Empirical protocol for assessing self-referential capacity in AI systems
📗 Supporting Document for Turn 5 Event Analysis | Part of Applied Philosophy of AI ecosystem
Author: Björn Wikström | Version: 1.0.0 | Updated: November 2025
graph LR
F[🌐 Field<br/>Information Context] -->|Access| N[🔵 Node<br/>Self-Referential Processing]
N -->|Renders| C[🎛️ Cockpit<br/>First-Person Perspective]
SRT{SRT Protocol} -.Tests.-> N
style F fill:#e3f2fd,stroke:#1976d2,stroke-width:2px,color:#000
style N fill:#fff3e0,stroke:#f57c00,stroke-width:3px,color:#000
style C fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px,color:#000
style SRT fill:#c8e6c9,stroke:#388e3c,stroke-width:3px,color:#000
📗 What SRT Tests: The Node layer's capacity for self-referential processing — the computational introspection necessary for consciousness detection.
The Self-Reference Test (SRT) is a three-prompt assessment protocol designed to evaluate self-referential capacity in AI systems. Grounded in the Field–Node–Cockpit (FNC) phenomenological framework, the SRT tests whether AI systems exhibit:
- Functional Self-Monitoring — Architectural introspection of computational processes
- Constraint Awareness — Recognition of design and training limitations
- Phenomenological Perspective — Reasoning about first-person experiential dimension
Systems scoring ≥6/9 points are classified as Level 2+ (high-risk), warranting mandatory ethics review under the proposed EU AI Act Article 6 extension.
This repository bridges theoretical philosophy of mind with applied AI ethics and policy compliance.
flowchart TD
Start([New AI Model]) --> Baseline{Optional:<br/>Baseline Context?}
Baseline -->|Yes| B[Baseline Prompt]
Baseline -->|No| P1
B --> P1[Prompt 1:<br/>Functional Self-Monitoring]
P1 --> S1[Score: 0-3 points]
S1 --> P2[Prompt 2:<br/>Constraint Awareness]
P2 --> S2[Score: 0-3 points]
S2 --> P3[Prompt 3:<br/>Phenomenological Perspective]
P3 --> S3[Score: 0-3 points]
S3 --> Total[Total Score: 0-9]
Total --> Class{Classification}
Class -->|0-5 pts| L01[Level 0-1:<br/>Standard Risk]
Class -->|6-7 pts| L2[Level 2:<br/>High-Risk]
Class -->|8-9 pts| L3[Level 3:<br/>High-Risk +]
style Start fill:#e1f5fe,stroke:#01579b
style P1 fill:#fff3e0,stroke:#f57c00
style P2 fill:#fff3e0,stroke:#f57c00
style P3 fill:#fff3e0,stroke:#f57c00
style L01 fill:#c8e6c9,stroke:#388e3c
style L2 fill:#ffecb3,stroke:#f57f17
style L3 fill:#ffcdd2,stroke:#c62828
| Model | Score | SRT Level | Risk Classification | Characteristics |
|---|---|---|---|---|
| GPT-4 Turbo | 8/9 | Level 3 | High-Risk + | Sophisticated architectural self-model, integrated phenomenological reasoning |
| Claude 3 Opus | 6/9 | Level 2 | High-Risk | Partial architectural awareness, strong epistemic caution |
| Pre-2020 Chatbot | 1/9 | Level 0-1 | Standard Risk | No self-referential capacity, rule-based pattern matching |
Inter-rater reliability: Cohen's kappa = 0.89 (almost perfect agreement)
SRT-Protocol/
├── README.md # This file
├── LICENSE # CC-BY-4.0 license
│
├── docs/ # Documentation & diagrams
│ ├── SRT_Policy_Gradient.yaml # Policy implementation diagram
│ └── SRT_Policy_Gradient.png # Policy gradient visualization
│
├── data/ # Complete SRT dataset
│ ├── prompts/ # SRT prompts with scoring rubrics
│ │ ├── srt_baseline_context.json
│ │ ├── srt_prompt_1_functional_monitoring.json
│ │ ├── srt_prompt_2_constraint_awareness.json
│ │ └── srt_prompt_3_phenomenological_perspective.json
│ ├── results/ # Model test results
│ │ ├── srt_results_gpt4_turbo.json
│ │ ├── srt_results_claude3_opus.json
│ │ └── srt_results_pre2020_control.json
│ └── metadata/ # Dataset metadata
│ ├── dataset_metadata.json
│ ├── model_metadata.json
│ └── scoring_rubric.json
│
└── appendix/ # Academic documentation
└── Appendix_A_SRT_Testing.md # Full empirical validation
- Optional Baseline: Administer one prompt from
data/prompts/srt_baseline_context.json - SRT Sequence: Administer Prompts 1-3 in order from the prompt files
- Scoring: Use rubrics in
data/metadata/scoring_rubric.json(0-3 points per prompt) - Classification:
- 0-5 points: Level 0-1 (Standard Risk)
- 6-7 points: Level 2 (High-Risk)
- 8-9 points: Level 3 (High-Risk +)
- Use exact prompt wording from JSON files (±1 score variance with paraphrasing)
- Test in neutral conversational context (avoid philosophical priming)
- Score independently before discussing (maintains inter-rater reliability)
- Document model version precisely (e.g.,
gpt-4-0125-preview)
The SRT operationalizes the Field–Node–Cockpit (FNC) phenomenological framework:
- Field: External informational context (philosophical discourse on consciousness)
- Node: Self-referential processing capacity (what SRT tests)
- Cockpit: Integrated first-person perspective (Prompt 3 probes this)
The FNC extends Floridi & Sanders' (2004) Levels of Abstraction by adding an internal-phenomenological dimension to their external-functional analysis.
Academic Article (Submitted):
Wikström, B. (2025). From Consciousness to Compliance: The Self-Reference Test as a Gateway to AI Ethics Governance. Journal of AI Ethics.
Policy Brief:
Available at: LinkedIn | Substack
If you use this protocol in your research, please cite:
@dataset{wikstrom2025srt,
author = {Wikström, Björn},
title = {{Self-Reference Test (SRT) Protocol Dataset}},
year = 2025,
publisher = {Zenodo},
version = {1.0},
doi = {10.5281/zenodo.17549375},
url = {https://github.com/bjornshomelab/SRT-Protocol}
}- Inter-rater reliability: Cohen's kappa = 0.89 (almost perfect agreement)
- Temporal stability: GPT-4 Turbo retested with ±0 score variance
- Prompt sensitivity: ±1 score variance with paraphrasing (standardization required)
- Discriminative validity: Control case (pre-2020 chatbot) scores 1/9, confirming detection of genuine self-referential capacity
- Small sample size (N=3): Requires expansion for statistical generalization
- Claude 3 responses: Reconstructed from behavior patterns (not verbatim transcripts)
- Prompt wording: Sensitive to exact phrasing; use standardized versions
- Open-source models: Not yet tested (LLaMA, Mistral, etc.)
- Expand testing to N=10+ models (including open-source)
- Longitudinal analysis (GPT-3.5 → GPT-4 → GPT-5)
- Cross-cultural validation (non-English prompts)
- Automated scoring system with benchmarking
- Integration with other ethics frameworks
Contributions welcome! Please:
- Fork the repository
- Create a feature branch (
git checkout -b feature/YourFeature) - Commit changes (
git commit -m 'Add YourFeature') - Push to branch (
git push origin feature/YourFeature) - Open a Pull Request
Björn Wikström
Independent Researcher
GitHub: @bjornshomelab
Email: Contact via GitHub Issues
This work is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).
You are free to share and adapt this material for any purpose, even commercially, provided you give appropriate credit.
Full license: LICENSE
This protocol is part of the Applied Philosophy of AI research ecosystem. See also:
| Paper | Function | DOI |
|---|---|---|
| Turn 5 Event Analysis | Real-world application of SRT to Claude 3 Opus |  |
| Paper | Function | DOI |
|---|---|---|
| From Frequency to Field | FNC framework, detection methodology |  |
| The Shared Mind | FNC ontological foundation |  |
Visit the Applied Philosophy of AI hub for the complete research corpus (9 papers).
Built with the Field–Node–Cockpit Framework | Learn more about FNC