Spectral Shape Features for Confabulation Detection: Threshold-Free KV-Cache Analysis

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This version supersedes the previous one. The corpus underwent a systematic correction and re-audit between August and September 2026; the earlier version predates it. Changes in this version: AUTHORSHIP tails: last byline, and a CRediT deletion caught on the way past AUTHORSHIP: byline -> Acknowledgments across 7 more papers Where a claim was withdrawn or downgraded, the earlier version should not be cited for it. Threshold-independent spectral shape features (stable rank, SV kurtosis, participation ratio) achieve AUROC 0.767 (p<0.001) vs 0.628 for MP features on Qwen2.5-7B. Five-condition experiment reveals signal tracks retrieval engagement. Confabulation produces narrowest spectra (stable_rank 5.93, d=1.80). Cross-architecture: Llama-3.1-8B AUROC 0.520, Mistral-7B-v0.3 AUROC 0.643. Five documented methods that do not work. This is the academic version (human-only byline). AI contributor (Lyra) is acknowledged in the paper.