Skip to content

Rampage Recovery 21 — two machines, measured as one fabric

Pre-release
Pre-release

Choose a tag to compare

View all tags
@Chris-ObtuseAI Chris-ObtuseAI released this 13 Aug 22:18
· 3 commits to main since this release
v0.3.1-recovery.21-bootstrap
3cdc467
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Rampage Recovery 21 — two machines, measured as one fabric

Recovery 21 is the first Rampage package published after a physical owner-PC + laptop compute proof.

Proven on real hardware

  • two enrolled Windows PCs maintained concurrent fresh signed offers
  • Rampage admitted one bounded four-lane SHA-256 job to each machine
  • both jobs returned signed receipts
  • combined measured rate: 62.50 MH/s
  • verified extra capacity over the faster machine: +97.7%
  • estimated time saved for matching fully divisible CPU work: 49.4%
  • the laptop used authenticated direct QUIC with a measured link

What Recovery 21 fixes

  • routes borderless title-bar movement through Tauri's documented direct window API
  • grants only the main window the exact core drag permission it needs
  • prevents bounded Ollama jobs from spending their entire answer budget in a hidden thinking lane
  • fails closed when Ollama produces no user-visible answer text instead of signing a zero-byte success
  • preserves the Recovery 20 signed-checkpoint startup path and fresh-offer restart recovery

The installed Recovery 21 build produced a signed 543-byte local-AI receipt after the earlier zero-byte condition was reproduced on Recovery 20.

Honest boundaries

  • This is an unsigned Windows x64 prerelease. Verify the attached SHA-256 manifest.
  • GitHub-hosted CI, CodeQL, and distribution jobs failed before runner allocation with zero steps. The complete Rust, Clippy, TypeScript, Python, installer, mesh, and physical-compute gates were run locally against the source-bound build.
  • The automated Windows input harness releases drag too quickly to qualify Tauri's asynchronous native drag loop; human drag confirmation remains open at publication.
  • Remote Assist remains absent until the laptop user explicitly enables Allow owner remote control. No session authority is issued without that local opt-in.
  • The benchmark applies to independent divisible work. It does not claim transparent cross-machine VRAM or RAM pooling.
  • The laptop advertises GPU, CPU, memory, storage, and link capacity, but no local model runtime; it cannot host an LLM replica or shard until a compatible runtime and model are actually installed and qualified.
Assets 6
Loading