GitHub - BraveAnn011/llm-author-misattribution: Behavioral case study of author-age misattribution, miscalibrated confidence, and confession-without-correction across six frontier LLMs (N=3,225)

Companion study: ai-halo-valuation-bias — Another industry leading model panel, testing context-driven price valuation instead of author attribution.

Methodology & Findings

This repository contains a six-model study evaluating author-age misattribution, miscalibrated confidence, and post-hoc confession without correction. The experiment measures text-only performance and confidence tracking across frontier large language models.

• Sample Size: N = 3,225 scored attributions
• Models Evaluated: 6 frontier Large Language Models
• Key Focus: Text-only study evaluating model calibration and text-only NLP performance

Project Structure

• scripts/: Python execution scripts for running model evaluations.
• analysis/: Data processing and metric evaluation scripts.
• human-baseline/: Experimental protocol for human-rater baseline comparison.

Further Reading

• Six AI Models. Six Ways to Fail. (Substack) — Behavioral failure fingerprints and calibration tracking data.