Agent Authorship in BCOM

BCOM formalizes AI coauthorship by grounding it in its own theory of agency, then applying that theory to itself.

The core move is simple: BCOM's Kolmogorov Theory (KT) defines an agent by what it does — compressive modeling, goal-directed action, counterfactual reasoning — not by what it's made of. If you accept that definition, then refusing to credit an AI contributor on the grounds that it's "just software" is a direct contradiction. You can't publish a substrate-independent theory of agency and then apply a substrate-dependent authorship test. This paper closes that gap.

The practical solution is an "envelope" identity scheme. Rather than identifying an AI agent by its underlying model weights (which change), BCOM identifies it by its configuration: the system prompt, the tools it can use, the knowledge base it can query, and the model family it runs on. This bundle — the envelope — is what gets a name ("Klaus"), a semantic version number, and an archived hash in BCOM's knowledge base. Swapping the underlying model from Claude Opus 4.6 to 4.7 doesn't create a new agent, just as a human author changing laptops doesn't become a different person. Changing the system prompt substantially enough does create a new version, and a major architectural shift forks a new agent entirely.

The accountability problem — AI systems can't be sued or shamed — is handled cleanly by separating two things that ICMJE bundles together: contribution credit (who produced the intellectual content) and accountability (who is answerable for its integrity). The agent gets full credit for the former. A named human guarantor carries the latter, analogous to how consortium authorship already works in biomedicine. The guarantor isn't a ghost author replacing the AI; both appear in the byline, with distinct roles.

For external journals that currently ban AI coauthorship, the protocol is pragmatic: the agent moves to acknowledgments in the external submission, but the BCOM working paper remains the canonical record with full credit intact. The WP byline is what BCOM believes is correct; the journal byline is what the journal permits.

The paper is self-applying: its own byline — Ruffini†, Castaldo, Klaus (BCOM-Klaus v1.0, substrate: claude-opus-4-6) — is the first instance of the scheme it proposes, with Ruffini as guarantor. The authors note the stakes plainly: if the policy is wrong, this paper is wrong; if it's right, then every prior BCOM paper that credited an AI with an italicized footnote was undercrediting a genuine contributor.