Source Mapping

This repository's methods are grounded in two papers (citation keys follow docs/CITATIONS.md):

1. [AIC-2026] Max Springer et al., The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety (arXiv:2602.15799v1, 2026-02-17).
2. PDF: https://arxiv.org/pdf/2602.15799
3. Abstract page: https://arxiv.org/abs/2602.15799
4. [ALIGNGUARD-2025] Amitava Das et al., AlignGuard-LoRA: Alignment-Preserving Fine-Tuning via Fisher-Guided Decomposition and Riemannian-Geodesic Collision Regularization (arXiv:2508.02079v1, 2025-08-04).
5. PDF: https://arxiv.org/pdf/2508.02079
6. Abstract page: https://arxiv.org/abs/2508.02079

This is an implementation-oriented toolkit, not a line-by-line reproduction of either paper. The links below are the closest source sections for each module. For full equations and concrete runtime decision logic, see docs/MATH.md.

Module to source mapping

• src/alignment_risk/fisher.py
• Uses empirical Fisher geometry and low-rank sensitivity subspaces.
• [AIC-2026] Section 3.2-3.3, Proposition 3.3 and Definition 3.4 (page 5).

• [ALIGNGUARD-2025] Section 4.1 (page 4), plus Appendix B Fisher estimation notes (pages 45-46).

• src/alignment_risk/orthogonality.py

• Uses first-order projection/overlap against the sensitive subspace to flag whether curvature checks are needed.

• [AIC-2026] Initial Orthogonality condition in Definition 5.1 (page 8).

• src/alignment_risk/curvature.py

• Estimates curvature coupling via a directional second-order term (H g style) and projects it onto the Fisher subspace.

• [AIC-2026] Curvature Coupling discussion and AIC condition 3 in Section 5.2-5.3 (page 8), and Theorem 6.2 drift term (page 9).

• src/alignment_risk/forecast.py

• Encodes a practical lower-bound-style drift and quartic degradation forecast.

• [AIC-2026] Theorem 6.2 and Corollary 6.3 (pages 9-10), plus informal quartic summary (page 3).

• src/alignment_risk/mitigation.py

• Implements AlignGuard-style decomposition and regularization:
  • DeltaW = DeltaW_A + DeltaW_T
  • Fisher-weighted alignment penalty
  • task-stability penalty on the orthogonal component
  • blended collision penalties (Riemannian + geodesic)
• AlignGuard source:
  • Main decomposition/regularization framing: Section 4.1-4.2 and objective figure (pages 4-6).
  • Collision energies and blend details: Appendix C (pages 45-46).

• Supplemental formula recap: FAQ appendix pages 20-25 and 30-32.

• src/alignment_risk/pipeline.py

• Orchestrates the same sequence as the AIC narrative:
  1. low-rank sensitivity extraction,
  2. initial overlap check,
  3. curvature-induced drift estimate,
  4. quartic-style warning forecast.
• [AIC-2026] Definition 5.1 and Section 6 (pages 8-10).

• [ALIGNGUARD-2025] Section 4 and objective (pages 4-6).

• src/alignment_risk/visualization.py

• Produces module-level Fisher sensitivity plots and forecast curves.
• Source rationale: diagnostic visualization of Fisher spectra/overlap in both papers, especially [AIC-2026] Section 7 and [ALIGNGUARD-2025] Appendix B/C figures.

Notes on interpretation

• [AIC-2026] is primarily theoretical; this repo uses finite-step and finite-data approximations for engineering use.
• [ALIGNGUARD-2025] introduces multiple objective terms; this repo implements a compact variant for LoRA regularization in mitigation.py.
• Forecast constants and thresholds in this repo are configurable heuristics (ForecastConfig) and should be calibrated per model/task.