Rampage 0.2.0 — Model Fabric

Rampage 0.2.0 — Model Fabric

Rampage 0.2 turns a group of trusted machines into an explicit strategy surface for local AI. The
desktop now separates the two goals people usually blur together: fitting the largest possible
model and producing the fastest possible response.

What changed

• Maximum Model plans the largest qualified, compatibility-matched aggregate memory placement.
• Speed Boost accepts distributed tensor peers only when signed link measurements predict a
  real gain; slow links fall back to the best whole-model host.
• Throughput, Efficiency, and Autonomous Balanced expose replica, smallest-fit, and
  proposal-only adaptive planning lanes.
• Signed agent offers can advertise bounded Ollama, Exo/MLX, or vLLM/Ray runtime profiles through a
  strict manifest and digest contract.
• The controller, CLI, Rust SDK, TypeScript SDK, Python SDK, and native desktop all share the same
  versioned model-session planning contract.
• The native Tauri shell now has a role-aware Windows tray, close-to-tray behavior, Start with
  Windows, quiet background launch, emergency stop, explicit Quit, and deterministic sidecar
  cleanup.
• MSI and NSIS packages create an automatic desktop launcher plus a Start-menu Rampage Shell whose
  session-local PATH makes the authenticated CLI immediately usable without mutating system PATH.

The boundary that matters

The 0.2 planner is real and read-only. It never turns a capacity estimate into execution authority.
Cross-host model launch and the proposed loopback inference gateway remain disabled until a backend
passes correctness, isolation, topology, streaming-failure, cleanup, and measured-benefit gates.
That distinction is the foundation for making Rampage ambitious without making it reckless.

Verification

The release candidate passes 44 Rust tests, strict all-target Clippy, 4 desktop tests and production
build, 5 TypeScript SDK tests, 13 intelligence tests, 5 Python SDK tests, authenticated direct-QUIC
mesh and artifact proofs, a real Ollama generation receipt, native owner/worker lifecycle probes,
and a silent NSIS install/shortcut/uninstall smoke.

The Windows artifacts are still unsigned. See release evidence before
distribution.

Rampage 0.1.0 — the fabric comes alive

Rampage 0.1.0 — the fabric comes alive

Rampage 0.1 turns Windows PCs controlled by one owner or trusted circle into a governed personal
compute fabric. It is the first complete release candidate of the desktop experience, Rust trust
kernel, Rampage-owned QUIC mesh, encrypted artifact plane, independent shard pooling, SDKs, and
proposal-only intelligence framework.

The headline capabilities

• Create or join a private fabric from the desktop with a complete, signed, ten-minute invitation.
• Discover real CPU, NVIDIA GPU/VRAM, RAM, storage, thermals, power state, runtimes, and local Ollama.
• Place whole jobs or atomically admit bounded independent shard sets with explicit success thresholds.
• Move declared inputs and receipt outputs through encrypted, content-addressed storage leases.
• Execute only compiled allowlisted adapters and return signed, restart-recoverable receipts.
• Inspect the fabric in the spatial Arena or the accessible Ops Grid.
• Run durable DBOS/Pydantic AI improvement workflows whose outputs remain proposals to deterministic policy.
• Stop local work without an AI model, controller response, or network connection.

Qualification

The release passed 35 Rust tests, strict Clippy, desktop and SDK tests, Python lint/type/test gates,
three-shard pooled execution and controller restart recovery, authenticated direct-QUIC control and
artifact transfer, real loopback Ollama work, packaged owner/worker lifecycle tests, and a clean
Windows installer/desktop-shortcut smoke test.

See docs/RELEASE_EVIDENCE.md for the complete evidence record and exact artifact SHA-256 hashes.

Important release boundary

These Windows x64 packages are unsigned release candidates. Windows may show an unknown-publisher
warning. Verify the SHA-256 checksums before installation. Public distribution should be upgraded
to an ObtuseAI Authenticode-signed build before this is described as a generally trusted installer.

Rampage 0.1 is for one owner or a deliberately trusted circle. It is not an anonymous public compute
marketplace, mobile/console support is not yet packaged, and cross-host tensor sharding remains gated.