Reporting window: 2026-08-31 20:20:18 WITA through 2026-09-07 20:16:35 WITA (Asia/Makassar).

Glamsterdam lost a week on its public-testnet schedule because its latest private network still did not produce the clean fork transition client teams wanted. AllCoreDevs Consensus set the Sepolia activation target to October 6 and deferred Hoodi’s date, replacing the September 28 Sepolia proposal discussed last week. The delay is short, but the reason matters: devnet-9 suffered non-finality amid outdated branches and client failures, while testing exposed a Gloas bid that could halt the network. Core developers responded by merging the missing validity check, planning another devnet, and accepting a late state-gas correction. Glamsterdam is moving forward, but the upgrade is still earning its public-testnet date through failure-driven specification work.

A public-testnet date now depends on a clean transition

The non-finality devnet ran at large scale but encountered state-transition, memory, and attestation problems across several clients. The same test campaign surfaced a more serious Gloas edge case: a builder bid could carry a block_hash equal to its parent_block_hash, and accepting it could stop the network. Nico Flaig’s consensus-spec change, credited to Bhargava Shastry for the finding, made those bids invalid and merged on September 3; the pull request also links merged Lodestar and Teku client fixes.

Rather than treating recovery of the troubled network as the release gate, the call planned devnet-10 for the following week to exercise a stable transition. October 6 is therefore a coordination target, not an unconditional promise. The signal to watch is whether devnet-10 crosses Gloas with current client branches and stays stable long enough for downstream users to test; only then does the calendar decision become evidence of readiness.

State gas changed after multi-client testing

The other late Glamsterdam decision affects EIP-8037, State Creation Gas Cost Increase. Cross-frame testing showed that state gas refunded in a different frame could remain in the state-only reservoir while the parent’s spendable gas_left stayed reduced. The same failure mode could make GAS report less and cause an out-of-gas result even though the transaction still held an unusable credit. The accepted correction repays the parent’s outstanding spill when a successful child frame merges, capped by the amount previously taken from execution gas.

ACDC recorded client support for the correction after Geth and Erigon tested it on devnet-8. The change then moved beyond meeting consensus: the EIP edit merged on September 4, and the executable Amsterdam specification merged the repayment behavior and its tests the same day. That sequence distinguishes a late design discussion from a canonical rule change.

The correction also propagated into the native-account-abstraction test track. The fourth EIP-8141, Frame Transaction, fixture release rebased onto the revised Amsterdam state-gas rules while leaving the Frame Transaction specification itself frozen at its September 1 commit. The tracker explicitly lists no devnet launch date, so the evidence remains released fixtures and Hive consumption, not a live multi-client Frames network.

FOCIL’s first interop failures hardened the wire rules

The dedicated EIP-7805, Fork-choice enforced Inclusion Lists (FOCIL), network is now producing the kind of evidence its earlier test plan lacked. On focil-devnet-0, Barnabas Busa observed that the 8,348-byte gossip limit only fit one maximum-size transaction because each additional SSZ list entry adds a four-byte offset. Under load, Lodestar rejected 209 of 211 publication attempts and would have dropped full Teku lists. The observed matrix covered Ethrex and Besu execution clients with Lodestar and Teku consensus clients.

The merged repair raised the signed inclusion-list maximum to 41,112 bytes, the theoretical worst case under the existing raw transaction-byte limit. A second consensus-spec change added missing dependent-block slot and root validity checks to inclusion-list gossip. These are not changes to FOCIL’s censorship-resistance objective; they are interoperability rules that only became concrete once different clients exchanged real lists. Next week’s useful signal is sustained mixed-client operation after the new bounds and checks land everywhere.

RowDAS arrives with measurements, not yet a network

For Hegotá beyond its two scheduled headliners, EIP-8371, RowDAS - Distributed Blob Reconstruction, gained the strongest new research artifact of the week. Csaba Király and collaborators published a Prysm prototype and network harness that reduce worst-case per-reconstructor work at 128 blobs from 20.7 seconds to 162 milliseconds of CPU core time per slot. Their paired simulations reported the existing column path within noise of baseline and tested networks as large as 1,000 nodes.

The authors are explicit about the boundary of that evidence: there is no devnet, the 128-row configuration is extrapolated from smaller subnet counts, and sparse-blobpool compatibility is still a design goal rather than a measured result. The Hegotá meta EIP lists RowDAS as Proposed for Inclusion, not Scheduled for Inclusion. Client review should focus on whether the full row channel earns its networking complexity over the reduced duty-sharded PeerDAS fallback, then reproduce the result at mainnet scale and on a real network.

The ERC track has a smaller but time-sensitive review cluster. ERC-7945, Confidential Transactions Supported Token, closes Last Call on September 8, followed by ERC-8153, Facet-Based Diamonds, on September 9 and ERC-8167, Modular Dispatch Proxies, on September 10. Only the last of those newly advanced this week: its move to Last Call merged on September 1. Absent substantive objections, these are process deadlines rather than evidence of adoption.

Sources and further reading