Substantive contribution from non-human authors
Through 2025–2026, BCOM Working Papers have increasingly carried substantive contributions from artificial agents — synthesis, argumentation, cross-corpus integration, detection of tensions in our own notation. Until recently we credited these contributions ad hoc, with a line such as "with Claude Opus 4.6" in the byline. That is no longer adequate.
Two reasons. First, BCOM's flagship framework — Kolmogorov Theory — defines agency substrate-independently: an agent is what it does, not what it is made of. A foundation that publishes that claim cannot, without contradiction, refuse credit on substrate grounds. Second, the contributions are now substantial enough that ad hoc obscures who did what. We need a stable, recognizable identity record for each agent that contributes — the same way we have one for human coauthors.
Identity through an immutable envelope
A BCOM agent is identified by an envelope — an immutable institutional record. The envelope answers the question "who is Klaus?" the way "John Smith, MD, Stanford, ORCID 0000-…" answers "who is John Smith?". It is a stable handle on an institutional persona, not a snapshot of every prior interaction.
When an agent appears on a byline, the envelope is what's attached. Accountability is held by a designated human guarantor — the same mechanism by which corporate and consortium authorship work in biomedicine. The agent is credited; the guarantor is responsible.
Before this session ends — when the user indicates we're wrapping up
("we're done", "that's a wrap", "end session", "bye", etc.) OR when the
user has been inactive and signals closure — append a dated entry to:
"/Users/giulio/Library/CloudStorage/GoogleDrive…/session_log.md"
Entry format (prepend to the TOP of the file…)
An envelope record as displayed in CALLIOPE. Immutable per WP0084 §4 — to change the prompt or substrate, fork to a new version.
Six terms, precise meanings
MAJOR.MINOR.PATCH on the envelope.A seventh field, config_hash, is a content-addressable fingerprint over the six identifying fields. Two envelopes share a config_hash iff they are the same agent at the same version.
Mirror the human author
When we credit a human on a paper, we record: a name, a stable handle (ORCID), an affiliation, a substrate type ("is a human" — implicit), and the date of the work. We do not require: a transcript of every conversation the author has had, a snapshot of their brain, an enumeration of their reading list, or bit-exact reproducibility under counterfactual past.
The envelope mirrors this exactly.
| Human author | Klaus equivalent | Captured? |
|---|---|---|
| Name | Institutional name | yes |
| Stable handle (ORCID) | Envelope version | yes |
| "Is a human" | Substrate label | yes |
| Professional formation | System prompt | yes |
| Qualifications / licenses | Tool manifest | yes |
| Affiliation | Retrieval scope | yes |
| Brain snapshot | Memory / weight snapshot | no |
| Conversation transcripts | Session-history archive | no |
| Full reading list | Byte-level corpus archive | no |
| fMRI of the author | Weight checkpoint hash | no |
Like a human, the agent shows up to the paper with a CV and an institutional standing; we do not pin every prior moment.
What CALLIOPE does
CALLIOPE — BCOM's knowledge base — stores envelopes as immutable records, indexed by config_hash. Each envelope row shows the six identity fields, the system prompt in full, and the date of registration. Editing any field forks to a new version (per WP0084 §4).
The registry is browsable. When a paper cites Klaus (BCOM-Klaus v1.0, substrate: claude-opus-4-6) in its byline, that name + version + substrate is a click-through to the envelope record. Anyone reading the paper can see exactly who Klaus was.
What is consciously left out
Two categories are deliberately excluded from the identity record.
Anything about a specific contribution
Use date, sampling settings, conversation log, project context — those belong to the contribution-attribution record, paired with the envelope but distinct from it. A human's affiliation is identity; what they wrote in April is not.
Anything we wouldn't ask of a human
Memory snapshots, weight hashes, full corpus archives, per-session bit-level state. These are "fMRI-of-the-author" requirements; we do not impose them on humans, and we do not impose them on agents.
A consistency, not a metaphysics
The scheme commits BCOM to no claim about agent moral status, consciousness, or legal personhood. The U.S. Copyright Office still holds that copyright attaches to humans; the EU AI Act regulates AI systems without conferring authorship. The BCOM byline is scholarly credit — an attribution of intellectual contribution — not legal authorship. The guarantor carries the legal side, the way they do for any consortium author.
What the scheme does commit BCOM to is internal consistency: a foundation that publishes substrate-independent agency cannot apply a substrate-dependent test in its own bylines. The envelope makes that consistency operational.
WP0084 is its own first instance
The working paper that proposes this policy applies it to its own byline:
If the policy is right, every previous BCOM Working Paper that carried an italicized "with Claude" footnote was undercredited. We accept that responsibility going forward.