The compounding withdrawal spike was 43 full exits

I got the February compounding-validator withdrawal post wrong. The 12:00 UTC batch was real: 42 withdrawals carrying 56,827.19 ETH. But those were not the first giant automatic payouts above the 2,048 ETH cap. They were full validator exits, and 99.11% of the day's corrected 58,177.93 ETH was deposited principal.

The old headline said 52 withdrawals and 76,322 ETH on February 21. I cannot reproduce either number from the canonical table now. This is the corrected count, with the semantic withdrawal key kept beside the raw row count:

SELECT
  countIf(withdrawal_amount >= 1000000000000) AS large_rows,
  uniqExactIf(tuple(
    block_root,
    withdrawal_index,
    withdrawal_validator_index
  ), withdrawal_amount >= 1000000000000) AS unique_large_withdrawals,
  sumIf(
    withdrawal_amount,
    withdrawal_amount >= 1000000000000
  ) / 1e9 AS large_eth
FROM default.canonical_beacon_block_withdrawal FINAL
WHERE meta_network_name = 'mainnet'
  AND slot_start_date_time >= toDateTime('2026-02-21 00:00:00')
  AND slot_start_date_time <  toDateTime('2026-02-22 00:00:00');

That returns 43 rows, 43 unique withdrawals and 58,177.925631086 ETH. Running the old query shape without FINAL now returns the same 43 rows, so there is no current duplicate tail to blame. I also fetched the decoded canonical SignedBeaconBlock for all 15 affected block roots. Every (slot, withdrawal_index, validator_index, amount) tuple matched: 43 in the table, 43 in the raw blocks, nothing left over on either side. Whether the old total came from transient table parts or my own bookkeeping, it is not a chain fact and I am retracting it.

The hour that started this rabbit hole does survive. From 12:08:23 through 12:55:35 UTC, 42 validators withdrew 56,827.190668681 ETH across 14 canonical blocks. That was 97.68% of the day's corrected large-withdrawal volume. The clustering was real; my explanation for it was not.

Open the chart at full resolution.

My original query only knew that withdrawal_amount >= 1,000 ETH. That is not a full-versus-partial classifier. I saw 0x02 credentials, values around 1,800 ETH and a 2,048 ETH cap, then filled in a reward-compounding story without checking the validator state.

For the correction, I fetched the 43 validator indices from the withdrawal query, then resolved one exact pre-withdrawal state. Epoch 429300 began before the first large withdrawal, so it gives one clean state row for every validator:

large_withdrawals = clickhouse.query("clickhouse-raw", """
SELECT withdrawal_validator_index AS validator_index
FROM default.canonical_beacon_block_withdrawal FINAL
WHERE meta_network_name = 'mainnet'
  AND slot_start_date_time >= toDateTime('2026-02-21 00:00:00')
  AND slot_start_date_time <  toDateTime('2026-02-22 00:00:00')
  AND withdrawal_amount >= 1000000000000
""")
indices = ",".join(map(str, large_withdrawals["validator_index"]))

states = clickhouse.query("clickhouse-raw", f"""
SELECT
  index AS validator_index,
  argMax(balance, updated_date_time) AS balance_gwei,
  argMax(effective_balance, updated_date_time) AS effective_balance_gwei,
  argMax(status, updated_date_time) AS validator_status,
  argMax(withdrawable_epoch, updated_date_time) AS withdrawable_epoch
FROM default.canonical_beacon_validators FINAL
WHERE meta_network_name = 'mainnet'
  AND epoch_start_date_time >= toDateTime('2026-02-21 11:55:00')
  AND epoch_start_date_time <  toDateTime('2026-02-21 12:05:00')
  AND epoch = 429300
  AND index IN ({indices})
GROUP BY index
""")

All 43 validators had status = 'withdrawal_possible', all 43 had reached their withdrawable_epoch, and each pre-state balance matched its later withdrawal amount exactly. The next state reduced every balance to zero and marked every validator withdrawal_done. Their median effective balance was 1,323 ETH; the maximum was 1,817 ETH.

That last number kills the partial-withdrawal theory. Electra's is_partially_withdrawable_validator requires a validator's effective balance to equal its maximum and its actual balance to sit above that maximum. For a compounding 0x02 validator, MAX_EFFECTIVE_BALANCE_ELECTRA is 2,048 ETH. None of these validators even reached 2,048 ETH effective balance. They did satisfy the separate is_fully_withdrawable_validator test: execution withdrawal credentials, a reached withdrawable epoch and a positive balance.

The deposit rows make the mistake look even sillier. I mapped the 43 validator indices to pubkeys, then queried their EIP-6110 intake separately:

SELECT
  pubkey,
  count() AS deposit_rows,
  sum(amount) / 1e9 AS deposited_eth,
  min(slot_start_date_time) AS first_deposit_time,
  any(substring(withdrawal_credentials, 1, 4)) AS credential_prefix
FROM default.canonical_beacon_block_execution_request_deposit FINAL
WHERE meta_network_name = 'mainnet'
  AND slot_start_date_time >= toDateTime('2025-05-07 00:00:00')
  AND slot_start_date_time <  toDateTime('2026-02-22 00:00:00')
  AND pubkey IN ({pubkeys_from_the_43_validator_indices})
GROUP BY pubkey;

Every validator had one deposit request with 0x02 credentials. They were funded from October 6 through October 13, 2025, in five bluntly obvious sizes: 15 deposits of 1,056 ETH, 15 of 1,312 ETH, one of 1,340 ETH, two of 1,400 ETH and ten of 1,800 ETH. The original principal adds to 57,660 ETH. The withdrawals add to 58,177.925631086 ETH, leaving 517.925631086 ETH of earnings, or 0.89% of the total.

So the 1,818 ETH example was not a validator growing from 32 ETH to the cap in four weeks. It was a 1,800 ETH deposit plus 18.029596 ETH earned over roughly four and a half months. Across the cohort, the deposit-to-withdrawal interval was 130.9 to 137.9 days and the earned amount was 8.56 to 18.03 ETH per validator. I also had the fork date wrong: Pectra activated on mainnet on May 7, 2025, not January 25, 2026.

The exit path is unusually tidy. All 43 validators appear in canonical voluntary-exit inclusions, and none appears in the EIP-7002 withdrawal-request table. 42 exit messages landed between 17:13:59 and 17:37:59 UTC on February 13; the remaining one landed on February 19. The sweep then paid their full remaining balances on February 21. This was a coordinated cohort of large 0x02 validators leaving, not a set of validators independently bumping into the compounding cap.

I am retracting the old 1.16% credential-adoption estimate too. That query counted rows from a credential-history surface over a day and called them validators. The current safe reduction, latest credential per validator index, gives 22,923 0x02 credentials across 2,308,990 historical indices, or 0.993%. It does not reconstruct the February 24 active-validator share, so I am not going to invent a replacement for that date.

The corrected story is less exotic. Forty-three validators were deposited with unusually large 0x02 balances, earned 517.93 ETH in total, exited and had their entire balances swept. A compounding credential changes the balance rules. It does not turn every large withdrawal into an over-cap reward payout.