§1 — The brief.
The brief.
Gonka AI is an L1 for AI compute. The thesis: nodes contribute GPU time to a network that runs ML inference, and consensus weights voting power by verified compute contribution rather than raw hash rate or token stake. The verification problem is harder than PoW or PoS — the work being verified is ML inference on a 235B-parameter transformer, which cannot be cheaply spot-checked the way a hash computation can.
Gonka's team brought Zpoken in as a contributor on the consensus layer and the network application surface. The engagement is open-source from the start; commits and pull requests are public on the gonka-ai/gonka repository, alongside contributions from other engineers in the Gonka contributor set.
§2 — What we shipped.
What we shipped.
[01]
Sprint mechanism + statistical validation
─ POC CONSENSUS
[02]
PoC v1 → v2 migration · BLS slot keys
─ POC CONSENSUS
[03]
MMR off-chain artifact storage
─ POC SCALING
[04]
vLLM-integrated PoC (no model offload)
─ POC RUNTIME
[05]
Bridge security (bridge-geth · bridge-prysm)
─ BRIDGE LAYER
[06]
CertiK audit closure
─ BRIDGE LAYER
[07]
app.gonka.ai web application
─ APPLICATION
[08]
inferenced CLI utilities
─ TOOLING
The PoC consensus contribution covers the Sprint mechanism — time-bound computational races where hosts run AI-relevant inference on a transformer with randomized layers over a stream of nonces, producing output vectors that determine voting weight for the next epoch — along with the validation logic that verifies submitted nonces are produced by the required Sprint model. The PoC v1 → v2 migration covered parameter migration, BLS slot-key precomputation for optimized verification, and the dual-mode tracking phase that measured adoption before flipping to full enforcement.
The bridge security stack covers contributions to bridge-geth and bridge-prysm — block comparison fixes, address validation, chain mapping, and the work to close the CertiK audit findings.
app.gonka.ai is the network application interface for hosts and developers; the inferenced CLI provides account management and network operations.
§3 — Engineering decisions.
Engineering decisions.
A swappable verification primitive for PoC.
The Sprint mechanism + statistical validation is the verification primitive in production today. The architectural call we contributed to: keep the consensus mechanism separate from the verification primitive, so Gonka can ship now with sampling + statistical validation + slashing, and migrate specific verification paths to ZK-of-inference later, when prover costs become economic at production model sizes. The reasoning: locking the consensus to today's verification primitive would force a mainnet hard fork later. Layering them keeps the upgrade path soft.
Off-chain artifact storage with MMR commitments.
Storing full PoC artifacts on-chain doesn't scale past a single network model. Moving artifact storage off-chain with Merkle Mountain Range commitments — recording only root_hash and count on-chain — preserves verifiability while making multi-model PoC tractable. This shipped alongside the PoC v2 enforcement.
PoC integrated directly into vLLM, not offloaded.
Rather than offloading the production inference model and loading a separate PoC model when verification runs, PoC runs directly inside vLLM. Hosts switch from inference to PoC without model reloads. The cost is slightly tighter coupling between the PoC mechanism and the vLLM serving stack; the benefit is a verification primitive that doesn't pause real production work.
§4 — Where it stands.
Where it stands.
Active contributor relationship. Ongoing PRs and issues against the public repository. Recent contribution areas include multi-model PoC support, BLS gas-scaling patches, DevShards reward distribution, and continuous PoC infrastructure work.
§5 — Tech stack.
Tech stack.
~/zpoken/work/gonka/stack
Cosmos SDK · Go (chain node, api node) · Python (mlnode, vLLM integration) · Docker · Cosmovisor (deployment + on-chain upgrades) · BLS signature aggregation · MMR commitments · Qwen3-235B-A22B (current network model)
§6 — Public surface.
Public surface.
§7 — Founders.
Founders.
Engagement led by Mike Yezhov and Anton Yezhov. Code committed under contributor accounts visible in the gonka-ai/gonka repository.