Mehmet Arda
Hakbilen

My main focus is trainscope: a post-mortem flight recorder for LLM training loss spikes. Its CUSUM change-point detector catches gradual loss drift 5-20 steps before it turns into a full spike, validated with 0.0% false positives across 280+ noise scenarios and 100% sensitivity for 0.10σ-0.50σ drift. When failure occurs, it reconstructs which layer moved first, whether from activation kurtosis or gradient explosion.

Where trainscope watches training runs, CERATA asks what a frozen model can still learn after deployment. Learning is an API call: a FAST key-value write for a single item, a MEDIUM closed-form edit from float64 sufficient statistics, and a SLOW consolidation into frozen experts. Every write and forget is guarded: locality, reversibility, order invariance, zero-parameter routing. Forget is an exact downdate. On frozen ViT-B/16 features (CIFAR-100, 20 tasks), its closed-form readout beats the v1 expert bank with zero trained parameters: 76.77 vs 59.30.

Earlier, I closed out NOESIS, testing if Titans-style associative memory matrices could inject continual learning directly into a frozen LLM manifold. The result was negative: frozen manifolds cannot assimilate untrained vector injections without distortion (0% recall on a 50-fact benchmark; full analysis in FINDINGS.md).