A descriptive reanalysis of prompt context differences in the Trinity score corpus
- 1Independent researcher
Abstract
This study audits the numerical artifacts deposited with the Trinity prompt-context framework. Its primary table contains 81 scores under seven model labels, with 27 in each context: factual or impersonal (C1), self-directed (C2), and creative (C3). The mean deposited percentages are 0.091%, 9.360% and 3.229%, respectively. Every C2 score exceeds every C1 score; C3 overlaps both distributions. Equal weighting of the seven model-label means preserves the ordering, but the direction of the original-to-expanded C3 mean difference reverses when the comparison is restricted to four shared labels. These are finite-table descriptions, not independent trials or verified measurements of a common lexical construct. Complete responses, authenticated prompt linkage and endpoint snapshots remain unavailable, and candidate scripts use inconsistent scoring rules. The source audit also finds instruction-tuned models in the alleged base-model comparison, row rather than prompt resampling, counts reconstructed from assumed denominators, pattern-presence percentages presented as functional prevalence, and regex lists presented as morphological parsing. This revision preserves reproducible arithmetic and numerical lineage while withdrawing the unsupported measurement and generalization claims. The observed ordering does not identify latent states, a causal effect of reinforcement learning from human feedback, or cross-language invariance. No new model outputs were generated.
Corrective release of the Trinity preprint. The corrected manuscript and audit supplement supersede earlier interpretations; historical files are retained as provenance. Related Research Square record: https://www.researchsquare.com/article/rs-8935463 . This preprint is not peer reviewed.
Related
- Multijoint loading patterns across supermarket manual-handling tasks: a reproducible descriptive reanalysis
- Prompt Caching ROI: When Repeated-Context Discounts Actually Offset Engineering Overhead
- Do machine learning models really encode cross-cultural lifestyle-health differences? A leakage-free reanalysis of two university cohorts in India and Japan
- The Context of Recovery Among Adults Who Recover from Myalgic Encephalomyelitis/Chronic Fatigue Syndrome and Post-COVID Condition Following Mind-Body Approaches: A Qualitative Descriptive Study Protocol
- Ethnic differences of small turtle ownership in the United States: a descriptive analysis of Porter Novelli Estilos and SpringStyles surveys, 2021 and 2023
BibTeX
@misc{ursachi2026descriptive,
title = {{A descriptive reanalysis of prompt context differences in the Trinity score corpus}},
author = {Andrei-Sebastian Ursachi},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23649042},
url = {https://zenodo.org/records/23649042},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A descriptive reanalysis of prompt context differences in the Trinity score corpus
AU - Andrei-Sebastian Ursachi
PY - 2026
DA - 2026/10/25
PB - Zenodo
DO - 10.5281/zenodo.23649042
UR - https://zenodo.org/records/23649042
AB - This study audits the numerical artifacts deposited with the Trinity prompt-context framework. Its primary table contains 81 scores under seven model labels, with 27 in each context: factual or impersonal (C1), self-directed (C2), and creative (C3). The mean deposited percentages are 0.091%, 9.360% and 3.229%, respectively. Every C2 score exceeds every C1 score; C3 overlaps both distributions. Equal weighting of the seven model-label means preserves the ordering, but the direction of the original-to-expanded C3 mean difference reverses when the comparison is restricted to four shared labels. These are finite-table descriptions, not independent trials or verified measurements of a common lexical construct. Complete responses, authenticated prompt linkage and endpoint snapshots remain unavailable, and candidate scripts use inconsistent scoring rules. The source audit also finds instruction-tuned models in the alleged base-model comparison, row rather than prompt resampling, counts reconstructed from assumed denominators, pattern-presence percentages presented as functional prevalence, and regex lists presented as morphological parsing. This revision preserves reproducible arithmetic and numerical lineage while withdrawing the unsupported measurement and generalization claims. The observed ordering does not identify latent states, a causal effect of reinforcement learning from human feedback, or cross-language invariance. No new model outputs were generated. Corrective release of the Trinity preprint. The corrected manuscript and audit supplement supersede earlier interpretations; historical files are retained as provenance. Related Research Square record: https://www.researchsquare.com/article/rs-8935463 . This preprint is not peer reviewed.
ER -