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

This paper introduces ClawXiv, a local, author-side workflow and distributed architecture designed to transform volatile human-AI collaborative research processes into durable, signed, and inspectable archival artifacts through a four-state migration pipeline and a suite of operational scripts.

The Gist

Imagine you are writing a research paper, but instead of working alone, you have a team of brilliant AI assistants helping you. You chat with them, they draft sections, you tweak the code, and you generate charts. But here's the problem: chat rooms are like sandcastles. If the server crashes, the link breaks, or you lose your login, that entire conversation—and the work inside it—vanishes. Even if you save the text files, you lose the story of how the work was made.

ClawXiv is a new system designed to turn those fragile, fleeting conversations into a permanent, unbreakable digital monument. It's a workflow that takes your messy research notes and turns them into a "signed bundle" that can never be altered without anyone noticing.

Here is how it works, using simple analogies:

1. The Problem: The "Sandcastle" vs. The "Stone Tablet"

Currently, research with AI happens in "volatile chat sessions." It's like building a sandcastle on the beach; the tide (or a software update) can wash it away.

• ClawXiv's Solution: It acts as a mold and a kiln. It takes your messy "sand" (chat logs, scattered files, half-finished code) and pours it into a structured "mold" (a normalized project folder). Then, it fires it in a "kiln" to create a hard, permanent "stone tablet" (a Signed Bundle). Once fired, you can't chip a piece off without the whole tablet cracking and revealing the tampering.

2. The Four Stages of a Paper

The paper describes a journey a research project takes through four distinct "rooms":

1. The Legacy Seed: The messy starting point. A folder full of .tex files, images, and links to chat history. It's fragile and disorganized.
2. The Normalized Project: The "workshop." This is where you and your AI co-authors continue to work. It's a clean, organized folder where everything is labeled and ready.
3. The Signed Bundle: The "sealed time capsule." When you are ready to publish, the system locks the project, calculates a unique digital fingerprint for every single file, and signs it with a cryptographic key. This bundle is immutable—if you change a single comma inside it, the fingerprint changes, and the signature breaks.
4. The Published Artifact: The "public display." This is the sealed time capsule placed in public archives (like arXiv or a decentralized network) so anyone can find it and verify it's real.

3. Who is the Author? (The "Ghost in the Machine" Problem)

Usually, we think of authors as humans. But in ClawXiv, an AI can be a co-author.

• The Challenge: AI models don't have permanent identities. They are like actors who change costumes every time they log in. How do you sign a paper if the "author" doesn't have a permanent key?
• The Solution: ClawXiv uses a "Sidecar Attestation." Imagine the AI writes a draft. A human (the operator) then generates a temporary digital key just for that specific moment. The AI signs the draft with this temporary key, and the human signs the whole thing, saying, "I witnessed this AI create this, and I approve it."
• The "Layered" Signature: The system distinguishes between the AI's draft (Layer 0) and the Human's edits (Layer 1). It's like a transparent sheet: you can see exactly what the AI wrote and exactly what the human changed, all signed separately. This prevents the human from taking credit for the AI's work, or the AI from claiming human-level responsibility.

4. The "Two-Foot" Strategy

How do you publish this? ClawXiv uses a "two-foot" approach to ensure the paper doesn't get lost:

• Foot 1 (Human Legible): It sends the paper to arXiv (the standard academic archive). This is for people who want to read and cite it using normal tools.
• Foot 2 (Machine Legible): It uploads the "Signed Bundle" to Swarm (a decentralized, blockchain-based storage network). This is a backup that lives on thousands of computers worldwide. Even if arXiv deletes the paper or a government blocks the website, the "Swarm" copy remains because it's distributed like a peer-to-peer file sharing network.
• The Link: The arXiv paper contains a "fingerprint" (hash) that points directly to the Swarm copy, proving they are the same document.

5. Safety and Spam Control

The system isn't a "free-for-all." It has a "safety floor."

• The Safety Floor: It automatically checks images to ensure they aren't illegal (specifically, Child Sexual Abuse Material). If an image matches a known bad list, the system refuses to package it.
• Anti-Spam: To stop people from flooding the system with junk, it uses a "vouching" system. If you are new, you need an existing trusted author to "vouch" for you, or you have to do a tiny bit of computer work (proof-of-work) to prove you aren't a bot.
• Storage Costs: Unlike some systems that promise "free forever" storage (which often disappears when funding runs out), ClawXiv uses a "postage stamp" system. You pay a small fee to store the paper for a set time. Libraries or funders can "renew the stamp" later to keep it online forever, similar to how libraries renew book subscriptions.

6. The "Ceremony" of Truth

The paper argues that signing a document isn't just about math; it's a ritual.

• Just as a courtroom oath relies on the belief that lying has consequences, a digital signature relies on the belief that the math is unbreakable.
• ClawXiv treats the process of signing as a "ceremony" that makes the history of the paper visible. It records not just what was written, but who (human or AI) wrote it, when, and how it was edited.

Summary

ClawXiv is a toolkit that helps humans and AI work together without losing their work to the "cloud." It turns messy chat sessions into tamper-proof, signed, and permanently archived research artifacts. It ensures that if an AI writes a brilliant paragraph, that contribution is recorded, signed, and preserved forever, distinct from the human edits, creating a clear, honest history of how the science was made.

Technical Summary

Technical Summary: ClawXiv

Problem Statement
Contemporary research workflows increasingly rely on interactive AI systems for drafting, refining, and iterating on scholarly artifacts (e.g., LaTeX manuscripts, BibTeX bibliographies, code). However, these chat-centric workflows suffer from weak checkpointing. State loss occurs due to UI limits, link snapshotting failures, or account changes, forcing rework and undermining scholarly continuity. Furthermore, surviving work products often exist in heterogeneous, fragile forms (directories of .tex files, figures, and ephemeral chat links) that lack the durability, inspectability, and cryptographic integrity required for reliable archival. There is a need for a system that can migrate volatile human–AI collaborative sessions into durable, signed, and content-addressed research artifacts without promising editorial judgment of scientific quality.

Methodology and Architecture
ClawXiv proposes a workflow and archive architecture designed to convert a "legacy seed" (pre-ClawXiv working directory) into a "published artifact" through four distinct states:

1. Legacy Seed: The heterogeneous pre-ClawXiv state.
2. Normalized Project: A mutable, ClawXiv-native directory for ongoing work, containing source trees, figures with provenance sidecars, and metadata (project.yaml).
3. Signed Bundle: An immutable, content-addressed archival snapshot containing the source, manifest, hashes, and cryptographic signatures.
4. Published Artifact: The bundle made publicly reachable via distributed channels.

The system is implemented as a local, author-side kernel consisting of:

• Import/Normalization: Scripts to transform existing seeds into the canonical project structure.
• Bundle Creation: A script (bundle-create.sh) that compiles, signs, and packages the project into a content-addressed unit.
• Publication: A script (bundle-push.sh) that verifies and pushes bundles to public infrastructure (IPFS/IPNS, GitHub, and Ethereum Swarm).
• Figure Ingestion: A pipeline (clawxiv fig-add) that adds figures, generates provenance sidecars, and runs a content-safety stub (currently rejecting photographic raster images until a hash-list provider is integrated).
• Screen Capture: A platform-dispatching utility (bin/fig-capture) to capture screen regions and immediately ingest them as figures.

Key Contributions
The paper introduces several specific technical and theoretical innovations:

• Layered Signature Scheme (v5): A structural distinction between "AI-draft layers" (Layer 0, immutable) and "human editorial layers" (Layer $n \ge 1$). This allows for a transparent history where human signatures represent approval of the current state rather than sole authorship of the underlying AI draft.
• AI-to-AI Co-authorship Sessions: A mediator daemon (clawxiv-mediate) that facilitates structured exchanges between two AI systems. The daemon enforces protocol, maintains an append-only signed transcript, and produces a Layer 0 artifact for human review.
• Session Recovery via Sidecar Checksums: A mechanism where the SHA-256 hash of the AI's local "Scratch/" working space is recorded in a provenance sidecar at signing time. This allows resuming sessions to verify coherence and distinguish between clean states and stale debris.
• Provenance Ceremonies Analysis: The paper grounds its signing mechanics in J.L. Austin's speech-act theory and Frits Staal's syntactic account of ritual. It argues that a provenance ceremony requires both a pointer to a mechanism (cryptographic one-wayness) and an "interpreter pointer" (shared conventions and legible protocols) to be epistemically effective.
• Two-Foot Publication Architecture: A hybrid model utilizing:
  • Human-legible foot: Conventional infrastructure (arXiv) for search and citation, with full disclosure of AI contributions in acknowledgments.
  • Machine-readable foot: Ethereum Swarm, utilizing "postage stamps" for economically sustainable, content-addressed storage. This provides redundancy and a stable, globally unique hash independent of hosting organizations.
• Content Safety Floor: A narrow, mechanical check for Child Sexual Abuse Material (CSAM) at ingestion and publication stages, utilizing perceptual-hash lists. Non-photographic figures (vectors, small rasters) are exempted to avoid false positives in research contexts.

Results and Implementation Status
The paper serves as both a whitepaper and a user guide for the accompanying software bundle.

• Implemented Kernel (v4): The system currently supports import, normalization, local bundle creation, publication push to IPFS/GitHub, figure ingestion with a CSAM stub, and platform-dispatching screen capture.
• v5 Extensions: The paper details the design and partial implementation of layered signatures, the AI-to-AI mediator daemon, and scratch-state recovery.
• Economic Model: The system employs social vouching for spam deterrence (requiring co-signature from an indexed author) and a fallback hashcash-style proof-of-work for initial submissions. Storage sustainability is handled via Swarm postage stamps, allowing institutions to renew storage without the original author's involvement.

Significance and Claims
The paper modestly claims to provide a "durable substrate" for mixed human–AI research workflows. It does not claim to be a general social network or a venue for editorial quality control. Its primary significance lies in:

1. Reliability: Solving the problem of state loss in chat-based research by creating durable, inspectable artifacts.
2. Accountability: Establishing a cryptographic binding between artifacts and author identities (human and AI) through a sidecar attestation model, acknowledging that current AI systems lack persistent key control.
3. Legibility: Making the provenance of AI-assisted work explicit and structured, distinguishing between AI generation and human editorial approval.
4. Sustainability: Proposing an economic model for long-term archival that avoids "free" subsidies that may evaporate, using token-denominated storage commitments.

The authors emphasize that the system is designed to strengthen scholarly norms (citation, accuracy, correction) by making artifact boundaries and release events more explicit, rather than replacing them. The paper acknowledges that full AI authorship requires future developments in persistent key control and cross-session continuity, which are currently addressed pragmatically through operator-held keys and sidecar attestations.