Tactical football commentary from a model running on your own device via QVAC — and when the phone is too weak, it borrows a laptop peer's compute over an encrypted Pear link.
"I was in the away end with no signal and a three-year-old phone — the cloud commentary app spun forever and I missed why the ref pointed to the spot."
→ so the fix can't be a bigger cloud. It has to be no cloud at all.
completion({stream:true}) narrates every press, overlap and penalty token-by-token from a football system prompt — entirely on your phone.→ the offload is the product. Everything else got cut.
Shown on Gaffer's disclosed sim engine (labelled on every surface, never presented as a model or cloud AI). The P2P transport is identical and real with --engine qvac — the same bench then measures real model tok/s on your hardware.
| QVAC | The mind, on-device. loadModel({modelSrc:'pear://…'}) · completion({stream:true}) → tokenStream · textToSpeech() (Piper) · startQVACProvider({topic}) when available. |
| Pear | The transport, peer-to-peer. Hyperswarm match rooms · Noise Secretstream E2E · Hyperdrive model sharing over pear:// · HyperDHT holepunching. |
| Core | The honesty layer. lib/state.js — every (state × event) pair defined and tested; lib/router.js — strict token ordering, token-exact resume on deterministic engines. |
Take QVAC + Pear out and you'd need a cloud LLM, a cloud TTS, a model CDN, a signalling/TURN server and a rented GPU — plus a privacy policy asking fans to trust you.
The ×7.8 transport figure is measured on disclosed sim-engine profiles (6 vs 48 tok/s) — CI also runs the bench artifact and a submission-readiness gate that re-verifies this deck's test count against a live run.
Even the one with a dead phone in the away end. Gaffer never once phones home.
github.com/edycutjong/gaffer · built by Edy Cu