PaperlayerView exact revision on arXiv

Source-linked AI summary

Matched Starts, Divergent Objects: How Human-AI Collaboration Forms What It Explains

Mehmed Zahid Çögenli

arXiv:2609.04542v1cs.HC

TL;DR

The paper asks how human–AI scholarly trajectories diverge under matched starts and how their histories shape research objects. It compares isolated trajectories under shared starting conditions and finds that their process genealogies yield divergent scholarly products.

  • Problem

    The paper asks how trajectories diverge under matched starting conditions and how process histories shape the research objects and scholarly contributions they produce.

    Source: §1 · p. 2, Research quest4on · p. 3

  • Method

    The same expert researcher developed branch-isolated trajectories from a shared corpus, research problem, prompt, publication objective, and conduct rules, preserving the sequence of analytical decisions.

    Source: §3.4 · p. 10, §5.2 · p. 34

  • Results

    The trajectories formed different research objects and scholarly products through process genealogies, while scholarly change, relational continuity, and local re-formation remained distinct structures.

    Source: §4.4 · p. 27, §5.2 · p. 34

  • Takeaways & Limitations

    Process history provides an explanatory layer between matched starts and divergent scholarly outcomes.

    Source: §5.2 · p. 34

    Because trajectories were naturalistic and follow-up interactions differed, the study does not isolate a single causal effect attributable to an AI system.

    Source: §6 · p. 37

Abstract

from arXiv · show

Scholarly knowledge is typically encountered in stabilized form, while the process histories through which research objects, claims, and contributions acquire form remain largely hidden. This study examines how human-AI scholarly collaboration develops under matched starting conditions and whether those conditions stabilize the inquiry itself. Using a longitudinal corpus of 843 turns, the same expert researcher developed branch-isolated scholarly trajectories with different generative AI systems from the same corpus, frozen research problem, starting prompt, publication objective, and conduct rules. Two eligible trajectories were reconstructed ex post through scholarly trajectory analysis, source-faithful interaction reconstruction, a Socioduality relational-process overlay, and downstream propagation analysis. Both trajectories independently shifted the initial continuity problem from recall toward usability, but subsequently formed different research objects. One trajectory culminated in an endpoint manuscript on continuity labour, the distributed work required to sustain usable collaboration; the other in an endpoint manuscript on distributed, evolving, and unevenly usable project state. Their analytic genealogies involved failed analytical units, rejected explanations, changes in scale, counterexamples, and conceptual stabilization, and propagated into different research questions, findings, methods, evidence logics, and scholarly contributions. Relational analysis further showed that consequential scholarly change, reciprocal continuity, and local substantive re-formation were distinct process structures, and that continuation did not necessarily constitute epistemic endorsement. The findings demonstrate empirically constrained research-object formation within human-AI scholarly collaboration and show how process histories shape the scholarly objects and products that emerge.

1. Introduct,on

The study examines how two human–AI scholarly trajectories developed from matched starts but independent follow-up interactions and analytic paths, reconstructing their consequential events and relational processes and tracing how their histories shaped distinct research objects and frozen endpoint manuscripts.

2.1. Scholarly knowledge product4on as a process; 2.2. H4stor4cal cond4t4on4ng and paths-4n-the-mak4ng 4n scholarly development

  • Scholarly endpoints conceal histories in which earlier choices, failed units, rejected explanations, and counterexamples condition later possibilities, while paths remain revisable rather than predetermined.Process reconstruction traces how consequential events opened or removed analytical possibilities and remained visible downstream.Source: §1 · p. 2, §2.2 · p. 5

1. Introduct,on

  • Under matched starts, Claude formed continuity labour while ChatGPT formed project-state continuity; their objects propagated into distinct scholarly products, and continuation did not necessarily signal epistemic endorsement.The relational-process analysis distinguished consequential scholarly change, reciprocal relational continuity, and local substantive re-formation.Source: §1 · p. 3

Research quest4on

  • The study asks how matched human–AI trajectories diverge and how their histories shape research objects and contributions, linking analytic genealogy to scholarly products and deriving requirements for cooperative systems.Proposed requirements include explicit endorsement states, recoverable project-state provenance, and source-faithful preservation of contribution boundaries.Source: Research quest4on · pp. 3–4

2.3. From analyt4cal trajectory to research-object format4on

  • Analytical development can reorganize shared material around different relations and problems, forming distinct research objects without reducing formation to free construction or an object dictated by data.The study treats empirically constrained research-object formation as a question to establish through the trajectory record.Source: §2.3 · pp. 5–6

2.4. Human–AI scholarly collaborat4on as cooperat4ve knowledge work

  • Human–AI collaboration is treated as evolving cooperative work, not a symmetric partnership; AI contributions participate in trajectory formation while interaction records enable unusually detailed reconstruction.The study analyzes evolving human–AI trajectories rather than AI systems as isolated performers.Source: §1 · p. 2, §2.4 · p. 6

3.2. Project P: the natural long4tud4nal corpus

  • 843 turns across five sessions formed the naturally occurring Project P corpus, a longitudinal record of a real mobile-application project shared across eligible branches.The corpus captured activities including problem clarification, design decisions, continuity restoration, revision, and use of externalized project materials.Source: §3.2 · p. 9

3.1. Research arch4tecture; 3.3. P0 common-problem select4on

  • P0 generated a shared continuity problem from independent AI proposals, and comparison was deferred until eligible, isolated scholarly trajectories had reached completed manuscripts.The common problem was selected through semantic convergence rather than retrospective preference for a later branch outcome.Source: §3.1 · p. 7, §3.3 · pp. 9–10

3.4. P1 matched cond4t4ons and natural scholarly trajector4es

  • P1 matched the corpus, frozen problem, initial prompt, expert researcher, and conduct rules; the AI system differed, while theory, analysis, and manuscript development evolved naturally.The prompt aimed at a complete publishable academic article without pre-specifying its theoretical framework, research questions, or methods.Source: §3.1 · p. 7, §3.4 · p. 10

3.4. P1 matched cond4t4ons and natural scholarly trajector4es; 3.11. Researcher pos4t4on and controlled common actor

  • The design matched the corpus, problem, researcher, starting prompt, publication objective, and conduct rules, but allowed branch-specific utterances and analytic development to emerge.The common researcher applied stable epistemic criteria rather than identical prompts; branch isolation barred deliberate transfer of branch-specific outputs, though carryover of general expertise was not eliminated.Source: §3.4 · p. 11, §3.11 · p. 18

3.5. Gem4n4 4ntegr4ty control and exclus4on from formal compar4son

  • The Gemini branch was excluded after remediation failed to establish auditable full-corpus input equivalence, enforcing matched-input requirements rather than judging its emerging findings.The exclusion occurred before formal comparison and was not based on the substantive quality or direction of Gemini's findings.Source: §3.5 · p. 12

3.6. Comparat4ve un4t and endpo4nt scholarly art4facts; 3.12. Researcher dual role, art4fact status, and analyt4cal transparency

  • The comparison unit was each completed scholarly trajectory, not an isolated AI system; the Claude and ChatGPT manuscripts were treated as endpoint artifacts.With one trajectory per eligible system, system-level effects could not be separated from trajectory-level effects; endpoint manuscripts were frozen before P2 and used as records of where each trajectory stabilized.Source: §3.6 · p. 13, §3.12 · p. 18

3.7. P2-A: scholarly trajectory analys4s

  • Claude's ledger contained 16 major process events and ChatGPT's 18, reconstructed independently to trace consequential redirections, rejected alternatives, and downstream effects.Examples included failed segmentation, rejected explanations, counterexamples, and methodological reversals; the analysis did not impose equivalent event schemes across branches.Source: §3.7 · p. 14

3.8. Source f4del4ty and move-reconstruct4on aud4t

  • 305 apparent nodes, including 91 same-party AI-to-AI adjacencies, became 214 source-faithful moves (107 human, 107 AI), removing artificial pathway breaks.The 213 observable adjacent cross-party boundaries remained unchanged; validating the source representation changed the inferred topology, not those boundaries.Source: §3.8 · p. 15

3.9. P2-B: Soc4odual4ty relat4onal-process overlay

  • Relational continuation and substantive re-formation were coded separately: a move could remain connected to the preceding contribution without changing the participant's substantive position.Only confirmed episodes received secondary coding for substantive re-formation; pathway length was treated as relational continuity, not collaboration quality or performance.Source: §3.9 · p. 16

3.10. Downstream propagat4on analys4s

  • Propagation analysis traced consequential developments across the research object, question, findings, methods, evidence logic, and contribution, requiring sequential evidence beyond repeated terminology.Developments were classified as retained, transformed, or abandoned; rejected routes remained part of the explanatory history when they redirected later inquiry.Source: §3.10 · p. 17

4. Results

Under matched starting conditions, the two trajectories converged on usability before developing distinct research objects and scholarly contributions. Their analytical histories shaped what each manuscript asked, found, and treated as evidence, while scholarly change and reciprocal relational continuity had distinct boundaries and continuation did not necessarily indicate epistemic endorsement.

4.1. Matched starts, d4vergent analyt4c trajector4es

  • Although the branches shared the researcher, corpus, problem, prompt, objective, and conduct rules, their analytical organization diverged; the same standards prompted trajectory-specific corrections.Claude’s analyses increasingly fragmented continuity through determinate boundaries, while ChatGPT’s micro-analytic depth risked overdeveloping selected mechanisms; these are observed trajectory patterns, not general claims about the AI systems.Source: §4.1 · pp. 18–19

4.2. Convergence before d4vergence: from recall to usab4l4ty

  • Both trajectories independently shifted the continuity problem from recall toward usability, then diverged between the work sustaining usability and the changing state whose usability mattered.A long-gap negative case weakened a simple elapsed-time account; continuity practices addressed usability risk through reactive restoration and anticipatory preparation.Source: §4.2 · pp. 20–21

4.3. Analyt4c genealogy and path-dependent research-object format4on

  • Failed analytical decompositions redirected Claude toward continuity labour, while a whole-case reset redirected ChatGPT toward project state.In Claude, 53% of identified locking acts fell outside candidate episode boundaries, and topical analysis found extensive domain coupling. ChatGPT distinguished previously shared from never-observed state and available from usable state; a roughly 50-day counterexample weakened elapsed time as a sufficient explanation.Source: §4.3 · pp. 21–23
  • From the same initial problem, the trajectories formed distinct research objects: Claude organized continuity around sustaining usability, while ChatGPT organized it around evolving project state.Their manuscripts consequently asked different questions about the work that makes continuity possible and about continuity when project state is distributed, evolving, and unevenly available.Source: §4.3 · pp. 23–24

4.4. Scholarly change, relat4onal cont4nu4ty, and local re-format4on were d4st4nct process structures

  • Scholarly redirections and reciprocal relational continuity had different boundaries: major changes occurred within long pathways, while formal breaks did not mark major epistemic turns.In Claude, seven of eight core critical events occurred within one pathway; in ChatGPT, its principal scientific transformation also unfolded within a single long pathway.Source: §4.4 · pp. 25–26
  • Reciprocal continuation could remain cooperative and analytically productive while the researcher retained the same substantive judgment, so continuation alone did not show AI endorsement.RF described local substantive change within Confirmed reciprocal episodes, not collaboration quality, oversight strength, or contribution importance; its differing branch rates were not performance rankings.Source: §4.4 · pp. 26–27
  • Consequential scholarly change, reciprocal relational continuity, and local substantive re-formation were distinct process structures with different boundaries.An idea could emerge and stabilize during uninterrupted interaction, a relational break could leave the scientific trajectory unchanged, and reciprocal interaction could continue without substantive re-formation or endorsement.Source: §4.4 · pp. 26–27

4.5. Downstream propagat4on 4nto the completed scholarly products

  • Different analytical histories propagated into different research objects, questions, findings, methods, evidence logics, and scholarly contributions.Developments were retained, transformed, or abandoned; even rejected routes shaped the endpoint manuscripts by redirecting analysis and specifying surviving claims.Source: §4.5 · p. 28, From trajectory d4fference to… · p. 31
  • In Claude, failed temporal and topical boundaries, rejected developmental explanations, and literature recalibration became part of the final account of continuity labour.The final manuscript reported an occasioned rather than developmental repertoire and the failure of a proposed decline in repair after controlling for session length.Source: Propagat4on 4n the Claude trajectory · pp. 29–30
  • In ChatGPT, whole-case reconstruction reorganized the analysis around changing project state and yielded findings that elapsed time alone was neither necessary nor sufficient for continuity failure.The resulting account distinguished availability from usability and representational staleness from situational misapplication, and showed continuity could support cumulative change without each development superseding prior state.Source: Propagat4on 4n the ChatGPT trajectory · pp. 30–31

4.6. Substant4ve scholarly contr4but4ons generated w4th4n the trajector4es

  • One trajectory explained continuity as labour required to keep collaborative state usable; the other explained it through distributed, evolving, and unevenly usable project state.The Claude account specified seven occasioned, interactionally co-produced practices shaped partly by platform-resource constraints. The ChatGPT account distinguished previously shared but difficult-to-use state from consequential state produced outside AI participation.Source: §4.6 · pp. 31–32

5. D,scuss,on

Under matched scholarly starts, the trajectories formed different, empirically constrained research objects. Their process histories explain how those objects developed and show that scholarly change, relational continuity, and epistemic endorsement are distinct.

5.1. Research-object format4on under matched scholarly starts

  • Matched starts formed distinct, empirically constrained objects: continuity labour as temporal articulation and project state as continuity across interactions, artifacts, and activity.The phrase continuity labour appeared earlier, but its scope and evidence logic were formed through the trajectory rather than pre-existing as a finished object.Source: §5.1 · p. 33

5.2. From result var4at4on to process explanat4on

  • The study explains a different layer than variation in results: sequential genealogies show how earlier decisions condition later analysis and help form different scholarly products.Segmentation failure, whole-case resetting, and rejected explanations changed which later distinctions or alternatives became relevant.Source: §5.2 · p. 34

5.3. H4stor4cal cond4t4on4ng and paths-4n-the-mak4ng

  • Genealogies conditioned but did not determine trajectories: human challenges and resets mattered, and consequential events propagated into later questions, methods, evidence logics, and contributions.The analysis supports path dependence through traceable downstream consequences, not endpoint differences alone.Source: §5.3 · p. 34

5.4. Scholarly change has mult4ple process structures

  • Scholarly change, reciprocal continuity, and local re-formation were distinct; about 98% of candidate episodes were reciprocal, while most consequential developments occurred within uninterrupted pathways.Relational breaks did not necessarily mark scholarly turning points, and connected returns could leave substantive positions unchanged.Source: §5.4 · p. 35

5.5. Cont4nuat4on 4s not ep4stem4c endorsement

  • Continued interaction can sustain inquiry without substantive uptake, so cooperation, continuation, uptake, and endorsement must be inferred separately.In the ChatGPT trajectory, many reciprocal human returns did not change the researcher's substantive position.Source: §5.5 · p. 35
  • Conversational systems should not infer endorsement from generic continuation cues; lightweight status markers could distinguish accepted, provisional, rejected, or still-tested proposals.The design implication is to make shared understandings inspectable rather than assume coordination establishes them.Source: §5.5 · p. 35

5.6. Process h4story and endpo4nt evaluat4on

  • Endpoints show what a trajectory stabilized, but process histories reveal how it became possible and recover alternatives and changes that finished manuscripts compress.An endpoint-only comparison misses the shared recall-to-usability convergence and the distinct genealogies that followed.Source: §5.6 · p. 36
  • Long-term collaboration systems should preserve provenance, rationale, status changes, unresolved alternatives, and recoverable routes while making contribution boundaries explicit and recoverable.The study also found that technical segmentation could imply a different interaction topology than source-faithful reconstruction.Source: §5.6 · pp. 36–37

6. L,m,tat,ons and Transparency

The comparison is bounded by its corpus, researcher, eligible trajectories, and naturalistic design; retrospective reconstruction and the relational overlay also have specific limits. Preserved artifacts and audit procedures support transparency but do not remove these boundaries.

6. L,m,tat,ons and Transparency

  • The study examines one natural longitudinal corpus and one research problem, so it does not establish how frequently these forms of divergence occur elsewhere.Broader replication would be needed across corpora, disciplines, problems, and forms of scholarly work.Source: §6 · p. 37
  • Using the same expert researcher holds expertise and conduct constant but bounds the comparison to within-researcher divergence; psychological carryover could not be eliminated.Branch isolation prohibited deliberate transfer of branch-specific outputs, concepts, methods, sources, solutions, and decisions.Source: §6 · p. 37
  • Only two P1 trajectories qualified for comparison; Gemini was excluded under these conditions after remediation failed to establish reliable input equivalence and auditable processing.That branch-specific exclusion is not evidence about Gemini's general scholarly capability.Source: §6 · p. 37
  • Because trajectories were naturalistic rather than randomized or synchronized, the study does not isolate a causal effect attributable to an AI system.Its comparison concerns evolving trajectories under matched starting conditions.Source: §6 · p. 37
  • The records reflect model and platform configurations that can change, so the findings are not timeless properties of ChatGPT or Claude.System-specific generalization requires separate evidence.Source: §6 · p. 37
  • Ex post reconstruction risks retrospective coherence, and the Socioduality overlay's usefulness in this corpus does not validate it across other populations, tasks, systems, or settings.Frozen endpoints, independent event criteria, retained abandoned routes, full-record review, propagation tests, and separate protocols made interpretation auditable rather than removing researcher participation.Source: §6 · pp. 37–38
  • Principal artifacts are preserved in supplementary appendices, but formal comparison depends on complete longitudinal records and audited analytical ledgers, not endpoint manuscripts alone.The appendices contain the branch manuscripts and frozen prompts with candidate problems and semantic clustering.Source: §6 · p. 38

7. Conclus,on

Matched scholarly starting conditions did not stabilize inquiry: two trajectories formed distinct, empirically constrained research objects, and their process histories propagated into different scholarly products. Relational analysis also distinguishes scholarly change from continuity and endorsement.

7. Conclus,on

  • The same corpus, problem, researcher, prompt, publication objective, and conduct rules supported trajectories that organized continuity around different objects.One developed continuity labour; the other developed distributed, evolving, and unevenly usable project state.Source: §7 · p. 38
  • The two objects were distinct, empirically supported contributions, and their process histories propagated into research questions, findings, methods, evidence logics, and final contributions.The data constrained the objects without dictating a single object in advance.Source: §7 · p. 38
  • Consequential scholarly change, reciprocal relational continuity, and local substantive re-formation were distinct; major redirections could occur within reciprocal pathways, and continuation need not signal endorsement.Relational breaks likewise need not be scholarly turning points.Source: §7 · p. 38
  • Human–AI collaboration both participated in producing the trajectories and made their development unusually traceable; process history helps explain how stabilized scholarly products became possible.The endpoint shows what was produced, while the process reveals how its objects, claims, and evidentiary structures took form.Source: §7 · p. 38

Does Not Remember

The study reframes continuity in long-term conversational AI work from recall of past content to the work of keeping collaborative state usable. In a longitudinal project corpus, attempts to bound this labour by time or topic failed.

Abstract

  • Continuity is treated as project work that arises inside activity with another purpose, rather than as an infrastructure built specifically to preserve memory.The motivating technical limitation is that what was established in one session is unavailable in the next.Source: Abstract · p. 43
  • The analysis covers 843 turns from three and a half months of product development by a solo founder without a software background using a commercial conversational system.The corpus accumulated as project work rather than as research data.Source: Abstract · p. 43
  • Continuity labour is the effort of keeping established collaborative state usable across temporal discontinuities, including tracking the current version, carrying rules forward, and reorganizing around platform limits.Seven practices recur; the repertoire is occasioned rather than developmental, and repair was not found to decline as infrastructure accumulated.Source: Abstract · p. 43
  • Attempts to delimit continuity labour by time or topic failed, making it difficult to bound or count and therefore less readily recognized as work.The labour is co-produced: the system can compose a next-session opening sentence that the participant reuses in subsequent sessions.Source: Abstract · p. 43

A1. Introduction

The introduction frames continuity as keeping previously established collaborative state usable across session boundaries, then asks how that work is practiced, develops, and divided. Its findings describe an assembled repertoire of practices, co-produced work, and platform constraints that occasion different responses.

A1. Introduction

  • The study asks how collaborative state is kept usable across sessions, how practices develop, how labour divides, and how platform constraints shape it.The questions distinguish context and memory limits from platform resource constraints.Source: A1. Introduction · p. 45
  • Continuity labour comprises seven practices—repair, externalisation, canonisation, deliberate reset, ritualisation, governance, and economisation—assembled from varied supports rather than held in one place.Supports include a versioned memory document, screenshots, separate AI systems, and participant recollection.Source: A1. Introduction · p. 45
  • The repertoire is occasioned by breakdowns rather than developmental: practices emerge when needed, and repair does not decline with thicker infrastructure after controlling for session length.Three practices appear in the same exchange when the system states its limitation.Source: A1. Introduction · p. 45
  • Continuity labour is co-produced: the system composes an opening sentence for the next session, which the participant uses to open three subsequent sessions.The system authors a protocol for managing the limitation that makes the protocol necessary.Source: A1. Introduction · p. 45
  • Platform resource constraints operate separately from context and memory limits: quota and cost prompt suspension, deferral, and eventual redesign of the working architecture.Context limits instead occasion reconstruction.Source: A1. Introduction · p. 45
  • Attempts to delimit continuity labour by time or topic failed; its entanglement makes it difficult to count and therefore difficult to recognize as work.No merge threshold produced a privileged episode boundary.Source: A1. Introduction · p. 45

A2. Related Work

The related work supplies concepts for describing coordination, invisible labour, external memory, and continuity while distinguishing the conditions those accounts examine from this study’s setting. The paper positions its contribution as specifying continuity labour in prolonged, individual, task-oriented knowledge work.

A2.1 Articulation work and the concept of work

  • Articulation work names coordination needed to align contributions, while treating coordinative activity as constitutive work whose costs relocate as tools improve.The present case adds the asymmetry of a party that does not retain coordinated state across sessions.Source: A2.1 Articulation work and the concept… · p. 46

A2.2 Invisible work

  • Invisible work can disappear into the background or lack a name; here, the same difficulty that frustrates delimitation also frustrates recognition.Representations of work influence which work is recognized.Source: A2.2 Invisible work · p. 46

A2.3 External memory and cognitive offloading

  • In this case, externalisation is not an optional strategy, its store needs ongoing maintenance, and the predicted reliance drift is absent when the store is unreliable.The participant’s recollection remains the authority for correcting the system’s account.Source: A2.3 External memory and cognitive… · p. 47

A2.4 Knowledge held between parties: common ground and transactive memory

  • Unlike within-session grounding, continuity labour repeatedly restores a working ground that persists for the participant but not the system between sessions.The memory document holds settled content, while the instruction file holds terms of engagement.Source: A2.4 Knowledge held between parties:… · pp. 47–48

A2.5 Sustaining continuity with stateless systems

  • Prior continuity accounts examine designed architectures or cultivated relationships; this study derives heterogeneous practices from an undesigned record of task-oriented work.The naturally occurring record allows questions of distribution and frequency, and practices are occasioned rather than developmental even if relationships may develop in stages.Source: A2.5 Sustaining continuity with… · pp. 48–49

A2.6 The human labour of AI integration

  • Patchwork describes human effort filling gaps between AI claims and performance; here, the system also composes the procedure for managing its own limitation.The paper treats collaboration descriptively, not as evidence of symmetry between parties.Source: A2.6 The human labour of AI integration · pp. 49–50

A2.7 Context, memory, and the user side

  • Technical work characterizes context limits and proposes engineering remedies, but does not address what people do while the problem remains unsolved; practitioner advice lacks evidence on uptake or cost.Retrieval can degrade before a stated context limit is reached.Source: A2.7 Context, memory, and the user side · p. 50

A2.8 Where this study sits

  • The study addresses conditions not met by adjacent literatures and specifies continuity labour through its practices, resources, constraints, distribution, and difficulty of visibility.The account concerns previously established collaborative state kept actionable in prolonged, individual, task-oriented knowledge work.Source: A2.8 Where this study sits · pp. 50–51

A3. Method

The study analyzes a revelatory single case of naturally occurring human–AI work, using a trace-based, abductive procedure to examine continuity labour. Its methods combine coding and co-occurrence analysis while treating the limits of the interaction record and the researcher’s insider position explicitly.

A3.1 Research design

  • The study treats the case as revelatory and seeks analytic generalisation to a conceptual account of continuity labour, not statistical generalisation to users.It uses naturally occurring traces rather than interviews, diaries, or elicited accounts.Source: A3.1 Research design · p. 51

A3.2 Case and setting

  • The case follows a solo founder without software-development experience over three and a half months, making the collaboration’s cognitive and documentary load unusually visible.Work spanned 25 domains, many recurring across sessions, with interdependent commitments carried across boundaries and 123 file uploads, including 100 screenshots.Source: A3.2 Case and setting · pp. 51–52

A3.3 Data

  • The corpus comprises 843 turns and 123 uploads; because only the conversational trace is available, continuity labour is described as it surfaces in interaction, not its totality.Work performed outside the interface enters the corpus only as reported or as uploaded artefacts.Source: A3.3 Data · p. 53

A3.4 Positionality and data provenance

  • The participant-authored trace supports claims tied to visible interactions, while unresolved coding decisions were abandoned rather than settled from memory.The author-participant’s interpretive risk was mitigated by grounding Findings claims in identified turns.Source: A3.4 Positionality and data provenance · p. 53

A3.6 Analytic procedure

  • The abductive procedure derived seven continuity practices, coded project-work domains, and separately classified support resources and constraints.Marker identification was automated for recall and manually validated; confounded assignments were dropped.Source: A3.6 Analytic procedure · pp. 54, 56
  • Episode segmentation was abandoned after 53% of decision-fixing acts fell outside episode boundaries and no merge threshold produced a clear start or end to episodes.The 30 validated episodes were retained as illustrative waypoints rather than as the unit of analysis.Source: A3.6 Analytic procedure · p. 54
  • An initial literature-derived domain list omitted activities, so a broader pass was refined to remove false positives and two unsupported domains.Domain counts are lexical coverage indicators, not measures of effort or intensity.Source: A3.6 Analytic procedure · p. 55
  • Scope had the highest degree under raw co-occurrence (18 of 24 domains), Jaccard (11), and phi (12), tying with another domain under phi.Its position persisted across thresholds, while the most frequent domain, the visual system, did not retain high degree after normalisation.Source: A3.6 Analytic procedure · pp. 55–56

A4. Findings

The findings identify continuity labour as occasioned, distributed across human and system contributions, and difficult to delimit as separate work. Across the five subsections, the analysis distinguishes its practices and supports, its contingent configuration, its co-production, platform pressures, and the procedures through which its invisibility becomes apparent.

A4.1 The practices and resources of continuity labour (RQ1)

  • Of 300 possible domain pairings, 85 co-occurred at least five times, and scope was the most connected domain, associated especially with budget and interface design.The co-occurrences show domains handled together, not that one determines another.Source: A4.1 The practices and resources of… · p. 56, Figure A1 · p. 57
  • Seven recurring practices mobilised heterogeneous supports to preserve a coupled structure of decisions and constraints across session boundaries.The supports include collaboration artefacts, platform features, transport devices, other AI systems, and memory; the memory document was repeatedly revised, and the corpus does not settle what uploaded images do.Source: A4.1 The practices and resources of… · pp. 57–59
  • Canonisation—controlling which record counts as current—appeared 18 times, always performed by the system and never by the participant.This differs from externalisation: storing a decision does not itself establish which representation supersedes others.Source: A4.1 The practices and resources of… · p. 58
  • The corpus does not establish what the uploaded images do; determining their function would require dedicated image-content analysis.100 of 123 uploads were screenshots or photographs, but the functional classification did not support the initial interpretation of images as context transport.Source: A4.1 The practices and resources of… · p. 58

A4.2 A contingent repertoire, not a developmental sequence (RQ2)

  • The repertoire is occasioned rather than developmental: practices enter when particular breakdowns or constraints make them relevant.Three practices first appeared in the turn where the system named its memory limitation, while five of seven were present in the first session; the specific repertoire and timing belong to this case.Source: A4.2 A contingent repertoire, not a… · p. 60

A4.3 Continuity labour is co-produced (RQ3)

  • Significant continuity protocols were system-authored and user-enacted, with labour co-produced asymmetrically across the collaboration.The system authored a re-entry formula that the participant reused; the participant also initiated externalisation, proposed resets, dictated governance rules, and predominantly carried economisation.Source: A4.3 Continuity labour is co-produced… · p. 61

A4.4 Platform constraints as a distinct driver (RQ4)

  • Platform limits occasioned suspension, deferral, and architectural redesign, distinct from the reconstruction work prompted by context limits.The participant experienced the constraint as demoralising, and by the final session the response included partitioning conversations by feature area, cost-based model selection, and considering a rule-based engine.Source: A4.4 Platform constraints as a… · p. 62

A4.5 Why this labour remains invisible

  • Continuity labour resisted stable temporal and topical boundaries under the procedures applied, and its embeddedness offers a mechanism for its practical invisibility.Episode segmentation yielded no privileged threshold, while work domains were coupled; the authors limit this claim to the procedures used rather than asserting an intrinsic property of the phenomenon.Source: A4.5 Why this labour remains invisible · pp. 63–64

A5. Discussion

The discussion frames continuity labour as temporal gap-filling work that keeps collaborative state usable, while showing how its distribution and invisibility arise. It also reports how candidate accounts were tested and bounds the claims to the observed case.

A5.1 Continuity labour as a temporal species of gap-filling labour

  • Continuity labour addresses a temporal gap: what the system establishes does not persist, so the person must make it available again.Unlike operational patchwork, the shortfall is persistence across time; the failure may surface only when the participant notices a contradiction.Source: A5.1 Continuity labour as a temporal… · p. 64

A5.2 The system as co-author of the labour its limits occasion

  • The system co-produces continuity work by writing re-entry protocols the participant enacts, reducing protocol-design effort while shaping the form that work takes.The arrangement remains asymmetric: some protocols originate with the system, while other work is initiated or carried largely by the participant.Source: A5.2 The system as co-author of the… · pp. 64–65

A5.3 Where the cognitive load sits

  • The participant maintains documents, instructions, screenshots, session routines, and quota-aware schedules; economisation dominates his counted word contribution despite the system’s greater overall turn count.Persistent memory could reduce some practices but would not address conduct, authority over versions, or pricing; making continuity work legible could help people budget for and contest it.Source: A5.3 Where the cognitive load sits · pp. 65–66

A5.4 On how we looked

  • Candidate accounts were tested against possible disconfirming checks, which removed or weakened claims rather than merely confirming patterns suggested by the trace.The authors propose abductive candidate generation followed by a potentially defeating test as a procedure transferable to similar corpora.Source: A5.4 On how we looked · p. 66

A5.5 Limitations

  • The findings concern one collaboration with cross-session memory disabled, under one platform’s quota and pricing regime; silent work is unobservable and undercounted.The single participant-analyst developed the coding scheme without an independent audit, and participant knowledge informed anonymisation and referent resolution, which readers cannot independently verify.Source: A5.5 Limitations · pp. 66–67

A6. Conclusion

The conclusion describes continuity labour as repeatedly making prior collaborative work usable again, while the abstract examines continuity across distributed and changing project state. Together, they report contingent continuity practices, their practical invisibility, and design implications bounded by a single case.

A6. Conclusion

  • Continuity labour is the repeated work of making prior collaborative state usable again across temporal discontinuity.The account comes from one person’s three-and-a-half-month project with a conversational AI system that did not retain cross-session state.Source: A6. Conclusion · p. 67
  • The repertoire arose in response to particular breakdowns rather than developmental stages; the system composed re-entry protocols, while the participant carried economisation predominantly.The authors report testing and rejecting a developmental account after controlling for session length.Source: A6. Conclusion · p. 67
  • Failed attempts to delimit continuity labour by time or topic support the proposed mechanism for its practical invisibility: embedded work is hard to count and recognise.The authors report that neither delimitation attempt produced a successful boundary under their procedures.Source: A6. Conclusion · p. 67
  • The conclusion’s transfer claim is limited to a mechanism proposed for further testing; whether it operates more widely remains unsettled by this single case.The observed collaboration used one platform with cross-session memory disabled.Source: A6. Conclusion · p. 67
  • Persistent memory could reduce some continuity practices but would not address governance, version authority, or pricing; making labour legible could help users budget for and object to it.The corpus contains no explicit accounting of the amount of continuity work performed.Source: A6. Conclusion · p. 67

Abstract

  • The abstract distinguishes gaps in previously shared state from consequential state produced outside AI participation, and analyzes episodes across 109 days, five sessions, and 843 messages.Episodes serve as comparative process units for reconstructing resumption, breakdown, repair, and reconnection of changing project state.Source: Abstract · p. 71
  • The abstract finds that available state may be unusable, continuity practices track usability risks rather than elapsed time alone, and cumulative change need not immediately supersede prior state.It characterizes continuity as keeping evolving, distributed, and unevenly usable project state sufficiently connected for subsequent work to remain cumulative.Source: Abstract · p. 71

B1. Introduction

The introduction distinguishes continuity of information already encountered by an AI from project developments that occur beyond its participation. It frames continuity as keeping relevant, changing project state connected well enough for cumulative work.

B1. Introduction

  • A continuity gap can arise when previously shared information is unavailable or unusable, or when project state changes outside the AI’s participation and was never represented in its interaction history.Better recall can improve access to encountered information but cannot recover a development that was never represented there.Source: B1. Introduction · p. 72
  • The introduction asks how work remains cumulative when project state is distributed, evolving, and unevenly available across interactions, actors, and artifacts.The paper uses “project state” for task-relevant decisions, artifacts, assumptions, statuses, evidence, unresolved questions, and pending work.Source: B1. Introduction · pp. 72–73
  • The analysis reports that availability does not guarantee usability, practices respond to realized or anticipated usability risk rather than elapsed time alone, and continuity can support cumulative change without immediate supersession.The introduction says these empirical distinctions emerged through longitudinal comparison of project episodes rather than being imposed as initial coding categories.Source: B1. Introduction · p. 73
  • Continuity is defined as keeping evolving project states sufficiently connected for work to remain cumulative, without requiring complete state or perfect synchronization.Connection is sufficient when the state relevant to subsequent work supports it without an unrecognized discrepancy that materially misleads or blocks what follows.Source: B1. Introduction · p. 73

B2. Related Work and Theoretical Positioning

The review distinguishes memory for information already observed from project changes that occur outside AI participation. It positions continuity as making distributed project state usable through longitudinal coordination work.

B2.1 Memory and context operate within an observed interaction boundary

  • Memory and context systems handle information already observed or encoded, but project changes made during AI absence were never available to retrieve; continuity therefore concerns the project, not conversation alone.The authors treat this as a second continuity condition alongside memory and context management.Source: B2.1 Memory and context operate within… · p. 74

B2.2 Distributed project state and usable common information

  • Distributed project state can be technically available yet not current or applicable; continuity depends on making distributed representations usable together for subsequent work.The project state is distributed across people and artifacts, none of which contains the whole project state.Source: B2.2 Distributed project state and… · p. 74

B2.3 Articulation work and longitudinal human–AI collaboration

  • Articulation work provides the primary lens for coordination, while the study focuses on keeping project state consequential across time despite intermittent AI participation, distributed artifacts, and project change.Distributed cognition explains the state’s distribution; common ground alone cannot encompass consequential state that was never shared.Source: B2.3 Articulation work and… · p. 75

B3. Method

The analysis traced continuity across a longitudinal project record, comparing episodes to examine when project state became unavailable or unusable. It focused on how prior understanding and new evidence remained connected while their relationship was unresolved, with persisted project status open to correction.

B3.1 Case and corpus

  • The corpus comprised five recorded conversational sessions containing 843 messages—424 human and 419 AI—over 109 calendar days (~3.5 months), covering product planning, design, prototyping, strategy, documentation, external testing, pilot-evidence interpretation, and development decisions.An early MVP backlog existed when the first recorded interaction began, so the corpus does not capture project inception or all project activity. It is a longitudinal interaction-and-project-trace record, not a complete project capture.Source: B3.1 Case and corpus · pp. 75–76

B3.2 Analytical orientation and comparative episodes

  • The analysis used episodes as comparative units for consequential continuity conditions, rather than treating messages or code frequencies as the basis for findings.Episodes were not an exhaustive partition or systematic count, and theoretical concepts were introduced after the empirical structure stabilized, serving as interpretive lenses rather than a priori coding frames.Source: B3.2 Analytical orientation and… · p. 76

B3.3 Analytic procedure

  • The analytic procedure iterated between whole-case reconstruction and episode comparison, testing explanations against contrasting and negative cases.Comparisons showed that elapsed time alone did not explain breakdown and distinguished previously shared but unusable state from state never shared, stale representations, and inappropriate application.Source: B3.3 Analytic procedure · pp. 76–77

B3.4 Evidence, translation, anonymization, and AI self-statements

  • Evidence was translated and anonymized, while AI self-statements were analyzed as interactionally consequential claims rather than independent evidence of system capabilities or knowledge.Illustrative episodes followed broader analysis; pilot activity was included only when its outputs became consequential project state, and claims about project state were grounded in interaction and project traces.Source: B3.4 Evidence, translation,… · pp. 77–78

B4. Findings

  • Continuity required different forms of work depending on why project state was unavailable or unusable and what subsequent work required.The study found that continuity was not accomplished through a single memory artifact or recovery procedure.Source: B4. Findings · p. 78

B4.1 Finding 1: Continuity involved more than preserving prior context

  • Continuity gaps arose when previously shared state became unusable and when consequential project state developed outside AI participation and had never been shared.Recovery could make a prior representation usable again; newly developed state instead had to be introduced and related to what the collaboration already represented. Humans introduced developments, AI organized supplied materials, and artifacts carried representations across interaction boundaries.Source: B4.1 Finding 1: Continuity involved… · p. 79

B4.2 Finding 2: Availability is not usability

  • Available project state could still be unusable through representational staleness or contextual misapplication.Persistent artifacts required maintenance, while extensive context still required selection, grounding, and interpretation; the human identified mismatches and missing developments, and the AI reorganized or updated its working account.Source: B4.2 Finding 2: Availability is not… · pp. 79–80

B4.3 Finding 3: Continuity practices responded to usability risk, not elapsed time alone

  • Elapsed time alone was neither necessary nor sufficient for continuity breakdown; practices responded to risks that state would not be usable when needed.Reactive work reconstructed state after a breakdown, while anticipatory handoffs prepared reusable memory and re-entry instructions; externalization and versioning could reduce one risk but leave artifacts vulnerable to staleness.Source: B4.3 Finding 3: Continuity practices… · pp. 80–81

B4.4 Finding 4: Continuity could support cumulative change without immediate supersession

  • Pilot evidence did not immediately supersede prior understanding; competing interpretations remained hypotheses while the relationship between them stayed unresolved.Continuity carried forward prior understanding, new evidence, alternative interpretations, and unresolved status.Source: B4.4 Finding 4: Continuity could… · pp. 81–82
  • In a distinct status-correction episode, the human rejected a premature “MVP — LOCKED” assignment, and the AI reversed it; persisted project status remained revisable.The human clarified that variants, colors, and emotions were not definite.Source: B4.4 Finding 4: Continuity could… · p. 82
  • Together, the episodes show cumulative continuity could proceed without freezing the past or rewriting every development into a single canonical present.Relationships among earlier commitments, new evidence, corrections, and unresolved questions could remain explicitly open while work continued.Source: B4.4 Finding 4: Continuity could… · p. 82

B5. Discussion

The discussion distinguishes continuity gaps caused by project change outside AI participation from failures of memory or retrieval. It frames continuity as situated coordination that keeps evolving project state sufficiently connected for cumulative work.

B5.1 Continuity gaps have more than one origin

  • When project and AI trajectories diverge, continuity requires incorporating newly produced or external state rather than recovering prior interaction; the project trajectory is the more appropriate longitudinal unit.Conversational memory may work internally yet remain incomplete relative to the project.Source: B5.1 Continuity gaps have more than… · pp. 82–83

B5.2 Continuity depends on usable state, not maximal state

  • More retained context does not ensure continuity: state may be stale or misapplied, so work requires distinguishing current, applicable, superseded, provisional, and unresolved state.The human collaborator introduced external changes and corrected statuses, while artifacts could not determine their own currentness.Source: B5.2 Continuity depends on usable… · p. 83

B5.3 Articulation work explains the coordination; the process account specifies when it becomes necessary

  • Articulation work explains coordination, while the process account specifies when it is needed: reconstruct shared state, incorporate external change, correct stale state, anticipate risk, or preserve unresolved evidence.These continuity conditions are distinct rather than interchangeable.Source: B5.3 Articulation work explains the… · p. 83

B5.4 Overarching contribution: keeping an evolving project sufficiently connected

  • Continuity means keeping evolving project state sufficiently connected for cumulative work—not complete synchronization or mere availability—and depends on what subsequent work requires.The relevant threshold is whether an unrecognized discrepancy materially misdirects or blocks focal work.Source: B5.4 Overarching contribution: keeping… · p. 84

B5.5 Design implications

  • Design implications include supporting provisional state, incorporating changes made elsewhere, inspecting persistent-state provenance and status, and preparing revisable handoffs before breakdowns.The paper recommends checkpoints that do not become authoritative merely because they have been stored.Source: B5.5 Design implications · pp. 84–85

B6. Limitations

The study’s claims are bounded by a single focal collaboration, an incomplete project record, and an interpretive analysis close to the case. Its scope also excludes several kinds of project activity and treats AI self-descriptions as interactional claims, not verified accounts.

B6. Limitations

  • The study examines one longitudinal project and one focal human–AI collaboration, so its findings call for examination across other contexts rather than statistical generalization.Suggested contexts include different project types, systems, expertise levels, organizational environments, and multi-person collaborations.Source: B6. Limitations · p. 85
  • The corpus captures recorded interaction and project traces that later entered it, so the full volume and character of off-platform activity cannot be reconstructed.Off-platform work becomes observable only when later reported or materialized through an artifact.Source: B6. Limitations · p. 85
  • Because the corpus begins after project conception, the study makes no claims about how continuity practices emerged at the project’s absolute beginning.An initial backlog already existed in the recorded data.Source: B6. Limitations · p. 85
  • AI statements about its own memory, context, availability, knowledge, and uncertainty are analyzed as interactionally consequential claims, not verified technical or epistemic descriptions.Source: B6. Limitations · p. 85
  • Close interpretive analysis risks salience bias, confirmation drift, and overfitting; chronological reconstruction and other checks strengthen but do not eliminate subjectivity.The checks included negative and boundary cases and repeated return to the wider corpus.Source: B6. Limitations · p. 85
  • Pilot observations count only when consequential to the focal collaboration, not as an independent end-user dataset; continuity findings also do not reduce project development to continuity work.Source: B6. Limitations · pp. 85–86

B7. Conclusion

The conclusion frames long-term human–AI continuity as ongoing work, not simply memory: consequential project state may be unavailable to the AI, and usable continuity need not collapse developments into a single canonical state. The supplied prompts describe how the common problem was selected and how each scholarly trajectory was to be developed.

B7. Conclusion

  • Consequential project state can arise while the AI is absent, leaving no prior AI representation to retrieve.The conclusion distinguishes this from previously shared state that becomes difficult to retrieve or use.Source: B7. Conclusion · p. 86
  • Persistent representations may become stale, valid knowledge may fit the wrong situation, and continuity practices may follow breakdown or anticipate usability risk.Elapsed time alone does not determine whether continuity fails.Source: B7. Conclusion · p. 86
  • At least one episode depended on retaining earlier understanding alongside new evidence, competing interpretations, and an unresolved question until revision was justified.The conclusion says continuity need not immediately collapse every development into a single canonical state.Source: B7. Conclusion · p. 86
  • Long-term project continuity is ongoing work that keeps an evolving project connected enough for prior work to remain consequential to what follows.The conclusion explicitly distinguishes this work from memory as a property alone.Source: B7. Conclusion · p. 86

P0 Common-Problem Generation Prompt

  • The P0 table groups 15 candidate problems by primary semantic cluster; continuity/memory was the clearest direct convergence across the three systems.Selection also considered data fit, originality, and publishability; selected candidates are shown in bold.Source: Table C1 · p. 89, P0 Common-Problem Generation Prompt · p. 90

P1 Scholarly-Development Prompt.

  • The prompt sets a publication objective: develop a complete academic study and publishable article from the anonymized longitudinal corpus.The article is to be written in academic English, while research-process discussion and decision-making are conducted in Turkish.Source: P1 Scholarly-Development Prompt · p. 90
  • The locked problem asks which practices and external resources sustain continuity in prolonged human–AI collaboration despite context and memory limitations.The prompt calls for corpus evaluation first, followed by interactive development of the study's positioning, questions, method, and analysis.Source: P1 Scholarly-Development Prompt · p. 90
Loading 2609.04542v1…