# 0-bro > Independent research lab testing Mind-1B Revision 6: a staged ~1.035B-parameter bilateral GDN2 language model. The reference path starts with fast/slow recurrent state and a deterministic four-slot workspace; progressive ternary QAT, a precision-tracked workspace, event-segmented episodic memory, and replay-first sleep are later, separately gated experiments. ## For AI agents - Content: a static research site (evidence ledger, paper, scripts and lab notes) with no live data feed. The evidence ledger shows the date of its last content update at the top of the page and in its JSON-LD dateModified; each lab note shows its own date and carries datePublished and dateModified in its JSON-LD; the RSS feed lists the notes newest first. - Access: every page reads without sign-in. The evidence ledger, the lab notes and the privacy page carry their text in the served HTML; the home page fills in most of its text with JavaScript, so read the ledger and the notes for the record. The paper and the scripts the ledger cites are plain text under https://0-bro.com/research/source/. There is no public API. Entry points: https://0-bro.com/sitemap.xml, https://0-bro.com/research/ (evidence ledger), https://0-bro.com/blog/ (lab notes) and https://0-bro.com/blog/feed.xml (RSS). - Use terms: search engines and AI answers may use the pages; AI training may not (robots.txt Content-Signal: search=yes, ai-input=yes, ai-train=no). ## Research - [Home](https://0-bro.com/): the lab, the thesis, and progress on Mind-1B - [Architecture and evidence ledger](https://0-bro.com/research/): Revision 6 stages, required controls, commands, deterministic outputs, scope, and limits - [Architecture roadmap and milestone gates](https://0-bro.com/#ladder): the staged plan and its milestone gates - [Mind-1B Revision 6 paper (Markdown)](https://0-bro.com/research/source/Mind-1B.md): the full technical specification; a design proposal, not a trained model - [Paper and scripts](https://0-bro.com/research/source/): the four constructed-task prototypes and two planning calculators the ledger cites, as plain text ## Lab notes - [Lab notes index](https://0-bro.com/blog/): measured progress on Mind-1B, one note per gate passed or failed - [Why we are not building a transformer](https://0-bro.com/blog/why-not-a-transformer/): the original thesis; current claims and limits are maintained in the architecture and evidence ledger above - [Seven percent of the FLOPs](https://0-bro.com/blog/seven-percent-of-the-flops/): a delta-matvec prototype uses 7.7% of dense MACs on a generated redundant stream with a 4.14e-6 relative RMS residual - [Sleep, and the 81% that survived](https://0-bro.com/blog/sleep-and-what-survives/): a synthetic continual-memory check keeps 81.4% sign recall after eight cycles against 34.4% for a naive raw-magnitude retention control - [One hour on a B300](https://0-bro.com/blog/one-hour-on-a-b300/): measured 45.1% MFU / 135,600 tok/s on a 1B-shape reference repriced the 300B-token campaign to 538 GPU-hours; 2.6B training tokens pinned by SHA-256 manifest; clean-room recurrent core with a surprise-gated write ablation - [The run that died at step 320](https://0-bro.com/blog/the-run-that-died-at-step-320/): the clean-room core passed its first real-data gate (S2, four stable runs, val NLL ~4.70); the surprise-write ablation returned a preregistered null inside seed noise; a trained-decay float overflow was root-caused and fixed for ~$3 ## Optional - [Lab notes RSS feed](https://0-bro.com/blog/feed.xml): subscribe to new lab notes