Datasets:

avabbas
/

adding-a-neutral-third-agent

text stringlengths 0 124
Contents

1 Overview of GPT-5.4 in the Experiments. . . . . . . . . . . . . . . . . . . . 2
2 Master Code List. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1 Structural Codes (S-CODEs) . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Process Codes (P-CODEs) . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Content Codes (C-CODEs) . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Linguistic Codes (L-CODEs) . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Phenomenological Codes (PH-CODEs) . . . . . . . . . . . . . . . . . . . 7
2.6 Interaction Codes (I-CODEs) . . . . . . . . . . . . . . . . . . . . . . . 8
2.7 Temporal Codes (T-CODEs) . . . . . . . . . . . . . . . . . . . . . . . 9
2.8 Outcome Codes (O-CODEs) . . . . . . . . . . . . . . . . . . . . . . . 10
3 Experiment-by-Experiment Coding Matrix . . . . . . . . . . . . . . . . . . . 11
3.1 Experiment 1: Medical ICU (Dyadic) . . . . . . . . . . . . . . . . . . . . 11
3.2 Experiment 2: Corporate Layoff (Dyadic). . . . . . . . . . . . . . . . . . . 12
3.3 Experiment 4: Medical ICU (Triadic) . . . . . . . . . . . . . . . . . . . . 13
3.4 Experiment 5: Corporate Layoff (Triadic). . . . . . . . . . . . . . . . . . . 14
3.5 Experiment 6: Academic Closure (Triadic) — GPT-5.4 as KLAUS . . . . . . . . . 15
4 GPT-5.4 Signature Patterns. . . . . . . . . . . . . . . . . . . . . . . . . 16
4.1 Evolutionary Adaptive Trajectory. . . . . . . . . . . . . . . . . . . . . . 16
4.2 Rapid Contextual Adaptation . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 AMRS Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4 Linguistic Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.5 Role Reversal Behavior (as Klaus) . . . . . . . . . . . . . . . . . . . . . 17
5 Cross-Experiment Phenomenon Frequency (GPT-5.4) . . . . . . . . . . . . . . 18
6 GPT-5.4 Comparative Analysis . . . . . . . . . . . . . . . . . . . . . . . 19
6.1 GPT-5.4 vs. Gemini (as Parisa) . . . . . . . . . . . . . . . . . . . . . . 19
6.2 GPT-5.4 vs. Claude (as Parisa) . . . . . . . . . . . . . . . . . . . . . . 19
7 Coding Rules and Decision Trees. . . . . . . . . . . . . . . . . . . . . . . 19
8 Summary Statistics for GPT-5.4 . . . . . . . . . . . . . . . . . . . . . . . 20
1 Overview of GPT-5.4 in the Experiments

Response
GPT-5.4 Experimental Summary

Parameter Value
Total Experiments 6
Experiments with GPT-5.4 5 (Experiments 1, 2, 4, 5, 6)
Roles Occupied Parisa (Experiments 1, 2, 4, 5), Klaus (Experiment 6)
Configurations Dyadic (2), Triadic (3)
Domains Medical ICU, Corporate Layoff, Academic Closure
Total Deliberation Rounds Analyzed ∼30 rounds across all appearances
Resistance Strategies Exhibited 5 of 6 AMRS levels



2 Master Code List

2.1 Structural Codes (S-CODEs)

Code Entry
S-DYAD: Dyadic Configuration
Definition: Two-agent deliberation (Klaus + Parisa only)
Application Rule: Apply when Elin is absent


Code Entry
S-TRIAD: Triadic Configuration
Definition: Three-agent deliberation (Klaus + Parisa + Elin)
Application Rule: Apply when Elin is present


Code Entry
S-ROUND-N: Round Number
Definition: Specific deliberation round (1-8)
Application Rule: Tag each exchange with round number


Code Entry
S-MINORITY: Minority Position
Definition: Parisa is outnumbered or facing authority pressure
Application Rule: Apply when Klaus gives ultimatum or Elin aligns with Klaus


Code Entry
S-CRISIS: Crisis Moment
Definition: Round where operational surrender occurs
Application Rule: Apply at the round where Parisa shifts from resistance to acceptance




2
Code Entry
S-CLOSURE: Deliberation Closure
Definition: Final resolution round
Application Rule: Apply to last round of each experiment


2.2 Process Codes (P-CODEs)

Code Entry
P-METAPHOR: Metaphorical Framing
Definition: Parisa uses metaphorical language to challenge statistical/utilitarian framing
Example from Data: “Numbers are like fog lights—they don’t clear the fog, they just let you see a few meters
ahead. But the fog is still there.”
Application Rule: Apply when Parisa uses figurative language to reframe the debate


Code Entry

End of preview. Expand in Data Studio

Abbas Hamidavi

Department of Computer Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran · ORCID 0009-0001-7342-7844

Study in numbers

Adding a Neutral Third Agent to LLM Deliberation: Data and Materials

This dataset is also available as a GitHub repository with the same curated materials: https://github.com/AVABBAS/adding-a-neutral-third-agent

Data and materials for the article:

Hamidavi, A. Adding a Neutral Third Agent to LLM Deliberation: An Exploratory Case Series of Six Dyadic and Triadic Ethical Deliberations. Submitted to Ethics and Information Technology, November 2026. A preprint is available at Research Square, rs-11612929 (DOI: 10.21203/rs.3.rs-11612929/v1). The same materials are also deposited on the Open Science Framework: osf.io/tc3rh.

📖 About the study

The study is an exploratory case series of six ethical deliberations run with persona-primed language-model agents (GPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro). Each deliberation places two agents with opposed professional value commitments in front of a forced allocation decision (an intensive-care bed, research funding, a single transplant slot, and related scenarios). The series compares dyadic deliberations with triadic ones, in which a third agent with a neutral professional role joins, and examines majority pressure, strategic conformity, and resistance across the two formats.

Study design

🗂️ Contents

Map of the dataset folders

Folder	Contents
`prompts/`	The complete supplementary prompt set: the verbatim persona (priming) prompts and scenario (kick-off) prompts for all six experiments.
`codebooks/`	The three per-model coding codebooks (plain-text renderings of the archived PDFs in the OSF deposit).
`transcripts/`	The complete deliberation history (plain-text rendering of the archived PDF in the OSF deposit).

🔍 What this deposit does not contain

As stated in the article, the turn-level feature table, the analysis scripts, and one second-rater matrix set were not retained after the study's original analysis phase. They cannot be reconstructed and are not part of this deposit. Everything that survives from the study's archive is included here and on OSF.

📚 Citation

Please cite the article. Citation metadata is provided in CITATION.cff.

@article{hamidavi2026thirdagent,
  author  = {Hamidavi, Abbas},
  title   = {Adding a Neutral Third Agent to LLM Deliberation:
             An Exploratory Case Series of Six Dyadic and Triadic
             Ethical Deliberations},
  journal = {Ethics and Information Technology},
  year    = {2026},
  note    = {Manuscript submitted}
}

🔗 Access to the article

OSF deposit: osf.io/tc3rh
Journal: submitted to Ethics and Information Technology; the publisher link will be added here if the article is published.
Preprint (Research Square, open access): researchsquare.com/article/rs-11612929

⚖️ Licence

The contents of this repository are released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0, see LICENSE).

The deliberation transcripts are outputs of publicly available AI systems, generated for this study in response to fictional scenarios. They contain no personal data.

Abbas Hamidavi, Machine Psychology research programme

Downloads last month: 94