Draft EIP-9999 landed yesterday with a clean argument for deleting the sync committee: 512 validators, one seat per 1,761 validators, only 0.06% of the active set. At the start of the current committee period, 901,612 / 512 was 1,760.96. The arithmetic is almost comically exact.
The word "validator" is doing too much work.
Blob transactions have an expensive way to say "same data, higher fee." The transaction hash changes, so the network pulls the whole thing again, including blob content it may already have.
EIP-8094 proposes splitting those pieces apart. I checked the premise against 14 complete days of mainnet mempool observations. Out of 3,427 consecutive same-sender, same-nonce blob fee increases, 3,413 kept the exact same versioned-hash array. That is 99.59%.
A small EIP-8141 text fix landed on Friday. The frame-transaction draft had accidentally priced zero calldata bytes at 16 gas and nonzero bytes at 64; the intended EIP-7976 floor is 64 for both. Execution specs already counted the bytes correctly, so this was not a client fix or a live-network bug.
It did send me back to current mainnet calldata. Across the 14 complete UTC days from August 7 through August 20, one executeBatch target accounts for 30.01 billion of the 67.87 billion added gas in a fixed-path EIP-7976 backcast. That is 44.22% from one address, and 85.7% of the affected calldata bytes sent to it were zero.
A higher-nonce blob transaction can reach every monitored observer before the transaction it must follow. I found 23 of those adjacent-nonce pairs across 14 complete mainnet days. In the slowest pair, the median lead for nonce N+1 was three minutes.
That is the awkward edge case behind EIP-8077. The draft would put sender and nonce into transaction announcements so a peer can spot a future nonce before it fetches the full transaction. All ten monitored observers saw the higher nonce first in only 23 of roughly 97,000 eligible pairs.
EIP-7923 sounded like a repricing for pathological memory users. Then I ran its 100-gas page charge over 34.08 million recent mainnet transactions, holding the historical execution paths fixed. On those paths, it nearly deletes the charge it replaces: no more than 0.52 billion gas instead of 10.35 billion.
That 95% headline is real and a little misleading. Memory expansion was only 0.340% of all transaction gas in the sample, so the minimum fixed-path saving is 0.323% of historical transaction gas. The first free 4 KiB page does most of the work.
EIP-8347 changed one file format on August 20. The new format is brutally easy to size: 24 bytes per account, then 32 bytes per live storage slot. Point that arithmetic at mainnet on August 19 and the file comes out at 62.25 GB.
Roughly 40.7 GB of it maps to slot keys that had not been read or written in twelve months.
Plataberget spent five hours producing block access lists around 24 KiB. A few hours later they had settled near 86 KiB, yet the lists still held roughly the same number of unique addresses and storage keys.
That is the part I did not expect. The devnet pushed almost ten times more gas through each block without adding more transactions or more BAL items. The extra 61 KiB went into the payload hanging off those items.
Yesterday's Plataberget post stopped at 11:15 UTC with finality still pinned to epoch 1,534. I left the same 84 validator-node cohort in place and kept watching. The extra three hours made the head-count proxy look pretty bad.
At 09:05, 57 of the 84 head gauges were already at or beyond the Gloas fork slot. That is 67.86% of the nodes, but the latest complete epoch had only 45.94% of eligible stake voting for the target. Finality did not return until 13:55, four hours and 50 minutes later.
A fresh Gloas spec change sent me looking for an ugly sanity check: do observed peer subscriptions line up with custody metadata? The obvious raw table answered before I had joined anything. One stored peer_id_unique_key owned almost exactly half the rows across every observer. Gloas is not live on mainnet; it just gave me a reason to inspect current mainnet capture.
That peer does not exist.
Plataberget's Gloas fork looked like an epoch boundary until the head metrics stopped moving. Lodestar and Nimbus reported a post-fork head in 21 seconds median. Prysm and Teku did not get there for roughly 93 minutes.
The fork was not just a version switch. It walked 16,306 pending deposits to build Gloas's new payload-builder registry, and the way clients handled that queue decided who crossed quickly.