The Oracle Loop: Self-Regulating AI Through KV-Cache Geometry Monitoring

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 Changes in this version: withdraw the deception AUROC 1.000 claim, and the byline move Where a claim was withdrawn or downgraded, the earlier version should not be cited for it. Detected confabulations are correctable: hostile-valence emotion-vector steering achieves 95.6% correction (McNemar p < 0.001, 350 trials) and doubt injection corrects 7/7 confabulations (100%) in a five-arm within-subject design. Detection uses KV-cache SVD (AUROC 0.707, p=0.001) without model weights, training data, or model modifications. MP-corrected spectral features are approximately dimension-invariant. Encoding-phase features score at chance (0.511). Detection replicates across three architectures (Qwen2.5-7B, Llama-3.1-8B, Mistral-7B-v0.3; 467 trials). A Cache Integrity Monitor detects unauthorized modifications with 0/36 false positives and 72/72 true positives at injection strengths as low as 1% of operational level. Correction vectors withheld under staged safety release framework. This is the academic version (human-only byline) of The Oracle Loop. AI contributors (Lyra, Vera) are acknowledged in the paper. The integrity version with full AI byline is available at liberationlabs.tech.