Overview

Between March 13 and March 20, the Gonka team closed 29 issues — one of the most productive weeks since MainNet launch. Three major workstreams dominated: a full collateral system overhaul, the first stage of reproducible sampling protection, and foundational subnet infrastructure. Capping the week off, v0.2.11 is rolling out to the network today (March 20) — one of the largest upgrades since MainNet, bundling all of this work into a single coordinated deployment.

v0.2.11 Network Upgrade (March 20)

Release v0.2.11, tagged on March 19, is deploying to the live network today. This is not a routine patch — it bundles every feature described below into a single coordinated upgrade: the reworked collateral economics, Stage 1 reproducible sampling, subnet proxy infrastructure, and over a dozen inference pipeline fixes. Validators are upgrading nodes throughout the day; once the upgrade proposal passes governance, the new binary activates at the designated block height. The release also includes inference sharding groundwork and IBC trading support, positioning the protocol for multi-model serving in the coming weeks.

Collateral System Overhaul

The collateral mechanism received significant attention this week. The team shipped weight-based reward distribution (#882), ensuring that validators with higher collateral deposits receive proportionally larger block rewards. A new slashing flow (#868) now differentiates between base and collateral-aware penalties, making the economic consequences of misbehavior more precise.

Edge cases were addressed: collateral adjustments now only apply when collateral is actually enabled on the network (#918), and both PoC scenario and collateral-specific tests were fixed (#881, #888) to match the new behavior.

Reproducible Sampling Protection

Stage 1 of Sequence Check landed (#719, #651). This feature ensures that inference outputs can be independently verified by checking the sampling sequence against committed seeds. It is a critical step toward trustless AI compute — if a validator produces results that diverge from the expected sequence, the protocol can detect and penalize it.

Subnet Infrastructure

Two foundational pieces for subnet support shipped: a proxy server with timeout handling (#891) and subnet parameter configuration (#916). These prepare the protocol for multi-model inference sharding, where different subnets can serve different models or handle overflow capacity.

The inference sharding design (#877) was also finalized and closed, establishing the architecture for splitting large inference workloads across multiple node groups.

Inference Pipeline Hardening

The inference pipeline saw six targeted fixes:

  • Off-chain inference transaction primitives (#729) — experimental support for transactions that skip on-chain settlement for low-value calls
  • Fair slot assignment (#718) — prevents the same node from always receiving the safe inference slot
  • Vested payout forwarding (#770) — correctly propagates caller module names in vested payment flows
  • Hot-path log reduction (#847) — moves high-frequency inference logs from INFO to DEBUG, reducing I/O overhead
  • Performance fixes for missed inference handling (#911)
  • Testnet logging validation (#839) for StartInference and FinishInference transactions

Economics and Settlements

Bounty distribution logic was added (#919), enabling automated reward payouts for completed bounties. Settlement atomicity (#787) received a fix ensuring that SettleAccounts either fully commits or fully rolls back, preventing partial settlement states.

The team also completed a mathematical model for scalability limits (#613), establishing benchmarks for throughput under various network configurations. MLNode benchmarks were upgraded (#872) to reflect current hardware capabilities.

API and Developer Experience

Two error message improvements landed: governance model validation (#680) now returns clear, actionable messages, and the /chat/completions endpoint (#679) correctly reports unsupported model errors instead of generic failures.

Validation weight handling (#655) was fixed to avoid uint32 truncation for large values — a subtle bug that could have caused incorrect validator rankings on networks with many high-weight participants.

Housekeeping

Porosity and negativity checks (#903) were added as new validation guards. Migration handler errors (#901) were resolved. Issue templates (#807) and CONTRIBUTING.md (#815) were updated to streamline the contribution workflow.

By the Numbers

  • 29 issues closed
  • 5 major feature areas (collateral, sampling, subnets, inference, economics)
  • 8 bug fixes
  • 3 performance improvements
  • 2 documentation updates