"AI on blockchain" is usually a slogan in search of a mechanism. Ailoos has three mechanisms that need a chain, and this post argues from them rather than from the slogan.
1. Provenance: which data trained which model
Every EmpoorioLM model has a provenance manifest recording which of the catalogued datasets went into it. The catalogue (SynoodoS, 421 datasets) is measured and versioned; the manifest is generated by the training pipeline, not written afterwards. Anchoring that manifest — and the hash of each adapter merged into a model — in pallet-ai-registry gives a verifiable answer to the question every AI regulator is starting to ask: what is this model made of? Today the manifest exists off chain; the registry pallet exists on chain; the anchoring is the next step.
2. Verification: did this node do the work?
A phone uploads an adapter signed with its ED25519 key, bound to a session by an upload secret. The coordinator validates geometry and loss trajectory and merges by FedAvg. For inference, a serving node can attach a Groth16 proof that aiServing.submit_inference_result verifies in the runtime (0.0035 DMS on the testnet). Without a chain, verification is a promise by the coordinator; with one, the proof is checkable by anyone.
3. Payment: proportional, per era, from a pool nobody can drain
pallet-ailoos-rewards holds ~300 M DMS. Each 24-hour era, the runtime freezes a budget (a release rate over the pool's balance) and pays each node budget × my_units / total_units. Work is measured in relative units, so inflating one's own claimed rate changes nothing — which is what made the ecosystem's four incompatible reward-rate tables unnecessary. The emission is set by the clock, not by demand: train twice as much and the pool pays the same total, split differently. Twenty-five percent of the chain's own emission flows into this pool, forever, under gap-decay.
A per-unit legacy path exists in the pallet and is intentionally zero; re-enabling it would revive a design that was retired. The audit's first draft recommended re-enabling it and was corrected.
Where the chain has not earned it yet
- The pool has distributed nothing. Operator authorisation was never performed. The first payout is an operational step away.
- Provenance manifests are not yet anchored.
- No on-chain inference has been fulfilled end to end on the public testnet.
- The Python signing library omitted the chain's custom signed extension; being fixed.
Robots
The previous version of this page asked whether robotics would be blockchain's breakout AI use case. The ecosystem has no robotics product and does not pretend to. It has phones — and a mesh-and-solar DePIN network (OmilooS) whose devices register in the same device-registry pallet the AI nodes could use for hardware attestation. If a robot ever joins the Ailoos network, it will register the way a phone does.
The honest thesis
A chain is worth its cost to an AI network when it makes three things checkable by strangers: what trained the model, who did the work, and what they were paid. Ailoos has built the pieces for all three and connected some of them. The rest is on the task list with dates, and this page will move items from promise to mechanism as they cross.
Based on the Ailoos production audit (2026-09-05), pallet-ailoos-rewards (cobrar_era), the training-provenance manifest and FEE_MODEL_AND_LOW_COST_STRATEGY.md.