PR #1143, merged into the main gonka repository on 2026-05-26 by @gmorgachev, bundles the v0.2.13 microrelease. It is a single upgrade that touches 169 files (+17,482 / -2,624 lines), adds a new AI model to the network, fixes a reward bug that could penalise honest miners, and reworks how off-chain inference stores its data.

What changed

• A new model joins the network. MiniMax-M2.7 (MiniMaxAI/MiniMax-M2.7) is added both as a governance model and as a Proof of Compute config. The release also recalibrates the per-model weight scale factors after the vLLM 0.20.1 update, where Kimi-K2.6 was scoring too high on Blackwell-class GPUs. Kimi is set to 0.78 (the Qwen-on-B200 reference plus a 10% top-tier premium) and MiniMax to 0.3024. Validation thresholds were re-measured across vLLM versions: Qwen 0.940, Kimi 0.900, MiniMax 0.922.
• A slashing bug during model rollout is fixed. Confirmation PoC is the step where the network re-checks that a miner really ran the compute it claimed. During a new-model bootstrap it used mismatched model sets for measured weight, preserved weight, and reward rescaling, which could slash honest miners that served both an eligible and a not-yet-eligible model. The fix takes one epoch snapshot of the confirmable models and weight factors, then uses that snapshot for every confirmation and reward calculation.
• Devshard storage stops growing forever. Devshards are the off-chain components that coordinate inference work. Their escrow data previously accumulated in a single SQLite file with no cleanup. Storage is now scoped per epoch and prunes old epochs in the background, keeping only the latest three. Postgres is added as an optional shared store, partitioned by epoch_id so pruning can drop whole epochs cleanly.
• Governance quorum is recalibrated. Genesis guardians held about 34% of adjusted voting power, which made quorum hard to reach when they abstained. The upgrade lowers guardian power to roughly 25% and sets the chain-wide quorum to 0.25. With guardians not voting, that yields an effective one-third quorum among the remaining 75% (0.25 / 0.75 = 0.334).

Why it matters

The slashing fix is the most consequential change for node operators. Adding a model to a live network is risky precisely because the verification math has to handle a period where some nodes serve the new model and some do not. Before this fix, that gap could cost honest miners their rewards through no fault of their own. The snapshot approach removes the ambiguity.

The devshard storage change removes a practical ceiling. Operators running long-lived devshards no longer watch a SQLite file grow without bound, and larger deployments can now point at a shared Postgres instance.

Tradeoffs and follow-ups

The upgrade deliberately disables confirmation PoC for the remainder of the upgrade epoch, reusing the v0.2.10 grace-epoch primitive with an extended protection window of 10,000 blocks, so the new snapshot logic starts cleanly from the next epoch. The release also adds a guardian-signed transaction (MsgSetDevshardRequestsEnabled) for emergency disabling and re-enabling of devshard inference requests.