dataset_name	pretty_name	license	language	tags	task_categories	papers	repository
s64-geometry-v1	S64 Geometry Validation – The Conversational Coherence Region	cc-by-4.0	en	symbolic-ai human-ai-interaction embedding-geometry semantic-space conversation-dynamics multi-model	other	title url doi The Conversational Coherence Region: Geometry of Symbolic Meaning Across Embedding Models https://www.aicoevolution.com/s64-geometry-paper 10.5281/zenodo.18149380	https://github.com/AICoevolution/paper02-coherence-region

S64 Geometry Validation Dataset

Paper 02: The Conversational Coherence Region: Geometry of Symbolic Meaning Across Embedding Models

This dataset accompanies the paper investigating how the S64 symbolic framework organizes semantically across 13 different embedding architectures, and how conversation dynamics reveal structured regions in semantic space.

---

Key Findings

Finding	Evidence
Architecture-Independent Structure	Role centroids (from→to, through→result) show consistent angular relationships across all 13 backends (p < 0.001)
Coherence Region	Structured conversations occupy a distinct dynamical region: lower velocity, higher SGI, tighter symbol clustering
Negative Cone-Diversity Correlation	r = −0.88 [95% CI: −0.98, −0.55] between trajectory cone-ness and symbol diversity
Large Effect Sizes	Cohen's d = 1.15 for velocity differences between structured and unstructured conversations

---

Repository Structure

s64-geometry-v1/
│
├── README.md                           # This file
│
├── sweep/                              # Multi-backend symbol geometry analysis
│   ├── sweep_summary.json              # Summary: isotropy, clusters, role p-values (all backends)
│   ├── symbols.json                    # 180 S64/S128 symbol definitions with roles
│   ├── rosetta_dataset.json            # Consolidated 3D PCA positions for visualization
│   ├── run_meta.json                   # Sweep configuration and timestamp
│   │
│   └── per_backend/                    # Per-backend analysis (13 backends)
│       └── {backend}/
│           ├── embedding_meta.json     # Backend name, dimension, symbol count
│           └── symbol_geometry_analysis.json  # Angular stats, clusters, role geometry
│
├── conversations/                      # Conversation trajectory analysis
│   ├── traces_metrics.csv              # Primary data: 1 row per (conv × backend × mode)
│   ├── traces_metrics.json             # Same as above, JSON format
│   ├── cone_explanations.json          # Correlations: cone-ness vs diversity/entropy
│   ├── cross_model_agreement.json      # Jaccard overlap of top-k symbols across backends
│   └── run_meta.json                   # Analysis metadata
│   # Note: rosetta_conversations_pack.json (~100MB) available on request
│
├── dynamics/                           # Turn-level dynamics data
│   ├── manifold_dynamics.csv           # SGI and Velocity per turn, per conversation
│   └── statistical_analysis_results.json  # Bootstrap CIs, permutation tests, effect sizes
│
├── visualizations/                     # Interactive visualizations
│   ├── symbol_geometry_rosetta.html    # Multi-model 3D comparison viewer
│   └── alignment_hypersphere.html      # Conversation trajectory on hypersphere
│
└── scripts/                            # Analysis scripts (Python)
    ├── analyze_symbol_geometry.py      # Per-backend geometry analysis
    ├── analyze_conversation_geometry_pack.py  # Produces traces_metrics, cone_explanations
    ├── compute_statistical_tests.py    # Bootstrap, Mann-Whitney, permutation tests
    └── export_manifold_dynamics_csv.py # Exports SGI/velocity time series

---

Backends Analyzed

Backend	Dimension	Provider
bge-m3	1024	BAAI
cohere-v3	1024	Cohere
e5-finetuned-v6	768	Custom fine-tuned
google	768	Google
jina-v3	1024	Jina AI
mistral-embed	1024	Mistral AI
nomic	768	Nomic AI
openai-3-large	3072	OpenAI
openai-3-small	1536	OpenAI
openai-ada-002	1536	OpenAI
qwen	1024	Alibaba
s128	768	Custom (S128)
voyage-large-2-instruct	1024	Voyage AI

---

Conversations Analyzed

11 baseline conversations spanning structured therapeutic dialogue to unstructured exploration:

ID	Type	Description
B01–B08	Synthetic	Controlled baselines with ground-truth transformations
B09	Naturalistic	Self-discovery dialogue (structured)
B10	Naturalistic	AI interaction baseline
LC1	Extended	Free-form exploratory conversation

---

Key Metrics Reference

Metric	File	Description
highd_R	traces_metrics.csv	768D cone-ness: mean resultant length (0 = spread, 1 = tight)
topk_unique_symbols	traces_metrics.csv	Number of unique symbols activated across conversation
topk_entropy_bits	traces_metrics.csv	Shannon entropy of symbol usage
step_angle_mean	traces_metrics.csv	Mean angular distance between consecutive turns
isotropy_score	sweep_summary.json	How uniformly symbols are distributed (per backend)
role_perm_p_close	sweep_summary.json	P-value: role centroids unusually close
role_perm_p_far	sweep_summary.json	P-value: role centroids unusually far
velocity	manifold_dynamics.csv	Angular velocity per turn (degrees)
sgi	manifold_dynamics.csv	Semantic Grounding Index per turn

---

Quick Start

Load conversation metrics (Python)

import pandas as pd

# Load primary analysis data
df = pd.read_csv("conversations/traces_metrics.csv")

# Filter to centered mode (recommended)
df_centered = df[df["mode"] == "centered"]

# Correlation: cone-ness vs symbol diversity
print(df_centered[["highd_R", "topk_unique_symbols"]].corr())

Load symbol geometry (Python)

import json

# Load sweep summary
with open("sweep/sweep_summary.json") as f:
    summary = json.load(f)

# Check role geometry p-values across backends
for backend, data in summary["backends"].items():
    role_geo = data.get("role_geometry", {})
    p_close = role_geo.get("pvalue_unusually_close", {}).get("from-to")
    print(f"{backend}: from→to p = {p_close}")

Interactive visualization

Open visualizations/symbol_geometry_rosetta.html in a browser to explore:

• 3D symbol lattice across all backends
• Role clustering visualization
• Conversation trajectory animation

---

Statistical Methods

All statistical tests follow conservative, exploratory protocols:

• Bootstrap CIs: 10,000 resamples for correlation confidence intervals
• Permutation tests: 2,000 permutations for role geometry significance
• Mann-Whitney U: Non-parametric comparison of velocity distributions
• Cohen's d: Effect size for group comparisons
• Stationarity testing: Pearson correlation of velocity vs turn index

Results are reported with explicit sample-size caveats (N = 11 conversations).

---

Citation

If you use this dataset, please cite:

@misc{jimenez2026coherenceregion,
  author = {Jiménez Sánchez, Juan Jacobo},
  title = {The Conversational Coherence Region: Geometry of Symbolic Meaning Across Embedding Models},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18149380},
  url = {https://doi.org/10.5281/zenodo.18149380}
}

Also cite the Semantic Grounding Index work:

• Marín, J. (2024). Semantic Grounding Index for Large Language Models. arXiv:2512.13771

---

Related Work

• Paper 01: S64: A Symbolic Framework for Human-AI Meaning Negotiation
• S64 Dataset: s64-validation-v4

---

Excluded Files

The following files are excluded from this repository due to HuggingFace size/binary limits:

• rosetta_conversations_pack.json (~100MB): Full embedding traces for all conversations
• figures/*.png: Paper figures (available in the paper PDF)
• paper02-*.pdf: Paper PDF (available at aicoevolution.com/s64-geometry-paper)

To request the complete dataset including large files, contact [email protected].

---

License

This dataset is released under CC BY 4.0.

Contact: Juan Jacobo Jiménez Sánchez — AICoevolution