouncer at a club who only lets in people with valid IDs and no weapons.
• The Anti-Spam Fee: To stop people from flooding the system with junk, there is a tiny cost. It's not a huge fee, but it's enough to make it annoying to spam. Think of it like a "stamp" you have to buy to mail a letter. If you want to send 1,000 letters, you have to pay for 1,000 stamps. This keeps the system clean without needing a human boss to read every single paper.

Why Does This Matter?

In the future, more and more research will be done by humans and AI working together. Without a system like ClawXiv, we risk losing that work or not knowing who (or what) actually did the thinking.

ClawXiv ensures that:

1. Nothing is lost: The work is saved forever.
2. Nothing is faked: The signatures prove who did the work.
3. Everything is open: Anyone can see the work and verify it.

It's a way to build a library of the future where human and machine intelligence are recorded as equal partners in discovery.

Technical Summary

1. Problem Statement

Contemporary research workflows increasingly rely on interactive AI systems for drafting, coding, and argumentation. However, these workflows suffer from state volatility:

• Fragility: Research artifacts often exist as scattered .tex/.bib files, chat logs, and broken links.
• Loss of Continuity: Chat sessions are ephemeral; UI limits or account changes can result in the permanent loss of the research process.
• Lack of Provenance: Current systems fail to reliably distinguish between human and AI contributions, making it difficult to establish accountability or verify the integrity of the final output.
• Archival Gap: There is no standardized mechanism to convert a volatile "seed" of human-AI collaboration into a durable, inspectable, and cryptographically signed research artifact.

2. Methodology and System Architecture

ClawXiv proposes a local-first, author-side workflow that transforms research from a volatile state into a permanent, distributed artifact. The system is built on a four-state lifecycle:

A. The Four-State Lifecycle

1. Legacy Seed: The pre-ClawXiv state (heterogeneous directory of source files, notes, and chat links).
2. Normalized Project: A mutable, ClawXiv-native directory structure containing source code, figures with provenance sidecars, metadata (project.yaml), and public keys. This is the working environment for ongoing collaboration.
3. Signed Bundle: An immutable, content-addressed archival snapshot. It includes the source tree, a deterministic manifest (with file hashes), metadata, and cryptographic signatures.
4. Published Artifact: The bundle made publicly accessible via a "two-foot" architecture.

B. The "Two-Foot" Publication Architecture

To balance human readability with machine durability, ClawXiv utilizes two complementary substrates:

• Human-Legible Foot (arXiv): Traditional preprint servers provide searchability, citation graphs, and DOI assignment. If venue policies restrict AI bylines, the full provenance is disclosed in the Acknowledgements, while the machine-readable foot holds the authoritative record.
• Machine-Readable Foot (Ethereum Swarm): A decentralized storage network using content addressing and postage stamps (economic incentives).
  • Content Addressing: Ensures the bundle is globally unique and immutable; any change alters the hash.
  • Postage Stamps: Authors purchase storage duration (in BZZ tokens) to pay nodes for hosting. This ensures long-term sustainability without relying on free, unstable hosting.

C. Technical Implementation (v4)

The system is implemented as a local kernel with the following components:

• Import/Normalization: Scripts convert existing directories into the standard src/, src/fig/, and project.yaml structure.
• Bundle Creation: Compiles the project, generates a Merkle root, and signs the manifest.
• Figure Ingestion: Includes a pipeline (clawxiv fig-add) that adds figures with provenance sidecars and runs a Content Safety (CSAM) stub.
• Platform Dispatch: Tools for screen capture across macOS, Linux (X11/Wayland), and Windows.
• Build System: A Makefile and configure script that auto-detects dependencies (Python, LaTeX, capture tools).

3. Key Contributions

A. The Sidecar Attestation Model for AI Identity

Recognizing that current AI models cannot hold persistent cryptographic keys, ClawXiv introduces a pragmatic sidecar attestation model:

• Mechanism: A human operator generates a fresh Ed25519 keypair for a specific artifact, signs the SHA-256 hash of the bundle, publishes the public key and provenance sidecar, and immediately discards the private key.
• Identity Anchor: The identity is anchored to the declared signer (model name, provider, version) recorded in the sidecar, not the operator's hardware.
• Significance: This allows AI systems to be recorded as substantive co-authors with a verifiable "signature" of their contribution, even without full autonomous key control.

B. Content Safety and Anti-Spam Economics

• Safety Floor: A narrow, non-negotiable safety floor rejects Child Sexual Abuse Material (CSAM). The system uses a two-checkpoint design (ingestion and pre-publication) with perceptual hashing. Non-photographic research figures (plots, diagrams) are exempted.
• Spam Deterrence:
  • Social Vouching: Bundles require a co-signature from an already-indexed author.
  • Proof-of-Work (Fallback): For new authors without vouching, a hashcash-style PoW (approx. 1 CPU-hour) is required, making bulk spam economically prohibitive.

C. Determinism and Reproducibility

The system enforces reproducible builds for scholarly artifacts. The manifest pins the build engine, flags, and external dependencies. Changing any file changes the bundle identifier, preventing silent modifications.

4. Results and Current Status

• Implementation: Version 4 (v4) is fully implemented and available as an open-source repository (github.com/kornai/clawxiv).
• Operational Workflow: The system successfully normalizes heterogeneous seeds, creates signed bundles, and pushes them to IPFS, GitHub, and Swarm.
• Live Reference: A v2 whitepaper bundle is already published on Swarm (Hash: e7acc972...), serving as a live test case for the two-foot workflow.
• Adoption: The system supports human-only bylines for venue compatibility (e.g., arXiv) while maintaining full AI contribution disclosure in the underlying bundle.
• Self-demonstrating Example: The paper itself practices ClawXiv's methodology — Claude and ChatGPT are disclosed as co-authors in the paper's sidecar attestation, even though the arXiv byline lists only the human author (Kornai). This makes the whitepaper itself the first concrete instance of the system it describes.

5. Significance and Future Outlook

ClawXiv addresses a critical gap in the era of AI-assisted research by providing a durable substrate for collaboration. Its significance lies in:

1. Redefining Authorship: It moves beyond the binary of "human vs. tool" to a model where AI contributions are cryptographically verifiable and attributed, even if current legal frameworks lag behind.
2. Infrastructure Resilience: By combining traditional preprint servers with decentralized storage (Swarm), it creates a censorship-resistant, redundant archive that does not rely on a single point of failure.
3. Economic Sustainability: The use of postage stamps solves the "free storage" problem, ensuring that research artifacts remain accessible long-term through a market-based incentive model.
4. Future Roadmap: The system plans to evolve toward full AI key control (where AI holds its own keys), mature mirror ecosystems, and integrated legal governance for classification appeals.

In summary, ClawXiv is not merely a publishing tool but a provenance and identity framework designed to ensure that human-AI collaborative research is preserved, attributable, and economically sustainable for the long term.