S³ FREE GPU FLEET
CONNECTINGBUDGET · £09 LOGICAL TABS
SOV33 · Sovereign distributed compute plane

Nine tabs.
Zero GPU bill.

One auditable control surface for local Apple Silicon, free T4 sessions and Oracle Always Free coordination. Every state and training manifest receives a deterministic SIGIL receipt.

Fleet receiptawaiting orchestrator…
Fleet VRAM—declared accelerator memory
Fleet throughput—T4-equivalent scheduling weight
Active models—unique runtime models
Training / queued—tabs with assigned work
Last sync—refreshes every 15 seconds

Execution mesh

Live state from /api/free-gpu-orchestrator

Waiting for truth plane…

Distributed training

Queue visibility + balanced shard dispatcher

Adapter-first · checkpoint-hashed · £0 sources only

Training jobs queue

No queue data yetWaiting for the orchestrator.

Dispatch sovereign job

The API queues a signed shard manifest. It never falsely reports external notebook execution before a worker heartbeat arrives.

01 · HONEST STATE

“Standby” means configured, not proven online. Training appears only after a signed queue assignment; live throughput remains null until workers publish measured samples/sec.

02 · DETERMINISTIC SHARDS

Dataset ranges are apportioned by worker compute weight. Five T4 workers receive full weight; local Macs and Oracle control-plane workers receive smaller slices.

03 · SOVEREIGN RECEIPTS

Fleet snapshots, per-tab states and job manifests are SHA-256 content-receipted, or HMAC-SHA256 when the deployment secret is configured.