Week in Numbers
17 pull requests merged across the Gonka monorepo during Feb 27 – Mar 6, 2026. The focus this week: inference speed, PoC reliability, and codebase cleanup.
Performance Overhaul
The biggest theme this week was raw speed. PR #812 reworked StartInference and FinishInference hot paths, while #788 moved inference statistics collection off the critical path entirely. To complement this, #861 disabled stats storage by default — operators who need metrics can opt in, but the chain no longer pays the cost on every block.
Two more PRs attacked storage overhead: #831 switched EnqueueFinishedInference to Cosmos SDK's TransientStore, avoiding unnecessary disk writes for ephemeral data. And #842 precomputed SPRT (Sequential Probability Ratio Test) values used in PoC validation, eliminating repeated calculation during block processing.
Continuous PoC and Consensus Fixes
PR #855 landed the Continuous PoC feature (#821), a significant milestone. Instead of discrete PoC challenges at fixed intervals, validators now face rolling verification — making it harder to game compute contributions.
On the consensus side, #824 fixed a subtle bug in the DKG (Distributed Key Generation) dealer logic that could cause consensus stalls under specific validator set configurations. PR #838 closed a loophole where validators with zero participation could still receive rewards, and #809 converted the SafetyWindow from a hardcoded constant to a chain parameter, allowing governance-controlled tuning.
Infrastructure and Architecture
PR #760 delivered Unified Permissions — a single permission model replacing the previous patchwork of per-module access checks. PR #862 refactored rolling window utilities into a centralized package, reducing code duplication across the inference and PoC modules.
Issue #776 (HA Infrastructure) was closed after the team confirmed that the current architecture meets high-availability requirements without additional redundancy layers at the chain level.
Bug Fixes
- #858 — Fixed timestamp inconsistencies in block metadata
- #740 — Resolved missing payloads in inference results (linked to #612)
- #808 — Enforced integer token amounts, preventing fractional token edge cases
- #525 — Merged all generated batches before forwarding to the dAPI, fixing a race condition in multi-batch inference
Cleanup
PR #841 removed obsolete migration tests that were no longer needed after the PoC V2 migration completed. Small but important — dead tests slow CI and confuse contributors.
What's Next
With Continuous PoC landed and inference paths optimized, the team is turning attention to validator tooling and network monitoring. Expect updates on the Node Manager rollout and further governance parameter tuning in the coming weeks.