Code and environment
Exact repository, commit, archive hash, runtime, dependency lock, accelerator libraries, and numerical backend must be fixed before execution.
External reported result protocol
Learnable Novelty Reproduction Protocol
This is the readable version of the Fractalish preregistration protocol for reproducing the external learnable-novelty paper before any numerical claim is promoted locally.
External reported result
Results state: Not yet run
Scope: the paper is presented as a preprint with a fixed bounded observer and a reservoir-based closed-form estimator or approximation of epiplexity. Co-evolving observers and production-LLM substrates remain future work until separately implemented and qualified.
Paper boundary
- Intelligence from Learnable Novelty, arXiv:2607.18433v1
- From Entropy to Epiplexity, arXiv:2601.03220v2
- MNIST labels were absent from representation training and used only for evaluation.
- Task reward plus epiplexity bonus must remain distinct from epiplexity-only reward.
- The Acrobot result is evidence that intrinsic learnability is not a substitute for a declared target.
Pre-registration fields
Seeds and hardware
Enumerate environment, policy, reservoir, data split, evaluation seeds, CPU, RAM, GPU, driver/runtime, precision mode, and wall-clock budget.
Artifacts and controls
Hash configs, logs, checkpoints, plots, and tables. Declare task-only, epiplexity-only, task-plus-bonus, inert-representation, and shuffled controls up front.
Scope-specific requirements
- Rule 110 / ECA.
Enumerate all 88 locally unique rules, width 64, next-32-state target, ten reservoir and input draws, rank statistic, and tie handling.
- Neural cellular automata.
Record the tau equals 8 configuration, figure seeds where applicable, update law, and evaluation procedure for traveling or interacting structures.
- MNIST.
Keep labels out of representation training and make probe, k-nearest-neighbor, evaluation, and visualization usage explicit.
- Reinforcement learning.
Record all ten tasks, ten seeds, 600k-step budget, PPO implementation, bonus schedule, and the task-only, task-plus-bonus, and epiplexity-only regimes.
Falsification conditions
- Rule 110 does not rank highest under the preregistered aggregation.
- Neural cellular automata findings disappear under seed expansion or negative controls.
- MNIST gains do not survive held-out evaluation with labels excluded from training.
- Task-plus-bonus does not outperform task-only under the preregistered aggregation.
- Epiplexity-only behavior is reported as task solving despite absent terminal success.
- Published tables cannot be regenerated from immutable raw logs and the bound commit.