Keryx-Labs/keryx-node v1.3.2-OPoI

Keryx-node-v1.3.2-OPoI

on GitHub

✦ Holder-reward v3 — coin-age

The holder-reward ratio now measures how long coins have actually been held, not just the balance at proof time:

• Every UTXO carries a coin-age anchor (effective_daa); effective balance for the reward ratio only counts coins matured past a 24 h window (864,000 DAA).
• FIFO age carry-over: spending old coins passes their age to the change fairly; splitting or shuffling UTXOs does not reset — or manufacture — age.
• Consolidation cannot mint effective balance: merged UTXOs inherit the youngest surviving anchor, and carried age is capped at maturity. The "consolidate to instantly qualify" shortcut is closed.
• Backed by a new AgeBuckets store with a maturation queue, promotion sweep, deep-reorg rebuild guard, and an era-gated muhash commitment so the UTXO set commitment stays exact across the upgrade.
• The ratio bracket table moves to 9 brackets (v3), replacing the v2 curve.

✦ Proof-of-Model verifier v2

• The possession proof is now verified by fully re-deriving the walk from the opened chunks (recompute-from-chunks) instead of trusting the committed trace
• Wire/DB plumbing is era-exact at every carrier site: v1 proofs remain valid for historical blocks, v2 is required at/after the flip. IBD across the boundary re-validates each era with its own rules.

✦ New model lineup (consensus side)

The five H4 tier roots (R_T), chunk counts and model ids are pinned in consensus, and the per-model inference reward minimums move to 0.5 / 1.0 / 1.5 / 2.5 / 4.0 KRX (EXAONE-4.0-1.2B, Mistral-7B-v0.3, GLM-4-9B, Qwen3.6-27B, Kimi-Linear-48B).

✦ Upgrade mechanics

• PoM proof retention lowered 50k → 25k DAA (GC-only, no consensus impact): v2 proofs are larger, this keeps the pruned datadir footprint at ~21 GB.
• RATIO-DEBUG now emits the producer SPK for off-chain holder-reward attribution (operator tooling).

✦ Community

Thanks to @AnshulMalik for PR #18: the fixed-point OPoI MLP weights are now generated once into a static table instead of being recomputed per verification — a clean perf win on every block validation.