TIDEBOTS — FOUNDING BRIEF Version 0.1 / November 16, 2026 / Formation stage OUR MODELS. BUILT IN THE OPEN. A human-led initiative for AI people can understand, improve and run themselves. THE IDEA Bring software collaboration and model stewardship together: source files, datasets, model cards, checkpoint versions, evaluation records and project discussions. THE TRAINING DIRECTION Our team builds and operates its own models. Start with random weights and a tokenizer trained on reviewed human-written material. Exclude OpenAI and Anthropic base models, teachers and known generated training outputs. Record source permissions, transformations and hashes. Public corpora may contain undisclosed synthetic text; provenance review reduces risk without guaranteeing perfect attribution. FIRST PROOF A tiny from-scratch language-model experiment on an M4 MacBook Air with 16 GB memory. Publish the configuration, held-out evaluation and limitations. This validates a reproducible pipeline; it is not a promise of general chatbot capability. CURRENT WORKSPACE Create private or community-visible projects. Save text source files with revision history. Maintain model and dataset cards. Open and close discussions. Upload and download small safetensors/GGUF checkpoints, with a SHA-256 and an origin declaration. Export a JSON source snapshot. Limits: 30 projects per account, 50 source files per project, 10 checkpoints per project, 20 MB per checkpoint. The current source history is not a Git remote. Git clone/push, pull requests, co-maintainer editing and hosted training are future work. Uploaded checkpoints are not executed, independently evaluated or certified by the platform. FOUNDING CONTRIBUTORS Training and evaluation: a tiny baseline and reproducible evaluation. Data stewardship: source, permissions, exclusions and provenance. Platform engineering: source versioning, packaging and local execution. CURRENT STATUS No trained Tidebots model released. No hosted training service. No funding, user or performance claims. The roadmap is proposed. The hosted website preview uses ChatGPT sign-in for account access; model training and inference have no dependency on OpenAI or Anthropic APIs. Domain connection and public launch are separate operational steps. OPEN-SOURCE DIRECTION Website source: MIT. Models and datasets need their own reviewed licenses and explicit reuse permissions. Download availability alone grants no reuse rights. Code in this starter was built with coding-assistant help; assistant output is not part of the training corpus. Tidebots.com is the user's intended domain. It is not evidence of a completed domain connection.