We fine-tuned Laya on Stellar.
If you're using Laya for anything on Stellar, try this before the base model. Open weights, same license.
https://t.co/s7HV4I8k61
It's already the model running Radar: 92.6 operations arriving per second, 92.6 typed per second, 0 skipped. One machine.
Open it and ask it something strange. Ask whether an operation is a bot, a treasury move, a first-time user.
Then tell us what it gets wrong. That is the fastest way this gets better.
This is the decision engine coming to Sorolog: open source, Kubernetes operator, self-hosted.
https://t.co/HyFtnEirBf
The first thinking indexer.
Every operation on Stellar, typed as it lands. Trade, transfer, payroll, contract call, spam. Judged live as each one arrives, with a confidence next to it. 20M+ judged so far, 112 per second.
π§΅Thread:
The whole thing runs on one machine.
Laya, a 322M open model, keeping pace with the entire Stellar network at 85% agreement with the ledger. No inference API, no per-call billing. That is the only reason judging every single operation is affordable at all.
Next: keeping it. Stellar RPC holds roughly 7 days of events, so the network forgets itself. A typed, permanent history is what we are building underneath this.