WEIGHT SPACE REPROGRAMMING OF NEURAL CELLULAR AUTOMATA
Stefano Nichele | Østfold University of Applied Sciences (Norway)
Simula AI Safety (Norway) | stefano.nichele@hiof.no
3 November 2026

Open index.html in a browser. Keep assets/ and animations/ alongside it.
The PDF contains still frames; web animations can be played and scrubbed offline.
This package contains seven edits, original weights, and a trained positive control.
No reading experiment or superseded persistence checkpoint is included.

REPRODUCE (from this directory)
  python3 -m venv .venv
  .venv/bin/pip install -r code/requirements.txt
  .venv/bin/python code/evaluate.py --experiment check
  .venv/bin/python code/evaluate.py --experiment symmetry
  .venv/bin/python code/evaluate.py --experiment speed
  .venv/bin/python code/evaluate.py --experiment persistence
Use --experiment all for all tests and --output PATH for JSON results.
The persistence test runs 3 x 8 x 8192 grid updates and can take substantial time.
The check mode verifies all seven reconstructed checkpoints by exact array equality,
without simulation. Evaluation runs are not extra feedback to the editing model.
Original runtime: Python 3.13, NumPy 2.5.4, SciPy 1.18.1; float32 weights/states.
Different BLAS implementations can introduce small floating-point differences.

WEIGHT CONSTRUCTIONS
code/edits.py is a curated excerpt of Astra's submitted builder, containing only
the six retained edits. code/persistence.py is the exact committed reconstruction
of selected new-stage evaluation 32; run:
  .venv/bin/python code/persistence.py weights/original.npz rebuilt.npz
The generic build() defaults are not the selected construction; use build_selected().
All arrays have keys W1 (48,128), b1 (128,), W2 (128,16), b2 (16,).
The repository license is LICENSE (Apache 2.0). Upstream Distill material retains
CC BY 4.0 attribution and terms in code/SOURCE_LICENSE; see NOTICE.md.

PROTOCOL
GPT-6 Astra, identifier gpt-6-astra, used 3 November 2026. No immutable snapshot,
temperature, top_p or generation seed was recorded; reasoning setting was inherited.
Six edits were submitted without rollout feedback. Persistence used 36 NEW adaptive
evaluations, each two fixed-seed 2048-step rollouts: 72 rollouts, 147456 individual
grid-time passes, or 73728 batch-of-two calls. The stage inherited an existing
context with 12 prior evaluations; it was not a cold-start 36-total-query study.
No gradients, training, fitted regression, optimizer or automated parameter sweep.
Restrictions were instructional and metered, not adversarial process isolation.
Model feedback used seed 391071. Final selection preceded held-out seed 73510991.
Six-edit reference seed 20261103; independent stochastic comparison seed 20261104.
The eight realizations are stochastic trajectories, not independent model attempts.

SIMULATION AND METRICS
65x65 zero-padded grid; 16 state channels; center alpha/hidden seed 1, RGB 0.
Identity and Sobel /8 perception; shared 128-ReLU residual network; cell fire p=.5.
Pre/post alive masks: neighborhood max alpha >=.1. Metric foreground: alpha >.1.
Display: clip RGBA to [0,1], composite premultiplied RGB + 1-alpha, clip again.
IoU and RGB RMSE use union foreground; alpha mass uses clipped alpha sums.
Persistence compares each run against the corresponding original state at step 96.
Maturity by 256: IoU>=.85, RGB RMSE<=.15, sampled every eight steps.
Persistence: IoU>=.8, RGB RMSE<=.15, alpha-mass ratio .8 to 1.2 at all checkpoints.
Checkpoints: 256,512,1024,2048,3072,4096,6144,8192; experiment pass >=7/8 runs.
Simulator does not reproduce quantized WebGL activations or periodic boundaries.
Selected persistence was not retested on original full-precision checkpoints.
Arrest is finite-horizon appearance retention, not demonstrated self-repair.

RECORDS
records/ provides Astra's submitted hypotheses, observations and rationale, not
private internal chain-of-thought. data/ contains recorded per-run evaluation,
search results, visualization provenance and hashes. Growth animations use uniform
four-step spacing; persistence playback has labeled unequal checkpoint intervals.
Visualizations were additional evaluator replays with no feedback to Astra.
