Notable updates to the benchmark: scoring changes, new methods and vantages, and fixes.
Quicknode's getTransactionsForAddress used to return a non-comparable shape (bare array instead of the {data, paginationToken} envelope, always-full details, slot pin ignored), so it couldn't vote. Re-probed live: all of that is fixed, and its answers are byte-equal with Helius and Alchemy across both buckets. The method goes from 2 voters back to 3 (Helius, Alchemy, Quicknode) — all three must answer and a 2-1 split is now decided with the deviator attributed. Second panel change for this method in August, after Triton dropped it on 2026-08-20. METHODOLOGY_VERSION stays at 4, so history is preserved: read a step in this method's correctness series as a rule change, not a provider regression.
Triton's endpoint began returning -32601 Method not found for getTransactionsForAddress, 100% of calls, while every other method on it stayed healthy. It's now declared unsupported for that method — scored on reliability, not marked wrong on a method its tier no longer serves. That took the panel to 2 voters and, for eleven days, made correctness there a pairwise byte-equal agreement check with no tie-breaker.
New Sends leaderboard (/sends): real transactions broadcast through each provider's standard endpoint and scored against the chain — landing rate and slot latency. Measures plain JSON-RPC sendTransaction across the 5 benchmarked providers (Helius, Alchemy, Triton, Quicknode, Chainstack) — no tips, no relays, apples-to-apples. Confirmation is a poll loop folded into the generator (getSignatureStatuses) — no separate service. Adds migration 0002 and a SENDS_ENABLED kill-switch. Additive: read scoring is unchanged. Versioned independently via SEND_METHODOLOGY_VERSION.
Chainstack joins the benchmarked panel (now five: Helius, Triton, Alchemy, Quicknode, Chainstack). METHODOLOGY_VERSION bumped to 4 — adding a fifth voter changes getStakeMinimumDelegation's consensus rule from a relaxed 2-of-3 to the default 3-of-4 strict majority, so pre/post results are scored under different semantics and not blended.
Open-source launch. Majority-consensus correctness across the benchmarked panel with honeypot spot-checks, ~45 read methods, per-region vantages, and a fully reproducible pipeline. Product releases follow semver from here; methodology changes are tracked separately in docs/methodology.md.