H1 Google AI Edge Gallery is an official Google app. PASS: publisher shown as Google on the Play Store / App Store listing AND repo under github.com/google-ai-edge. H2 A Gemma "e4b" model exists and the user's "gamma 4 e4b" resolves to a real name. PASS: official model card/doc page found with that name; else report the real nearest name (do not assume). H3 Claim "free, no usage limit, prompts stay on device": each sub-claim marked verified only with a cited primary source (listing, privacy policy, source code); otherwise "not verified". H4 The model runs on this workstation (Intel Core Ultra 7 155H, 22 threads, 30 GB RAM, no NVIDIA GPU). PASS: generates coherent output with a local runtime. If it cannot, report numbers; no forcing, no silent substitute model. H5 Usable: PASS per task type if the checklist in prompts.md is met in >=2 of 3 trials AND decode speed >= 5 tok/s (readable chat speed). Quality measured by fixed checklist, compared to the cloud control. H6 Control: one small OpenRouter model on identical prompts. Download budget: total <= 8 GB, else stop and report. Every claim in RESULTS.md cites a file in outputs/ and is marked verified / not verified.