Skip to content

Latest commit

 

History

41 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

information-as-alignment

arXiv

code, configurations, and result artifacts for Information as Structural Alignment: A Dynamical Theory of Continual Learning.

OVERVIEW

this repository is the code and artifact companion to the Information as Alignment paper. it contains the toy model, the three validation domains, and the saved outputs needed to inspect, audit, or reproduce the reported results.

at the level of this release, this is not the full Informational Buildup Framework. it is the continual-learning instantiation developed in the paper: a first computational slice of a broader framework grounded in the premise that information is structural alignment rather than stored content.

the repository is organized as a research release, not as a production package. the fastest path is to inspect the saved artifacts first, then the notebook outputs, and only then rerun the experiments you actually care about.

REPOSITORY LAYOUT

.
├── README.md
├── REPRODUCIBILITY.md
├── (pre-print)information-as-alignment-v1.pdf
├── (seminar-slides)information-as-alignment.pdf
│
├── (IBF)Toy-Model.ipynb
├── (IBF)Domain-I-RRW.ipynb
├── (IBF)Domain-II-Chess.ipynb
├── (IBF)Domain-III-CIFAR-100.ipynb
│
├── RRW-paper-results.json
├── chess-results-seeds.json
├── chess-paper-results.json
└── CIFAR-paper-results.json

the four notebooks correspond to the toy model, RRW, chess, and CIFAR-100. the json artifacts contain the saved outputs used for the paper-facing comparisons, especially in chess.

WHAT IS IN HERE

the toy model reproduces the full mechanism lifecycle in two dimensions.

RRW reproduces the controlled mechanism-confirmation domain and the five-seed summary behind table 1.

chess reproduces the independent-oracle strategic domain, including the main comparison table, the readout sweep, seed replication artifacts, and the agency / Crucible diagnostics.

CIFAR-100 reproduces the high-dimensional continual-learning domain, including the main benchmark results, ablations, class-IL evaluation, and the weak-head analysis.

TRAINING ENVIRONMENT & RUNTIMES

all reported experiments were developed and executed on a rented cloud compute instance provided by RunPod. specifically, we provisioned a single-GPU pod running a standard PyTorch + JupyterLab template. the notebooks provided in this repository were executed directly within that Jupyter environment without requiring any custom containerization or complex cluster setups.

if you wish to perfectly replicate our environment on RunPod (or a similar cloud provider), the reference pod specifications were:

  • 1 × RTX 5090
  • 21 vCPU (AMD EPYC 9354 32-Core Processor)
  • 125 GB system memory
  • 30 GB container disk space

approximate wall-clock estimates on this reference hardware:

experiment estimate
toy model < 1 min
RRW ~30 min
chess ~50 h
CIFAR-100 ~75 h

DEPENDENCIES

recommended environment:

  • python 3.12+
  • jupyter notebook or jupyterlab
  • pytorch 2.x
  • torchvision
  • numpy
  • scipy
  • scikit-learn
  • matplotlib
  • python-chess

the chess notebook also requires stockfish 16.

install examples:

# ubuntu / debian
sudo apt install stockfish

# macOS
brew install stockfish

the CIFAR-100 dataset is handled through torchvision and can be downloaded automatically if needed.

for chess data, use the public lichess database:

https://database.lichess.org/

download a suitable PGN archive and point the notebook to your local file.

QUICK START

if you want the fastest understanding of the mechanism, start with:

  1. (IBF)Toy-Model.ipynb
  2. (IBF)Domain-I-RRW.ipynb

if you want the fastest audit of the paper, read REPRODUCIBILITY.md and inspect the saved artifacts before attempting any long reruns.

RESULT ARTIFACTS

  • RRW-paper-results.json: contains the 5-seed metrics (accuracy and backward transfer) used to generate the Phase A/B/C comparisons in Table 1.
  • chess-paper-results.json: the main chess artifact. it contains the outputs used for the main chess comparison table and the paper-facing centipawn advantages.
  • chess-results-seeds.json: contains the chess seed-replication artifact, including the seed-level behavioral and backward-transfer summaries.
  • CIFAR-paper-results.json: contains the 20-task benchmark metrics, ablation results, class-IL evaluation, and the weak-head analysis.

the notebooks themselves also contain embedded reported outputs. in many cases, direct inspection of the notebook output is enough to verify the relevant claim without needing to parse the json files.

PAPER

this repository accompanies:

Information as Structural Alignment: a Dynamical Theory of Continual Learning
Radu Negulescu
The Informational Buildup Foundation
april 2026

LICENSE

code in this repository is licensed under apache-2.0.

documentation, paper text, figures, and result artifacts are licensed under CC BY 4.0, unless otherwise noted.

About

Code, configurations, and result artifacts for Information as Structural Alignment - A Dynamical Theory of Continual Learning paper, including the toy model, RRW, chess, and CIFAR-100 experiments.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages