Evidence-based on-device inference

EdgeFit AI is a prototype built for the Meta Global AI Developer Hackathon. It measures real language-model inference on physical Android phones and works toward a global evidence network where only verified, consented, comparable measurements count as evidence.

How it works

  1. Measure on the phone. The EdgeFitDiag Android app runs inference benchmarks on the physical device and exports a signed-shape JSON profile. No permissions, no network, no telemetry.
  2. Review and consent. Contributors preview exactly what would be shared and grant explicit per-submission consent before anything leaves the phone.
  3. Submit for qualification. Consented results enter a server pipeline that validates, deduplicates, and qualifies them through audited eligibility states.
  4. Publish exact cohorts only. Statistics are published only for exact cohorts — same model, build, runtime, workload, and configuration — once 5 eligible runs from 3 independent installations agree the cohort is real.

Current status

Staging backend under test. No verified global aggregates have been published yet. See the evidence status page for what exists today and what does not.

Provenance honesty

All EdgeFit AI source and functionality in this repository is new work started October 26, 2026, for this hackathon. Honey LLM and BeeBench are pre-existing independent projects; EdgeFit does not reuse their code or present their results as its own.