Trust & economics

The moat. A stake-weighted spot-check makes cheating negative-EV: a failed challenge slashes the node's stake and drops its reputation, which raises its sampling rate. Reputation drives the sellable fraction — the dominant term in the unit economics.

Sellable fraction—%

Verification tax—%

Stake at risk—MYC

Cheats slashed—

Contributor unit economics (per node-hour, RTX 4070-class)

requester $0.15/sellable GPU-h · 20% platform fee

regime	electricity	requester gross	contributor receives	contributor NET

The spread is positive in the GPU + cheap-power + proven-node regime and goes to zero/negative in high-kWh or unproven-node regimes — which is exactly why driving the sellable fraction up (better verification → less replication tax) is the whole business.

Region-aware payouts

NET = gross − local electricity · off-peak gets a bonus

region	$/kWh	renewable	window	NET / node-hr

We recruit GPU + cheap-power + proven supply, and pay a bonus for off-peak/renewable windows — shifting when and where the incremental draw lands.

Refereed-delegation recompute

verification cost → logarithmic

When two nodes disagree, the referee doesn't recompute the whole tile (the 2× replication tax). It treats the computation as a sequential trace, binary-searches to the first divergent step, and recomputes only that one step to convict the cheater — driving verification from O(n) toward O(log n).

running a live challenge…

Host protection · capability sandbox

untrusted jobs run capability-denied + capped

running sandbox checks…

Slice of the Wasmtime/WASI design (Firecracker/gVisor for native/GPU jobs) — capability model + resource caps.

Proven nodes · low spot-check rate

node	rep	stake	spot-check

Flagged nodes · caught cheating

node	rep	stake	spot-check	failed