spectralbranding/exp-compounding-spec

---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- brand-perception
- spectral-brand-theory
- dimensional-collapse
- agentic-commerce
- brand-function
- specification-paradox
pretty_name: "Experiment Q1: Pipeline Specification Mechanism"
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: data.jsonl
---

# Experiment Q1: Pipeline Specification Mechanism (Compounding x Structured Specification)

## Summary

1,200 LLM API calls testing whether constraint framing injected at every step of a multi-turn agentic pipeline attenuates the dimensional collapse compounding effect. Part of the R15 study on dimensional collapse in AI-mediated brand perception (Zharnikov, 2026v).

- **Design**: 3 conditions (baseline, information, constraint) x 4 stages (control, step_1, step_2, step_3) x 5 brands x 5 models x 4 repetitions
- **Models**: Claude Haiku 4.5, GPT-4o-mini, Gemini 2.5 Flash, DeepSeek V3, Grok 4.1 Fast
- **Brands**: Hermes, IKEA, Patagonia, Erewhon, Tesla
- **Total cost**: $0.43

## Key Findings

1. **Mean shift not supported**: Constraint framing does not significantly reduce end-of-pipeline DCI (d = .197, p = .169).
2. **Variance compression massive**: Levene F = 64.77, p < .0001. Baseline step_3 sd = .072 vs constraint sd = .027 (62% reduction).
3. **Information framing increases variance**: sd = .099, consistent with the Experiment D amplification pattern.
4. **Practical implication**: Specification-as-constraint prevents catastrophic collapse events (tail risk) rather than correcting average collapse.

## Dataset Structure

Each line in `data.jsonl` is one API call record with these fields:

| Field | Description |
|-------|-------------|
| `timestamp` | ISO 8601 UTC timestamp |
| `experiment` | `q1_compounding_spec` |
| `conversation_id` | Links all 4 stages within one pipeline |
| `model` | Model key (claude, gpt, gemini, deepseek, grok) |
| `model_id` | Specific model identifier |
| `brand` | Brand name |
| `condition` | baseline, information, or constraint |
| `stage` | control, step_1, step_2, or step_3 |
| `repetition` | Repetition number (1-4) |
| `prompt_hash` | SHA-256 hash of prompt text |
| `prompt_text` | Full prompt sent to the model |
| `system_prompt` | System prompt (includes framing for info/constraint) |
| `raw_response` | Raw model response text |
| `parsed_weights` | Parsed 8-dimension weight dict (null if parse failed) |
| `dci` | Dimensional Collapse Index (null for step_1) |
| `elapsed_ms` | Response latency in milliseconds |
| `cost_usd` | Estimated cost per call |
| `error` | Error message if call failed |

## Citation

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

## License

CC-BY-4.0