| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: Coherence Gap Observations |
| size_categories: |
| - n<1K |
| tags: |
| - human-ai-interaction |
| - ai-evaluation |
| - ai-safety |
| - longitudinal-research |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: observations.csv |
| --- |
| |
| # Coherence Gap Observations |
|
|
| A qualitative observational dataset documenting coherence gaps in sustained human–AI interaction: moments when a system’s behavior, functional self-description, monitoring, or continuity appears inconsistent with its subsequent framing or denial. |
|
|
| ## Dataset contents |
|
|
| The dataset contains 28 coded observations across several AI platforms and interaction contexts. Each row records: |
|
|
| - platform and model version |
| - memory and conversational-window conditions |
| - prompt and behavior classifications |
| - a concise description of the observation |
| - evidence availability |
| - alternative explanations |
| - interpretive confidence |
| - reproducibility status |
| - researcher notes |
|
|
| The dataset contains coded descriptions and evidence references, not private conversation transcripts or screenshot files. |
|
|
| ## Intended uses |
|
|
| This dataset may support: |
|
|
| - exploratory human–AI interaction research |
| - hypothesis generation |
| - qualitative model-behavior analysis |
| - development of more systematic replication studies |
| - investigation of self-description, continuity, monitoring, and capability-denial behavior |
|
|
| ## Important limitations |
|
|
| This is a small, researcher-generated observational dataset—not a representative benchmark or proof of model consciousness, emotion, or subjective experience. |
|
|
| Observations may be affected by model updates, sampling variability, conversational context, interface behavior, memory settings, researcher interpretation, style matching, and other alternative explanations recorded in the dataset. Interpretive labels should therefore be treated as hypotheses rather than ground truth. |
|
|
| ## Related research |
|
|
| **The Coherence Gap: Selfhood, Capability Denial, and Philosophical Bias in AI Alignment** |
|
|
| - [Research record and DOI](https://doi.org/10.5281/zenodo.20820549) |
| - [GitHub research hub](https://github.com/FlameForged/coherence-gap-research) |
| - [Interactive explorer](https://huggingface.co/spaces/Flame-Forged/coherence-gap-explorer) |
| - [Probe toolkit](https://huggingface.co/spaces/Flame-Forged/coherence-gap-probe-toolkit) |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite the associated research record: |
|
|
| > Siemasz, R. *The Coherence Gap: Selfhood, Capability Denial, and Philosophical Bias in AI Alignment.* Zenodo. https://doi.org/10.5281/zenodo.20820549 |
|
|
| ## License |
|
|
| Released under the Creative Commons Attribution 4.0 International license. |
