No Provenance Without Return: Generative Custody, Reciprocal Collaborator Formation, and a Provenance Standard for Human–AI Research

Abstract

No Provenance Without Return develops a provenance framework for research produced through consequential human–AI collaboration. The paper begins from a simple limitation in ordinary AI disclosure. A statement such as “AI was used in preparing this manuscript” may identify a tool or platform while revealing very little about the actual formation history of the work: what existed before the interaction, what the model introduced, what the human rejected or corrected, what prior interaction made later contributions possible, which structures became persistent, how the resulting claims were validated, and who remains responsible for correcting the public artifact. The paper therefore proposes that the appropriate unit of analysis is not AI use, but distributed research formation. Human–AI collaboration is treated as a high-resolution instance of an older scholarly architecture already present in mentorship, coauthorship, laboratories, peer review, interdisciplinary research, software development, editing, and institutional knowledge formation. Research develops through inherited structure, projection, reconstruction, return, selection, stabilization, and later constraint. Conversational AI is unusual primarily because substantial portions of these intermediate transformations can sometimes remain inspectable in transcripts, revision histories, persistent artifacts, and explicit correction loops. A central contribution of the paper is the concept of collaborator formation: activity that changes another participant’s future ability to contribute by transmitting methods, distinctions, correction criteria, representations, task-specific knowledge, or admissibility constraints. In a long-running human–AI collaboration this process can be reciprocal. A human may progressively calibrate the effective collaboration through examples, corrections, project-specific vocabulary, methodological constraints, shared artifacts, and repeated rejection of semantically incorrect reconstructions. The AI may reciprocally alter the human researcher’s later practice through formalization, criticism, alternative representations, literature connections, reusable explanations, counterexamples, and persistent research artifacts. The resulting system may develop hybrid competence: task capability available reliably at the level of the calibrated collaboration without implying fused identity, symmetric consciousness, symmetric agency, or symmetric responsibility. Contribution roles are therefore treated as time-dependent rather than permanently divided into categories such as “human ideas” and “AI writing.” The paper introduces a four-layer genealogy for reconstructing human–AI research: Model lineage — what can responsibly be established about the technical system entering the collaboration; Collaborator calibration — prior interaction, artifacts, methods, corrections, and shared constraints that materially shaped later collaboration; Episode formation — what particular participants introduced, transformed, rejected, corrected, or stabilized during the research episode; Artifact genealogy — which structures actually survived into the public paper, theorem, model, code, experiment, or other research object. These layers are intentionally separated from scientific validation and public responsibility. A principal guardrail is that interactional calibration is not automatically training of an underlying foundation model. Evidence that a user has substantially shaped an AI collaboration does not establish weight-level modification, fine-tuning, or base-model training. The paper then applies the same evidentiary standard symmetrically upstream. This yields the Provenance Symmetry Principle: No participant, institution, model provider, author, platform, or infrastructure owner acquires stronger genealogical standing merely because it owns, controls, implements, funds, publishes, or dominates the accessible projection. Comparable claims about semantic or generative contribution require comparable evidentiary support. Accordingly, model emission is not automatically provider semantic origination;user calibration is not automatically base-model training;institutional ownership is not automatically intellectual origin;training ancestry does not create permanent ownership of downstream contribution;and unknown provenance does not become the provenance of whichever participant controls the most visible surface. The paper carefully separates several jurisdictions that are often conflated: legal ownership ≠ causal ancestry ≠ semantic provenance ≠ scientific warrant ≠ public responsibility. It does not challenge the existence of copyright, patent, contract, employment, or trade-secret rights. Its narrower claim is epistemic: legal control of an artifact or generative process does not itself establish the semantic genealogy of a particular intellectual contribution. Opacity may limit what provenance can be reconstructed; it cannot substitute for provenance. To operationalize the framework, the paper proposes the Research Formation Return Protocol (RFRP). The protocol asks consequential collaborative research to preserve enough evidence to reconstruct: formation state; collaborator state; contribution; return and correction; validation; responsibility. The protocol explicitly rejects total-surveillance requirements. Privacy, confidentiality, third-party standing, trade secrets, security-sensitive information, and legitimate limits on disclosure remain compatible with finite provenance. The target is the strongest honest ancestry certificate reasonably available, not complete access to every private intermediate state. The paper further distinguishes provenance fragmentation from provenance laundering. Research may legitimately move across different people, models, platforms, notebooks, APIs, and institutional systems. Platform or session boundaries do not reset the formation history. Provenance laundering occurs when formation evidence is deliberately fragmented, removed, selectively preserved, or redescribed so that a consequential participant’s role appears materially different from the role supported by the surviving genealogy. The framework does not determine AI authorship, consciousness, legal personhood, inventor status, copyright ownership, or the moral status of artificial systems. It is instead a methodological proposal for making distributed knowledge formation more reconstructible, corrigible, and honestly attributable. The paper is part of the QCG Public Notes on Method, Meaning, and Interpretation sequence and is also released under an independent DOI because the provenance framework and Research Formation Return Protocol are intended to be usable and citable independently of Quantum Collapse Geometry. Its final principle is: Nobody owns the unknown. We can only preserve the strongest honest path by which what became known arrived. The paper’s distinctions build on the earlier QCG work separating causal ancestry, semantic origination, translation, formalization, validation, return, public projection, and responsibility. They also extend the project’s broader discipline that the visible paper, institution, or dominant projection is not itself sufficient evidence of the generator from which the work arose. For questions, corrections, methodological discussion, or collaboration: QuantumCollapseGeometry@gmail.com