The surface
| Component | What it is | Status |
|---|---|---|
pallet-ai-registry, -oracles, -serving, -verification, -remuneration | Register models, request inference, submit verified results, pay servers | In runtime 220; no on-chain inference fulfilled yet on the public testnet |
pallet-ailoos-rewards | Per-era proportional payout for verified work | Funded (~300 M DMS); has paid nothing |
Precompile 0xAA | Solidity → AI inference | Reachable since runtime 211 |
| zkML toolchain | Groth16 on BLS12-381; aiServing.submit_inference_result verifies in-runtime | Benchmarked; 0.0035 DMS per verified result |
api.ailoos.com | Coordinator: sessions, validation, merge, OpenAI-compatible inference | Live |
| EmpoorioLM | 25M / 40M / 64M models, trained by the ecosystem | Served; small |
| SynoodoS | 421 datasets, sized | Live |
Ideas from the plan
- Inference with a Groth16 proof verified on chain: run a supported model, attach the proof, submit; the runtime verifies.
- A reputation system for Ailoos nodes from on-chain contribution records —
pallet-reputationexists as a base. - A dataset marketplace paid in testnet DMS with zkML integrity proofs, drawing from SynoodoS's catalogue.
What a team should know
- The models are small. A 64M-parameter model completes short text and code and follows simple instructions unevenly. Design for what a phone can train, not for a chatbot.
- The tokenizer is part of the model. Two of the ecosystem's tokenizers share 3.4 % of ids; decoding with the wrong one produces fluent nonsense with no error. Pair them correctly.
- The reward pool pays per era, proportional to relative units of work; there is no per-unit rate to game (the legacy per-unit path is deliberately zero).
- The Python Substrate library omits the chain's custom signed extension; use the Dart or Rust SDK, or Polkadot.js, to sign.
What has run
On 8 September 2026 a phone downloaded the 64M base via SynoodoS, trained 30 LoRA steps (loss 6.87 → 6.44), signed, uploaded, passed validation and had its adapter implemented into the production model. That is the loop. What has not run is an on-chain inference request fulfilled with a verified result on the public testnet — the first team to do it will have done something the core team has not.
The winners this page used to list
Belonged to another ecosystem's hackathon. This one has not happened; when it does, its winners will be listed with their transaction hashes.
Based on TESTNET_HACKATHON_PLAN.md (Track 2), the Ailoos production audit and the mobile verification of 2026-09-08.