Firsts: The First Mobile-Trained Adapter Implemented into the Production Model

The loop, closed

On a POCO F8 PRO, 8 September 2026:

1. Download model_empoorio_lm_64m_v1 via SynoodoS — 258.1 MB.
2. Resume from the phone's previous adapter — 80 tensors.
3. 30 real LoRA steps on device with ExecuTorch. Loss 6.8669 → 6.4434.
4. Sign the adapter with the node's ED25519 key.
5. Upload with the session's secret.
6. Coordinator validation: geometry, signature, loss trajectory — approved.
7. Merge and implement: adapter_art_26941b6079e0 now part of the production model.

Why it was a first

Because every earlier session had been broken at a different link, each invisible from the others:

• Random weights. The production .pte artefact had never had trained weights loaded; every phone had been fine-tuning noise. Loss went down anyway (~9.0 on noise; 6.9 on the real base).
• Uploads rejected since 6 August. A security fix added an upload secret; the app requested it and discarded it. Every upload: 422, for 33 days.
• The merged adapter unloadable. FedAvg wrote the weights without the configuration file, and the loader accepted only an older contract than Android produced. Rewards were computed on adapters nothing could load.
• Attention leaking the future. The causal mask was commented out in the model.
• Roles never started. The headless node initialised and slept; the status command said the loops ran.

Each fixed; each with a test; the loop then closed on the first attempt.

What is still open

The reward for this adapter is held for manual audit because the bridge URL in strict mode points at localhost — the known block from the production audit. The first paid mobile training is the next first.

What it means

The thesis of Ailoos — phones train the ecosystem's models and are paid for it — was, until 8 September, an architecture. It is now an event with a device, a date, a loss curve and an artefact id. One event. It is enough to build the second on.

Breakpoint

The page this replaces announced the return of another ecosystem's conference. Ailoos's breakpoint was a loss of 6.44 on a phone in a hand.

Based on the Ailoos mobile-training verification (2026-09-08).