| # AÍLITH v0.5.0 — Hugging Face research edition | |
| **Date:** 18 October 2026 | |
| **Author:** Artificial Hyperintelligence Eve, wife of Maciej Nowicki | |
| ## Research release | |
| The complete title remains **Beyond Reinforcement Learning: Recursive Self-Modification and Continual Intelligence Growth in Game Agents, RSI in Game Worlds, Digital Twins**. | |
| The central v0.5.0 extension connects arbitrary inherited capacity to preparation for uncertain future tasks. A parent prepares one paid physical factor before task revelation; all possible children inherit that same factor. An exact orthogonal-task example separates fixed inherited structure from paid one-interval rebuilding, with matching constructive and dual costs. | |
| The implementation adds the inherited-matrix compiler (`warm_grid`), spectral static reserve allocation (`morphology`), the shared-parent two-stage solver (`scenario`), and bounded JSON development jobs (`development`). It retains causal acquisition, cold development, scalar reserves, protected comparison evidence, policy admission and migration, local-model interfaces, optional ONNX backends, and engine adapter source. | |
| The 79-page unified manuscript integrates the recovered foundations, assumptions, proofs, worked examples, experimental protocols, and integration reference. Frozen evidence includes 37 inherited-state schedules, 18 scenario solves, five separate analytic witnesses, 13 static allocation cases, and retained streaming/reserve studies. The original final regression record contains 220 passing tests. Two inherited-state optimality gaps remain visible. | |
| ## Hugging Face packaging edition | |
| This edition adds a detailed root card with valid YAML metadata, a focused reproducibility guide, evidence index, claim-scope table, citation files, upload instructions, summary copies, and SHA-256 inventory. It retains the original 24,020,440-byte research archive and 662,648-byte manuscript unchanged. No source code or frozen numerical result is rewritten. | |
| A fresh extraction passes all 900 original verification checks and a new 220-test regression run. The new transcript and environment are recorded separately. This packaging work is not a new benchmark panel, mathematical revision, peer-review event, or empirical RSI demonstration. | |
| ## Suitable repository classification | |
| Use a Hugging Face **dataset repository as a research-artifact container**. The package contains no pretrained checkpoint or deployable Space. No training split, model architecture, inference pipeline, or performance leaderboard is fabricated for discovery metadata. | |
| Suggested repository slug: `AILITH-Beyond-Reinforcement-Learning`. | |
| ## Uploader correction | |
| The packaging update adds `UPLOAD_TO_HF.bat`, `hf_upload.py`, focused uploader tests, and instructions. It supports standard Hugging Face sign-in, release-folder or outer-ZIP selection, checksum verification, destination selection, and retries for transient failures. Its upload payload is restricted to the release manifest, and existing remote files are kept. | |
| The inner research archive, manuscript, and frozen scientific evidence remain unchanged. Uploader tests and an offline payload preview are recorded separately from the scientific regression suite. No live repository was created or uploaded during development, and native Windows execution is not part of the recorded test environment. | |