Continual & Nested Learning: Structural Stability Through Clarus
A coherence‑aware learning framework that detects and prevents catastrophic forgetting before it occurs.
Continual and nested learning systems remain vulnerable to catastrophic forgetting — the erosion of prior knowledge as new tasks overwrite shared representations. Existing methods such as EWC, SI, GEM, and replay strategies slow degradation but remain reactive: they repair after damage rather than preventing it. They cannot detect when internal stability begins to fail.
Clarus introduces a structural feedback layer that makes coherence measurable and controllable during training. It defines a live invariant K, the normalized reciprocal of feature‑space dispersion across sentinel layers and anchor exemplars. Two auxiliary metrics — alias (gradient interference) and drift (temporal centroid displacement) — expose coherence loss before accuracy declines.
Clarus operationalizes stability through a three‑layer stack:
External Monitoring for early drift detection,
In‑Loop Regulation that gates updates when K weakens,
Architectural Shaping that builds a high‑K spine and low‑K adapters for safe adaptation.
Modelled results across continual‑learning benchmarks suggest Clarus could reduce forgetting by 70–90%, lower gradient instability by 45–65%, and cut compute cost by 25–35%. Clarus transforms continual learning from reactive repair into proactive coherence management — enabling systems that learn longer, remember better, and train leaner.
Index
1. Abstract
Catastrophic forgetting, limitations of current methods, and Clarus as a structural stability layer.
2. Introduction
Why continual learning fails: representational drift, reactive mitigation, and blind spots in Nested Learning.
3. Core Claim — Nested Learning + Clarus = Structural Stability
NL provides modular adaptation; Clarus provides coherence awareness. Together they form a closed stability loop.
4. The Clarus Three‑Layer Fix
A full stability stack:
Layer 1: External Monitor — K, alias, drift, AK triggers
Layer 2: In‑Loop Regulator — learning‑rate gating, freeze policies, gradient routing
Layer 3: Architectural Shaping — high‑K spine, low‑K adapters, K‑aware routing, K‑rebasing
5. Measurement Protocol & Evaluation Framework
Definitions, proxies, thresholds, anchor sets, sentinel layers, benchmarks, baselines, ablations, pass/fail gates.
6. Modelled Results & Expected Impact
Projected improvements in retention, stability, collapse frequency, and compute cost.
7. Discussion & Implications
Stability–plasticity reframed, coherence as a system variable, regulation vs optimization, structural integrity as intelligence.
8. Compatibility
Integration with LLMs, vision models, RL policies, and control systems.