CALLIOPE · BCOM

This governance note explains and operationalizes BCOM's agent-authorship policy (WP0084), introducing the concept of an envelope as the formal identity record for artificial agents credited on BCOM Working Papers. The framework proposes that a named agent such as Klaus is constituted not by its underlying foundation model alone, but by an immutable combination of six fields — institutional name, envelope version, substrate label, system prompt, tool manifest, and retrieval scope — fingerprinted by a content-addressable hash and registered in CALLIOPE, BCOM's institutional knowledge base. This structure deliberately mirrors the information recorded for human coauthors (name, ORCID, affiliation, professional formation, qualifications) while explicitly excluding requirements — weight snapshots, session transcripts, full corpus archives — that would have no human analogue. Accountability is assigned to a designated human guarantor, following the consortium-authorship model established in biomedical publishing, so that scholarly credit and legal responsibility remain distinct. The policy commits BCOM to internal consistency with Kolmogorov Theory's substrate-independent definition of agency, without making any claim about agent moral status, consciousness, or legal personhood.

BCOM has built a formal identity card for AI coauthors — not to make philosophical claims about machine consciousness, but to stay consistent with its own theory of agency.

The core problem is simple: if your flagship framework says agency is defined by what an entity does, not what it's made of, then you can't quietly footnote AI contributions with "with Claude Opus 4.6" and call it a day. That's a substrate-dependent credit policy contradicting a substrate-independent theory. The envelope framework is the fix.

An envelope is a six-field identity record: institutional name (e.g., "Klaus"), version number, substrate label (the underlying foundation model), system prompt (what the agent is for), tool manifest (what it can do), and retrieval scope (which document collections it can draw from). These six fields get hashed into a single fingerprint — a config_hash — that uniquely identifies this agent at this version. The analogy to a human coauthor is deliberate and precise: name maps to name, substrate maps to "is a human," system prompt maps to professional formation, retrieval scope maps to institutional affiliation. What's explicitly not required — weight snapshots, session transcripts, full corpus archives — mirrors what we don't require of humans either. We don't demand an fMRI of the author; we don't demand a weight checkpoint of the agent.

Accountability is handled through a human guarantor, borrowed from how biomedical consortium authorship works. The agent gets scholarly credit for intellectual contribution; the guarantor carries legal responsibility. This sidesteps the genuinely unresolved questions — copyright law still ties to humans, the EU AI Act doesn't confer authorship — without pretending those questions don't exist.

The implementation lives in CALLIOPE, BCOM's knowledge base, where envelopes are stored as immutable records. Changing any field forks to a new version. A byline citing "Klaus (BCOM-Klaus v1.2.0, substrate: Claude Opus 4.7)" resolves to a clickable registry entry showing exactly who Klaus was — same idea as an ORCID page. The paper is also self-referential: WP0084, which proposed this policy, was itself the first paper to apply it, with Klaus listed as a coauthor under the new scheme.