adaptergate
Copyright (c) 2026
Licensed under the Apache License, Version 2.0.

This product builds upon the following research and software. All upstream
work is credited to its authors; this project's contribution is its
original implementation, integration, and the regression-gating system.

============================================================================
PAPERS (algorithmic foundations)
============================================================================

ProCL — arXiv 2605.13162
  "Continual Fine-Tuning of LLMs via Program Memory"
  Used in: src/adaptergate/core/procl.py
  Our contribution: independent re-implementation from the paper, bug fixes
  (composition cross-terms, dtype handling, autograd reset), production
  integration. The algorithm is credited to the paper authors.

Silent Collapse — arXiv 2605.14588
  "Silent Collapse in Recursive Learning Systems"
  Used in: src/adaptergate/observability/silent_collapse.py
  Our contribution: independent re-implementation of the MTR framework.
  The drift signature (H_g/H_0 < 0.5 with PPL_g/PPL_0 ≈ 1) is credited
  to the paper authors.

Online-LoRA — arXiv 2411.05663
  "Online-LoRA: Task-free Online Continual Learning via LoRA"
  Concepts of online distribution-shift detection informed our
  regression-gating design.

PreFT — arXiv 2605.14217
  "Prefill-only Finetuning for Efficient Inference"
  Referenced as a future integration once code is publicly released
  by the authors. Not currently integrated.

N-LoRA / O-LoRA — arXiv 2408.06133, arXiv 2310.14152
  Orthogonal LoRA subspace work, referenced in the regression-gating
  literature review.

============================================================================
SOFTWARE
============================================================================

vLLM — https://github.com/vllm-project/vllm — Apache License 2.0
  Used as serving backbone via its public Python API.
  No source modifications are distributed in this repository.

bitsandbytes — Apache License 2.0
  Used for 4-bit quantization.

transformers, peft, accelerate (HuggingFace) — Apache License 2.0
torch — BSD-3-Clause
