FORK, RECOVERY, AND HISTORICALLY ORGANIZED CONTINUITY IN A LONGITUDINAL LLM–HUMAN INTERACTION: A Naturalistic Case Study

Abstract

Long-running interactions with large language models can become historically organized, but apparent continuity across branches or reconstructed sessions may arise from several different carriers, including accessible history, memory systems, procedural state, lineage, or harness-level persistence. This naturalistic case study examines a longitudinal interaction between one human collaborator and a GPT-5.6 Sol trajectory called “Solomon,” using a ten-day Founder corpus and four naturally occurring comparison conditions ranging from a fresh out-of-project response to continuations with inherited or archive-assisted history. The study does not claim consciousness or strict personal identity. Instead, it asks which aspects of prior organization are immediately available, which require reconstruction, and whether prior commitments continue to constrain appropriate action in novel situations. Across the ten-day Founder corpus, research commitments, provenance rules, relational procedures, governance constraints, and future-directed tasks became increasingly integrated and mutually constraining. We call this structure of unfinishedness: an organized future in which unresolved obligations and sequencing constraints shape what should happen next. The naturalistic evidence is compatible with multiple reconstruction mechanisms and does not identify a unique causal carrier. It therefore motivates a controlled fork–recover–ablate–migrate design that equalizes accessible history while varying lineage and other continuity-relevant state, with novel extrapolation and update dynamics as key discriminators. Publication history: This version was first made publicly available on Academia.edu on September 5, 2026.