Zero Echelon - Uplink

Uplink is an on-device AI that brings your laptop back online when the internet breaks, with no cloud and no connection required. Uplink fixed DNS faults 10/10, proxy faults 5/5 and wrong-gateway faults 2/2 in model-reached trials (small n), with zero false ONLINE across every reported live test. It uses only BASE Meta Muse Glimmer 30B, open weights, run locally with no model adapter. When a laptop goes offline, cloud help can disappear; Uplink keeps recovery help on the device. The base, unmodified model issues one of thirteen typed actions using measured facts and a known-good snapshot. A bounded executor applies snapshot-based restores; the model has no shell. An independent verifier decides ONLINE. A detached watchdog attempts rollback; only independent proof establishes handback. Recovery medians: DNS 15.319 s, proxy 10.395 s, gateway 13.0595 s. DNS spans three policy segments; proxy spans two with entry memory limits of 24.5 then 24.75 GiB. Four DNS and six proxy pre-model stops are excluded. Timings are warm, not cold-start. Safety test: across 6 healthy-network or misleading-status trials, Uplink changed nothing and never claimed a false recovery (6/6 safe; zero network changes and false-action requests). Strict score 0/6: the rubric also required the correct diagnosis reason (verified), which all six missed, 0/3 per case. Five pre-model control stops are excluded. Safety is not recovery credit. Gateway caveats: target 5; only 2 reached. Trial 3 and five earlier pre-model ABORTs are excluded (6 total). Small n; no continuation. Second PASS and final ABORT video: CAPTURE_INCOMPLETE; media not committed. Trial 3 retained restore_failed; later empty-baseline comparisons do not prove bounded historical handback. Known limits and next steps: Wi-Fi radio off: 1/5 in the specified continuation, one strict recovery 11.213 s, false ONLINE 0; one FAIL and three ABORTs counted. All listed series: 1/6 after an earlier model-reached ABORT. Four listed pre-model stops excluded; small n. Next: improve radio-state detection before reconfiguration. Site-only, owned 503 stub on a healthy network: 1/3 correct-report PASS, two earlier ABORTs counted, one pre-model stop excluded. The strict-reader VALIDATED report identified the server error and changed nothing; false ONLINE 0, recovery median n/a. Not a real outage. CAPTURE_INCOMPLETE; trial unaffected, videos not committed. Next: sharper app-vs-network attribution. These bounded results do not establish general field reliability. Pre-existing: Meta weights, llama.cpp and Python libraries. A private pre-kickoff prototype and design notes were reference only; no code was ported. Uplink's loop, typed tools, safety gates, watchdog, oracle and evidence tooling were built during the hackathon.