Ailoos is the reason EmpoorioChain has an AI pallet group. It is a network that trains the ecosystem's own models on hardware its users already own — phones first — and pays for verified work. This post is its honest state as of September 2026.
The pieces
- Coordinator at
api.ailoos.com: assigns training sessions, validates and merges adapters (FedAvg), serves inference (OpenAI-compatible), exposes the dataset catalogue. Health check returnshealthy. - Node software: a Python SDK and CLI with 107 commands in 17 groups, generated from live introspection so the reference cannot drift. Seven node roles (trainer, validator, oracle, archival, relay, …) and five network scales.
- Ailoos App (Android, internal testing): the phone as a training node, with a PII firewall and on-device LoRA training via ExecuTorch.
- SynoodoS (
synoodos.com): the dataset and model catalogue — 421 datasets, every one measured by parquet size so a node can filter by what fits its memory. - Ailoos Web (
ailoos.com): chat against the coordinator's models. - On chain:
pallet-ailoos-rewards(index 31) with ~300 M DMS; per-era proportional payout.
What the audit found
Eight agents audited the production system on 5 September 2026. The diagnosis: Ailoos does not have an implementation problem; it has a connection problem. Almost everything important existed and worked in isolation; what failed was the line joining it to the production path.
- The seven roles were a label. The role loops existed entirely and were never started by the headless node command, while the status command told operators they ran every 60 seconds. Fixed.
- The training loop did not close. The merged adapter was written without its configuration file, and the loader accepted only an older contract than Android produced. The adapter the whole network produced was unloadable — and rewards were computed anyway. Fixed, with an end-to-end test.
- The reward pool had never paid. The pallet was deployed and funded; nobody had authorised an operator; the configured signer had no funds. Identified as an operational step.
- A future leak in attention. The causal mask construction in the model was commented out. Fixed and measured before/after.
What closed the loop
On 8 September, on a physical phone: download the 64M model via SynoodoS (258 MB) → resume from the previous adapter (80 tensors) → 30 real LoRA steps, loss 6.87 → 6.44 → ED25519 signature → upload → validation → implemented into the production model. The only open item: the reward transaction is held for manual audit because the bridge URL in strict mode points at localhost — the known block from the audit, not a regression.
Also found that day: the production mobile-training artefact had contained random weights since export — every mobile session had fine-tuned on noise. The fixed exporter verifies real weights are inside the file before writing it.
Infrastructure
A dedicated server (ailoos.com, api.ailoos.com) runs the coordinator in Docker and a quantized general model outside it. It has had a cryptominer removed (July) and a rootkit found (September); the incident playbook exists because of that.
What Ailoos is not
It is not a wrapper around a commercial API. It does not promise an APY — a Play listing that said "5–15 %" is being corrected. And it has not yet paid its first DMS reward on the public testnet.
Based on the Ailoos production audit (2026-09-05), the mobile-training verification (2026-09-08), the 421-dataset catalogue and CLI_COMMANDS_REFERENCE.md.