spectralbranding/r15-synthetic-cohorts

---
license: cc-by-nc-nd-4.0
task_categories:
- text-generation
tags:
- doi:10.57967/hf/8441
- brand-perception
- synthetic-cohorts
- llm-evaluation
- spectral-brand-theory
- prism-b
- primacy-effect
pretty_name: R15 Synthetic Cohort Differentiation
size_categories:
- 1K<n<10K
configs:
- config_name: run15_main
data_files:
- split: train
path: data/run15_synthetic_cohorts.jsonl
- config_name: run15b_robustness
data_files:
- split: train
path: data/run15b_robustness_latin_square.jsonl
---

# R15 Synthetic Cohort Differentiation

Synthetic cohort differentiation experiment (Run 15 + Run 15b Latin-square robustness) from the Spectral Brand Theory research program.

## Dataset Description

- **Paper**: [Dimensional Collapse in AI-Mediated Brand Perception: Large Language Models as Metameric Observers (Zharnikov, 2026v)](https://doi.org/10.5281/zenodo.19422427)
- **Repository**: [sbt-papers/r15-ai-search-metamerism](https://github.com/spectralbranding/sbt-papers/tree/main/r15-ai-search-metamerism)
- **Point of Contact**: [e-mail]

## Dataset Summary

1,200 API calls (800 main + 400 robustness) testing whether the PRISM-B instrument differentiates synthetic observer cohorts defined by behavioral vignettes (no SBT dimension vocabulary in prompts).

- **Run 15 (main)**: 10 cohorts x 5 brands x 5 models x 3 repetitions = 800 calls
- **Run 15b (robustness)**: 8x8 Latin-square balanced dimension ordering x 10 cohorts x 5 brands x 1 model = 400 calls

### Key Findings

- **H1 SUPPORTED**: Cohorts produce significantly different spectral profiles (ANOVA p < .001 on 7/8 dimensions)
- **H2 SUPPORTED**: Trait-profile similarity predicts spectral-profile similarity (Mantel r = .43, p < .001)
- **H3 SUPPORTED**: Dimension-specific sensitivity confirmed (Economic eta-sq = .394, Ideological = .264)
- **Primacy effect discovered**: Serial position bias in JSON elicitation (eta-sq = .217). First-listed dimension gets ~15.4%, last gets ~8.0%. Cohort effects dominate on 5/8 dimensions.

### Models

Claude Haiku 4.5, GPT-4o-mini, Gemini 2.5 Flash, DeepSeek V3, Grok 4.1 Fast (xAI).

### Brands

Hermes, IKEA, Patagonia, Erewhon, Tesla (canonical SBT profiles).

## Data Files

- `data/run15_synthetic_cohorts.jsonl` — 800 records, main experiment
- `data/run15b_robustness_latin_square.jsonl` — 400 records, Latin-square robustness
- `prompts/` — cohort profiles, brand profiles, experiment configuration
- `protocol/` — pre-registered hypotheses and design
- `analysis/` — results JSON and human-readable summaries

## Citation

```bibtex
@article{zharnikov2026v,
author = {Zharnikov, Dmitry},
title = {Dimensional Collapse in AI-Mediated Brand Perception: Large Language Models as Metameric Observers},
year = {2026},
doi = {10.5281/zenodo.19422427}
}
```