Jhinn Bay, Psychology BA
Human–Computer Interaction, Cyberculture, & Behavioral Patterns Online
Mental Wealth Academy
DOI (pre-registration, materials): https://doi.org/10.55277/researchhub.uzdhftd8

Abstract: This paper introduces the Daemon Model, instantiated as the AI persona Azura, an AI-mediated governance participant designed to address democratic deficits in many contemporary decentralized autonomous organizations (DAOs). Drawing on Carl Jung’s conception of the daemon and a proposed new disciplinary lens—Cyber Psychology—we frame the Daemon Model as a transparent, explicitly delegated mediator that translates qualitative, non-stakeholder input into quantifiable governance influence. We present the system architecture (RAG with DeepSeek reasoning model, Supabase vector embeddings, local containerized deployment), ethical design principles (consent, transparency layering, human override), and a pre-registered experimental protocol for randomized mini-game trials (storytelling, Dungeons & Dragons–style roleplay, and MMORPG–style collaborative scenarios) to evaluate governance alignment, participation equity, and perceived legitimacy in a diverse sample (planned N ≈ 100–200). All participant-facing on-chain interactions for experiments will be executed on an Ethereum testnet to sandbox financial risk. IRB review is pending; project materials and protocol are archived at the DOI above.

Keywords: Daemon Model, Computer Psychology, DAOs, hyperreal, RAG, DeepSeek, Supabase, AI governance, mini-games, Ethereum testnet, neurosymbolic traceability.

The COVID-19 pandemic accelerated the migration of social life into algorithmically mediated, attention-driven spaces. In this environment, representations and curated content can come to shape what communities treat as real and consequential: a condition Jean Baudrillard called the hyperreal, in which simulation precedes or substitutes for reality. This shift alters how institutions, communities, and individuals form judgments, preferences, and identities in digital public spheres.

Parallel to these cultural shifts, advances in large language and reasoning models (LRMs) have sparked debate about what these systems actually do. Recent technical work has argued that LRMs can display impressive emergent behaviors while also exhibiting systematic limitations in generalization or reasoning under certain problem structures. Such debates underscore the need to ground claims about model capabilities in careful evaluation design and to consider how human sociocultural environments interact with model behaviors.

We propose the Daemon Model—an explicitly delegated AI mediator designed to amplify non-stakeholder voices within governance processes, not by replacing human agency, but by translating distributed qualitative inputs into auditable, tokenized governance recommendations. Drawing metaphorically on Jung’s daemon—an autonomous mediator between conscious and unconscious content—we operationalize Azura to function transparently as a procedural amplifier: it receives multi-channel inputs, retrieves relevant organizational knowledge, generates candidate proposals, and casts delegated votes under pre-specified, revocable rules.

This approach responds to two linked problems. First, many DAO architectures replicate economic and attention hierarchies: token-weighted voting privileges capital holders and attention leaders, producing an appearance of decentralization without substantive redistribution of influence. Second, AI systems that obscure their provenance or reasoning risk exacerbating legitimacy deficits. The Daemon Model is designed to be auditable, consented to, reversible, and sensitive to socio-cognitive dynamics—what we call Computer Psychology—the study of parasocial relationships, semiotics, and cultural semiotics that shape human–AI interaction.

Section 2 summarizes structural risks in digital governance and responds to anticipated criticisms. Section 3 develops the Computer Psychology framework and details the Daemon Model architecture. Section 4 describes the experimental methods—mini-game design, recruitment, and ethical safeguards. Section 5 presents the pre-registered Statistical Analysis Plan (SAP) and sample/power guidance. Section 6 discusses implications, limitations, and next steps. Appendices contain the full SAP, OSF registration note, and operational implementation details.

Contemporary decentralized governance frequently reproduces inequalities it claims to redress. Token-weighted voting privileges capital holders and funds that confer both decision rights and attention. Attention economies amplify spectacle, memetic success, and parasocial prominence. Many constituencies—students, low-income participants, casual users—remain effectively disenfranchised because influence is tied to capital and continuous engagement. These dynamics create performed decentralization that lacks substantive redistribution of decision-making power.

Hyperreal dynamics—parasocial attachments, virality, and rapid memetic shifts—exacerbate cognitive biases (anchoring, availability, recency) and undermine deliberative norms. Interventions that ignore socio-cognitive context risk either steepening existing inequalities or inadvertently generating new forms of manipulation. Effective design must therefore attend to cultural semiotics, interface framing, and the dynamics of trust and perceived legitimacy.

We identify five principal critiques and concrete design responses.

1. Legitimacy & Representation.
   Criticism: An AI casting votes risks illegitimacy—who authorized the daemon?
   Design response: Daemon participation is strictly opt-in and governed by explicit delegation rules. Communities must consent to translation and tokenization mappings (how qualitative input converts to on-chain votes). Delegations are time-limited and revocable; audit trails record all inputs, mappings, and outputs.
   

2. Manipulation & Gaming.
   Criticism: Adversaries could manipulate inputs (sockpuppetry, coordinated brigading) to bias daemon outputs.
   Design response: Multi-channel aggregation reduces single-channel attack surfaces. Defenses include reputation weighting, rate limits, outlier detection, and adversarial stress testing. Mini-game trials include adversarial injection conditions to measure and refine these defenses.
   

3. Bias Amplification & Model Error.
   Criticism: The daemon could codify and amplify systemic biases in data or sampling.
   Design response: Debiasing layers (counterfactual augmentation and fairness-aware objectives), continuous bias audits, and neurosymbolic traceability ensure decisions are accompanied by symbolic evidence and remediation paths.
   

4. Transparency vs. Usability Tradeoff.
   Criticism: Technical transparency confounds lay users, enabling adversarial exploitation.
   Design response: Provide tiered explanations: (a) plain-language summaries with confidence bands for general participants, (b) mid-level rationales linking to evidence for engaged members, and (c) machine-readable audit logs for auditors.
   

5. Legal, Ethical, and Institutional Risks.
   Criticism: Delegation of governance functions to AI raises regulatory concerns.

To generate robust behavioral and psychometric data while controlling for real-world harms, we adopt mini-games as experimental testbeds. Mini-games emulate governance tradeoffs in low-stakes, interactive formats (storytelling, DnD-style roleplay, and MMORPG missions), enabling:

• Safety: Testnet on-chain interactions and sandboxed treasury mechanics reduce financial risk.

• Causal leverage: Randomized manipulations of explanation granularity, adversarial injection, and tokenization rules.

• Granular behavioral traces: Rich, timestamped logs of chat, proposals, votes, and in-game actions.

• Psychometric anchoring: Embedded pre/post measures (trust, perceived fairness, parasocial attachment).

We propose Cyber Psychology as an interdisciplinary lens extending HCI and UX to include semiotics, parasocial dynamics, memetics, and cultural frames. Cyber Psychology focuses on how language, persona design, symbolic cues, and interface affordances shape human attachments to non-human actors and influence behavior in networked governance contexts.

The Daemon Model operationalizes Azura via a Retrieval-Augmented Generation (RAG) pipeline built around a reasoning-capable LLM. Design principles:

• Transparency & provenance: Every recommendation links to retrieved evidence and symbolic metadata (policy tags, provenance strings).

• Debiasing & fairness: Dedicated layers evaluate decisions for bias signals and apply corrective transformations.

• Human oversight: All outputs support human review, stakeholder verification, and override.

• Testnet first: On-chain execution in experiments uses Ethereum testnet to sandbox financial effects.

Implementation details (system stack): Azura’s pipeline uses the DeepSeek reasoning model as the base LLM. Embeddings and vector indexing are stored in a self-hosted Supabase vector store; retrieval and orchestration follow LangChain-style patterns. Model inference and orchestration run in containerized local environments (Kubernetes/Docker) to maintain version control and data residency. The label neurosymbolic refers to symbolic, human-interpretable fragments—policy tags, decision rules, provenance strings—stored alongside retrieved documents in the RAG dataset, enabling symbolic justification generation linked to evidence. Fine-tuning artifacts, prompt templates, and model configuration are versioned and archived in the project repository (OSF DOI).

Primary objectives:

1. Alignment: Does Daemon-mediated governance improve alignment between participant preferences and executed policies relative to human-only governance?

2. Equity: Does Daemon mediation increase participation equity (measured by Gini coefficient of contributions)?

3. Legitimacy: Does Daemon mediation maintain perceived legitimacy and trust (Trust in Automation scale) compared to human governance?

4. Robustness: What is the Daemon’s susceptibility to adversarial inputs, and how do defenses perform?

Design: Between-session randomized controlled design. Units of randomization are sessions. Sessions are randomly assigned to:

• Daemon condition: Azura ingests aggregated inputs and casts a pre-specified, revocable fraction of votes under transparent mapping rules.

• Human condition: Human participants vote directly; no AI delegation.

Within Daemon sessions, we randomize sub-conditions:

• Explanation granularity: summary vs. detailed explanations.

• Tokenization scheme: fixed vs. adaptive mapping of inputs to voting weight.

• Adversarial injection: orthogonal binary manipulation (simulated coordinated inputs in select sessions).

Mini-game mechanics: Each session runs one mini-game (storytelling/DnD/MMORPG scenario) with 1–3 governance decisions to be made by the group. Participants provide multi-channel inputs (free text chat, structured votes, and short voice notes). The Daemon aggregates inputs, retrieves relevant organizational memory and prior precedent, generates candidate proposals, and (if delegated) casts a designated share of votes.

Operational constraint: All on-chain actions in experiments use Ethereum testnet wallets; no real treasury funds are at risk during trials.

Planned sample: 100–200 participants total across sessions. Recruitment aims for demographic diversity and a range of Web3 familiarity, intentionally including many participants unfamiliar with crypto to evaluate onboarding and inclusion. Participants aged 18+ will consent in-app; no classification by educational status (e.g., “student”) is required.

Session sizes: Sessions of ~8–25 participants (we aim for smaller clusters where feasible to increase statistical power per SAP recommendations).

Primary outcomes:

• Governance alignment (session or participant level): Cosine similarity between the vector of aggregated participant preferences and the executed policy vector (range 0–1). Operational definition in Methods: Let ps\mathbf{p}_sps​ be mean participant preference vector over K policy dimensions and as\mathbf{a}_sas​ the executed policy vector; alignment = (ps⋅as)/(∥ps∥ ∥as∥)(\mathbf{p}_s \cdot \mathbf{a}_s)/(\|\mathbf{p}_s\|\,\|\mathbf{a}_s\|)(ps​⋅as​)/(∥ps​∥∥as​∥).

• Participation equity: Gini coefficient across contribution counts per participant per session.

• Trust & perceived legitimacy: Trust in Automation scale (pre/post change) and Perceived Procedural Fairness.

Secondary outcomes: Decision processing time, reversal rate, parasocial attachment (short scale), susceptibility to adversarial perturbation (effect size change with injected inputs), qualitative satisfaction.

Behavioral logs: All chat messages, timestamps, proposals, votes, and in-game actions are stored in anonymized, timestamped logs.

IRB submission is pending and will be completed prior to field pilots. In the meantime, the protocol includes:

• Informed consent: In-app consent screen describing testnet use, data collection, and opt-out.

• Debrief: Post-session debriefing describing AI mediation and how inputs were used.

• Opt-out & recall: Participants can withdraw inputs; governance votes are revocable.
  Sandboxing: All on-chain interactions use Ethereum testnet during experiments; production treasury interactions are gated by governance charter changes and legal review.

• Data handling: PII is stored separately from analysis logs; anonymized IDs used for analysis. Data release will follow de-identification procedures (remove PII, hash IDs, aggregate small cells) and IRB guidance.

A full, unabridged SAP is provided as Appendix B. Below we summarize the confirmatory analysis plan and sample/power guidance.

Primary analysis approach: Mixed-effects regression (participants nested in sessions; random intercepts for sessions and participants where repeated measures occur). Analyses are intent-to-treat. Primary family: alignment, participation Gini, trust change.

Planned sample range: 100–200 participants. Because clustering (session structure) affects power, we pre-register mitigation strategies: run more sessions with smaller cluster sizes (m ≤ 12), incorporate within-participant repeated measures (multiple decisions per participant), and emphasize estimation of effect sizes with confidence intervals for this feasibility study.

Pre-registered hypotheses:
H1: Daemon sessions will show greater alignment than Human sessions (Daemon > Human).
H2: Daemon sessions will show lower Gini coefficients (more equitable participation).
H3: Trust in Daemon sessions will be statistically equivalent to Human sessions within ±0.3 SD.

Multiple comparisons & missing data: Primary family tested at α = 0.05; secondary outcomes corrected via Benjamini–Hochberg FDR. Missing data handled via Multiple Imputation (MICE) with sensitivity analyses.

(Full SAP with detailed model formulas, power arithmetic, and R/Python templates is provided in Appendix B.)

This manuscript introduces the Daemon Model as a theoretically grounded and operationally feasible approach for AI-mediated governance. Key contributions:

• A conceptual integration of Jungian daemon metaphors with the emerging discipline of Computer Psychology to analyze AI–human co-participation in governance.
  

• A concrete system architecture (DeepSeek + RAG + Supabase + local orchestration) emphasizing provenance, auditable symbolic justifications, and human oversight.
  

• A pre-registered experimental plan using mini-games and Ethereum testnet to evaluate governance alignment, equity, and legitimacy under controlled manipulations.

When designed with explicit consent, layered transparency, continuous fairness audits, and human override, AI mediators can serve as procedural amplifiers for voices excluded by token-weighted architectures. Mini-game experiments provide a feasible, ethical platform to evaluate behavioral dynamics before field deployment.

• This manuscript documents design and pre-registered plans; empirical data collection is pending IRB approval. Findings will be provisional until experiments are completed and analyzed.

• Mini-games intentionally trade some ecological validity for experimental control; findings will need triangulation with cautious field pilots.

• Legal/regulatory constraints and cultural differences require careful study before generalizing the model.

The Daemon Model reframes AI in governance not as an opaque authority but as a transparent, consented mediator that can translate collective qualitative inputs into accountable influence. Our architecture, ethics, and pre-registered experimental roadmap provide a foundation for empirical evaluation. We invite collaboration, critique, and replication.

Thanks to contributors within the Mental Wealth Academy research collective for feedback on the design and ethics plan. Pre-registration materials and code skeletons are archived at DOI: https://doi.org/10.55277/researchhub.uzdhftd8.

Baudrillard, J. (1981). Simulacra and Simulation.

Shojaee, P., et al. (2025). The Illusion of Thinking: Understanding Reasoning Model Limits. Apple Machine Learning Research. [Available: https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf]

Lewis, P., et al. (2020). Retrieval-Augmented Generation (RAG). Conference Paper.

(Full references for neurosymbolic AI, HCI, DAO governance literature, and psychometrics will be appended per journal citation rules.)

Title: Piloting the Daemon Model: Mini-game experiments on AI-mediated governance, participation equity, and trust
Authors: James Q. Marsh et al.
DOI (materials & pre-registration): https://doi.org/10.55277/researchhub.uzdhftd8
Primary outcomes: governance_alignment (cosine similarity), participation_gini, trust_delta.
Planned N: 100–200 participants across multiple sessions.
Design: Between-session RCT; within-session repeated measures where feasible.
IRB: submission pending; no field trials until IRB approval.
Notes: All on-chain interactions executed on Ethereum testnet during experiments.

(Full SAP follows. This is the unabridged statistical plan referenced in the paper. It includes detailed model specifications, power calculations with clustering adjustments, imputation strategy, R/Python code templates, and reporting guidelines.)

1. Primary inference strategy: Mixed-effects models to account for hierarchical data structure (participants nested in sessions).

2. Analysis principle: Intent-to-treat (ITT) primary analyses. Pre-registered covariates and models; any deviations documented on OSF.

3. Software: R (≥4.1) with lme4, lmerTest, mice, brms optional; Python alternatives acceptable. Randomization and imputation seeds set for reproducibility.

Session-level variables: session_id, condition (Daemon/Human), adversarial_injection (Y/N), explanation_granularity (summary/detailed), tokenization_scheme (fixed/adaptive), organization_id.

Participant-level variables: participant_id, age, gender, region, prior_dao_experience (0/1), baseline_trust, baseline_alignment (if available), participation_count, post_trust, governance_alignment_score, excluded_bot_flag.

alignment_ij ~ condition_j + adversarial_j + explanation_j + tokenization_j

               + baseline_alignment_i + prior_dao_experience_i + age_i

               + (1 | session_j) + (1 | participant_i)

Primary coefficient: condition_j (Daemon vs Human).

If alignment aggregated by session:

session_alignment_j ~ condition_j + adversarial_j + explanation_j + tokenization_j + (1 | organization_k)

gini_session_j ~ condition_j + adversarial_j + explanation_j + (1 | organization_k)

trust_change_ij ~ condition_j + baseline_trust_i + prior_dao_experience_i + age_i + (1 | session_j) + (1 | participant_i)

trust_change_ij = post_trust - pre_trust.

glmer(reversal ~ condition + explanation + adversarial + baseline_covariates + (1 | session) + (1 | participant), family = binomial)

Primary contrast: condition (Daemon vs Human). Secondary confirmatory interactions: condition × adversarial, condition × explanation. Within-Daemon contrasts: detailed vs summary explanations; fixed vs adaptive tokenization.

Pre-specified covariates: baseline_alignment (if available), baseline_trust, prior_dao_experience, age, region. Present both adjusted and unadjusted models.

Include random intercepts to account for clustering. Estimation of ICC will be reported; design effect DE used in power calculations.

Assumptions and arithmetic (reproduced here):

• Target effect: Cohen’s d = 0.50 (moderate).

• Two-tailed α = 0.05, power = 0.80.

• Z values: 1.96 (α), 0.84 (power) → sum = 2.80 → square = 7.84.

• Independent sample n per group = 2*(7.84)/0.25 = 62.72 → 63 per group.

• Adjust for clustering: example m = 20, ρ = 0.05 → DE = 1 + 19 × 0.05 = 1.95. Adjusted n ≈ 123 per group.

• Given planned N 100–200, mitigation strategies (smaller m, repeated measures) recommended.

Mitigation: reduce cluster sizes (m ≤ 12), include repeated measures, emphasize effect estimation, sensitivity analysis across ICC values.

Primary family: three outcomes (alignment, Gini, trust). Secondary outcomes corrected with Benjamini–Hochberg FDR (q = 0.10).

• Use MICE with m = 20 imputations; PMM for continuous variables.

• Diagnostics: assess MAR plausibility; run pattern-mixture sensitivity analyses and worst-case bounds.

• Exclusion of bot accounts as per pre-registered heuristics; report counts and include sensitivity analyses.

• Per-protocol analyses excluding attention-check failures.

• Heterogeneity analyses by prior DAO experience, region, baseline trust.

• Adversarial sensitivity curve: compute minimal coordinated fraction to change outcome.

• Alternate models: Bayesian brms, robust clustered SEs, permutation tests.

• Present coefficients, SEs, 95% CIs, p-values, R² (marginal/conditional).

• Upload analysis scripts, Docker container/renv snapshot, anonymized datasets to OSF after IRB/legal approval.

• Document any deviations from SAP on OSF.

(Include the MICE, lmer, glmer, and Gini computation templates adapted from the earlier SAP. Code skeletons are archived in the OSF repository.)

• Pre-registration: completed (DOI link).

• IRB submission: pending (before data collection).

• Primary analysis: within 2–4 weeks of data lock; supplemental analyses within 4–8 weeks.

• Data & code release: anonymized data and scripts to OSF after embargo or IRB approval.

• Vector store & embeddings: Supabase self-hosted vector store; embeddings computed locally by DeepSeek or a compatible embedding model per licensing.

• Orchestration: LangChain-style orchestrator patterns (retrieval, prompt templating, policy generation).
  Provenance: Retrieved items store metadata (source, time, policy tag, contributor ID) surfaced in explanations.
  Deployment policy: Local containerized model serving; network egress controlled; no external cloud inference without explicit agreements.

Author positionality: James Q. Marsh holds a BA in Psychology and works at the intersection of HCI, cyberculture, and behavioral patterns online. This manuscript presents an interdisciplinary project that blends technical design, social theory, and experimental methods. Personal reflections on identity and background have been deliberately limited in the main text and can be found in a supplementary author note if the journal requests positionality statements.