Dual-Gradient Learning Series

# Release v2.0.0 This release marks a **major milestone** for the Dual‑Gradient Learning repository.Paper A has reached a **submission‑ready state** for *Nature Machine Intelligence*,with a fully audited experimental ledger, finalized figures, and a consolidated reproducibility package. --- ## Highlights ### Paper A — Finalized Core Contribution- **Title**: *Designing Robust Learning as Phase Control: Identical Training Produces Stable Success or Irreversible Failure*- All experiments, figures, and statistics are now **internally consistent and frozen**- Results validated across **vision and NLP**, with explicit domain‑specific limits ### Regime‑Structured Robust Learning- Sharp boundary and variance peak in λ- Bimodality, hysteresis (22–53 pp), and metastability in vision- Absorbing high‑error failure states at high noise- Cross‑domain validation in NLP with causal decomposition: - time‑driven corrosion - path‑dependent damage --- ## What’s New in v2.0.0 ### Paper A: Experimental Ledger v4.1- Introduced **run‑level accounting** (Total: 3,916 runs)- Unified branch definitions (L‑branch, H‑branch, T‑state)- Corrected and audited all noise‑rate and geometry metrics- Removed invalid cross‑condition comparisons- Explicitly documented domain‑dependent phenomenology ### Figures- Added **final Main Figures (Fig. 1–5)** for Paper A- Added **Extended Data Figures (ED Fig. 1–8)**- All figures are publication‑ready (PDF) and numerically reproducible ### Repository Structure- Standardized Paper‑A layout: - notebooks / results / figures (main + extended_data)- Top‑level README updated to reflect submission‑ready status --- ## Backward Compatibility - Papers B–E remain unchanged- Existing DOIs remain valid- No breaking changes to earlier notebooks --- ## Next Steps - Zenodo DOI assignment for Paper A (final snapshot)- Submission to *Nature Machine Intelligence*- Subsequent public release of the final DOI‑linked archive --- **Version**: v2.0.0 **Date**: March 2026