Graded Continuity and Operator Reconstruction Burden in Long-Horizon Human–AI Collaboration: An Exploratory Case Series

Long-horizon human–AI collaboration depends on more than whether isolated facts are available. This v1.0 release presents an embedded exploratory case series drawn from the creator's own longitudinal collaboration. Amber Nicholson is both the author/researcher and the human participant/operator in Human–AI Pair 01. Within the public case evidence, Operator 01 is a role-level identifier for Nicholson; it does not represent an anonymous or undisclosed participant. The corpus comprises seven Continuity-Recovery Episodes (CREs), five Positive Continuity Cases (PCs), and three explicitly linked PC/CRE comparisons. Each case was evaluated against a bounded state target and coded for continuity outcome, recovery fidelity, recovery condition, reconstruction turns, and qualitative operator burden.

Across the frozen corpus, all five PCs achieved substantive first-response recovery with zero reconstruction turns and Low burden. The seven CREs required one to four material operator repair messages before substantive or partial substantive recovery; three were rated High burden, three Moderate, and one Low. The cases show that broad topical or project-state continuity can coexist with loss of an exact routine, rationale, prior correction, role division, or current priority. Confident false reconstruction can add verification work beyond simple omission.

The release supports treating continuity as a graded, target-specific property of collaboration rather than a binary memory state. It does not estimate prevalence, compare models, validate a calibrated benchmark, or identify hidden system mechanisms. Reconstruction turns describe the observed pair-level repair process and are confounded with operator skill; they are not a system-only performance metric.

The public package contains the structured 12-case dataset, cleaned human-readable case records, README, codebook and data dictionary, public cleaning and anonymization protocol, research note, license, and package manifest. Private raw-evidence records, internal trackers, excluded high-sensitivity material, and private evidence links are not included.

Project website and continuing research updates: https://amberannnicholson.com/research