cherry: P0 data-quality → main

benchmarks/evm-block-builders.yml

@@ -13,7 +13,7 @@ unit: pct
 higher_is_better: true
 
 disclaimer: |
-  Attribution relies on the extraData tag each builder voluntarily stamps into its blocks. A builder can change or strip its tag at any time; unrecognized tags land in the Other row and every raw string is logged so the table grows. The Vanilla row (locally built blocks) could be marginally overstated if a builder deliberately uses a client-default tag to avoid attribution — the relay cross-check is the independent signal that surfaces this divergence. High share is a centralization datapoint, not a quality ranking.
+  Attribution relies on the extraData tag each builder voluntarily stamps. A builder can change or strip its tag; unrecognized tags land in Other and raw strings are logged so the table grows. The Vanilla row (locally built) could be marginally overstated if a builder uses a client-default tag; the relay cross-check surfaces this. High share is a centralization datapoint, not a quality ranking.
 
 seo_intro: |
   Who actually builds Ethereum? Since MEV-Boost, validators outsource
@@ -207,8 +207,8 @@ providers:
 
   - slug: vanilla
     name: Vanilla
-    tag: Locally built by the proposer — or a builder who blanked/faked a client-default tag (rare; relay cross-check surfaces divergence)
-    formula: "Blocks with empty or execution-client default extraData (geth, reth, nethermind, besu, erigon), meaning the proposer built locally, as a percentage of all blocks observed in 24h. Could be marginally overstated if a builder fakes a client-default tag; the relay cross-check is the independent verification signal."
+    tag: Locally built by the proposer, or a builder who blanked/faked a client-default tag (rare; relay cross-check surfaces divergence)
+    formula: "Blocks with empty or client-default extraData (geth, reth, nethermind, besu, erigon), meaning the proposer built locally, as a percent of all 24h blocks. Marginally overstated if a builder fakes a client tag; relay cross-check verifies."
     queries:
       p50: 100 * sum(increase(ebb_blocks_total{builder="vanilla"}[24h])) / scalar(sum(increase(ebb_blocks_total[24h])))
       p90: 100 * sum(increase(ebb_blocks_total{builder="vanilla"}[24h])) / scalar(sum(increase(ebb_blocks_total[24h])))

benchmarks/nft-collection-metadata.yml

@@ -66,7 +66,7 @@ findings:
   - "external_url is the field most likely to differ between providers: OpenSea and Moralis surface the collection's website on most blue-chips, Alchemy's `openSeaMetadata.externalUrl` is structurally null on the same set. p99 ({{p99:moralis}} / {{p99:alchemy}} / {{p99:opensea}}) is what a portfolio tracker sees during the worst 6h window in the day."
   - "{{worst_name}} trails at {{worst_p50}} (p50, 24 h). A 20-point gap on this leaderboard means roughly one in five field renders falls back to a placeholder when displaying a blue-chip collection."
 
-source: https://github.com/ChainBench/OpenChainBench/tree/main/harnesses/nft-collection-metadata
+source: https://github.com/ChainBench/OpenChainBench/tree/main/harnesses/nft-metadata-coverage
 
 prometheus:
   window: 24h

benchmarks/pm-api-latency.yml

@@ -105,7 +105,7 @@ per_chain_explainer:
     h2: Predexon API latency
     body: "Predexon is a multi-venue aggregator that unifies Polymarket, Kalshi and Limitless behind one schema. Its API answers at {{p50:predexon}} p50 (24h, all regions), measured on the same market-price endpoint the harness probes on the native venues, so the direct comparison of a unified aggregator against three venue-native APIs is head to head. Predexon aggregates upstream data; a Predexon latency higher than the sum of its three underliers is the cost of unification. Lower is the value of having one contract instead of three."
 
-source: https://github.com/ChainBench/OpenChainBench/tree/main/harnesses/pm-api-latency
+source: https://github.com/ChainBench/OpenChainBench/tree/main/harnesses/pm-rate-limits
 
 prometheus:
   window: 24h

benchmarks/rwa-yield-accuracy.yml

@@ -20,7 +20,7 @@ higher_is_better: true
 
 seo_intro: |
   Every tokenized U.S. Treasury on the market publishes an APY.
-  Ondo USDY, Superstate USTB, Ondo OUSG, Maple SyrupUSDC — each
+  Ondo USDY, Superstate USTB, Ondo OUSG, Maple SyrupUSDC, each
   advertises a yield on its public dashboard. This page measures what
   they actually pay their token holders on-chain, using each issuer's
   own distribution mechanism (daily rebase, monthly USDC dividend,

benchmarks/validator-yield.yml

@@ -93,7 +93,7 @@ methodology:
   - "MEV treatment. Solana: MEV tips from Jito are already folded into Stakewiz `total_apy`. `ocb_validator_mev_share_bps` is exposed for transparency (a reader sees 'Helius APR 600 bps of which 80 bps MEV') but NOT subtracted to produce a 'staking-only' figure. Hyperliquid: no separate MEV layer (centralised sequencer captures order-flow value at the protocol level), so the MEV gauge is 0 by convention."
   - "Ethereum. Source: `GET ultrasound.money/api/v2/fees/effective-balance-sum` (free, no key), live total beacon-chain effective balance. APR = 64 × epochs_per_year / sqrt(total_gwei), the consensus spec reward formula at 100% participation (realized ~99.5%, under 2 bps of drift). One network-average series: consensus rewards are uniform per 32 ETH increment, so the average equals the solo validator nominal APR. Execution tips and MEV excluded. Price: CoinGecko (ETH/USD)."
   - "Jailed / inactive handling. Hyperliquid: `jailed=1` validators (currently 4) land in the dataset with `predictedApr=0`, so `net_yield_bps=0`. Solana: ~42 validators report `total_apy=0`. Live data shows two causes: commission 100% (operator routes everything to themselves, delegators get 0) or zero leader slots this epoch (no block-production rewards). Both legitimate transient states, both count toward total + median."
-  - "Excluded by design. Per-validator Ethereum gauges (~1M active indices; needs beacon-node ingestion, deferred to v2). Ethereum execution tips and MEV (needs per-relay MEV-Boost scraping; the published figure is consensus only). Lido / Rocket Pool / Coinbase staked-ETH are products with own fee structures, not validator yields (separate bench `liquid-staking-yield`, future). Polkadot not yet active (requires separate source ingestion, deferred to v2). Cosmos Hub, Avalanche, Cardano and Sui now tracked in v1."
+  - "Excluded by design. Per-validator Ethereum gauges (~1M indices; beacon-node ingestion, v2). Ethereum execution tips + MEV (per-relay MEV-Boost scraping needed; published figure is consensus only). Lido/RocketPool/Coinbase are products with fees, not validator yields (separate bench). Polkadot v2. Cosmos Hub, Avalanche, Cardano, Sui now tracked."
 
 findings:
   - "{{best_name}} currently leads at {{best_p50}} (median net yield, 24 h) across {{count}} measured chains. The cross-chain gap is structural: Solana's ~5.6% median reflects PoS inflation + Jito MEV tips minus average ~24% validator commission, while Hyperliquid's ~2.2% reflects staking-only rewards on a deflationary HYPE token with no MEV layer to top it up."

src/lib/removed-benches.ts

@@ -48,6 +48,13 @@ export const RENAMED_BENCH_SLUGS: Record<string, string> = {
 export const REMOVED_ANSWER_SLUGS = new Set([
   "which-evm-aggregator-has-the-fastest-quote",
   "which-solana-rpc-lands-the-most-transactions",
+  // References rpc-keyed-latency (bench 069, staging-only, in
+  // REMOVED_BENCH_SLUGS below). Prod was returning 404 on this answer
+  // page because the underlying bench 410s. Answer stays on staging.
+  "alchemy-vs-quicknode-vs-infura-latency",
+  // References solana-dex-quote-latency (staging-only, in
+  // REMOVED_BENCH_SLUGS). Same pattern.
+  "which-solana-dex-aggregator-is-the-fastest",
 ]);
 
 export const REMOVED_BENCH_SLUGS = new Set([