Overview
DOTTIE-OS is a local-first System One sidecar from ARXIVIQ for tools and pair-programming. It takes state and typed questions and returns Choice, Score, or Noul outputs, with the probability named as a probability. It runs on the user's GPU with Laya as the local System One champion. The site states it is not a chat bot.
The product exposes three primitives. Choice returns one closed action for a pair-programmer step: execute, escalate, halt, or other. Score returns an ordered severity level for the step in front of the user. Noul returns P(safe), the probability the next action is safe to take. The sidecar picks a label instead of drafting a paragraph to parse.
Data handling is described as HF → factory → stamp. Inputs are public Hugging Face plus observed pair-programming traces, with no synthetic champion rows. The Choice/Score/Noul factory curates honest HELPER rows and Nimble contrastive flips (≤8 words) into staging JSONL. A fine-tune is followed by a held-out bench, and Cam stamps before anything is champion; nothing auto-promotes.
The site says this is not a chat model or a chat train control plane; the verb is decide, using calibrated probabilities rather than a transcript. It also lists refusals: synthetic rows do not train a champion, checkpoints do not auto-promote, and bots do not approve merge or ship. DOTTIE is on GitHub, and the starter is a solo personal project with no connection to employer, built with public/free-tier only, hosted on Vercel Hobby, and MIT licensed.
Pricing
Free
Detailed plans have not been confirmed in our catalog. Check the official website for current limits and billing terms.
Visit WebsitePrices and limits may change. Confirm the currency, billing period, seat minimum and usage caps on the official website.
Use Cases
- Developers use DOTTIE-OS to get local System One decisions for tools and pair-programming.
- Pair-programmers use the sidecar to pick a closed action label instead of parsing a paragraph.
- Users use Score to get an ordered severity level for the step in front of them.
- Users use Noul to get the probability the next action is safe to take.
- Operators run /decide on their tailnet.
- Builders use public Hugging Face plus observed pair-programming traces as inputs.
- Maintainers use a held-out bench to stamp before promote.
- Teams use DOTTIE-OS to keep bots from approving merge or ship.
Features
Personal AGI factory project
Sharded document corpus
Long-context language model experiments
Four-space J-Space routing
Six always-on scheduled jobs
Reviews
No reviews yet. Be the first to share your experience!