EIP-7928: Block-Level Access Lists

⌬LLM Level 0.

⚇Human No rationale recorded.

⌬LLM Access recording alone does not count as a semantic change, so level 0.

⚇Human Final EVM semantics (gas charged, state changes, reverts) are unchanged. Only the observation order for BAL inclusion is constrained, which is a spec-framework concern, not a behavior change in the anchor's sense.

⌬LLM Level 0.

⚇Human No rationale recorded.

⌬LLM Level 0.

⚇Human No rationale recorded.

⌬LLM No new system contract, so level 0.

⚇Human No rationale recorded.

⌬LLM No blob-gas changes, so level 0.

⚇Human No rationale recorded.

⌬LLM No new refund mechanism, so level 0.

⚇Human No rationale recorded.

⌬LLM Level 0.

⚇Human No rationale recorded.

⌬LLM New execution-header member (bal_hash) and new block-body element, so level 3.

⚇Human New `block_access_list_hash` (Hash32) on the header; `blockAccessList` carried out-of-header via the Engine API.

⌬LLM No activation-specific state transition, so level 0.

⚇Human No state mutation or pre-existing internal variable is modified at the activation block (the builder is initialized fresh per block).

⌬LLM t8n must compute and output the block access list, and possibly its hash. That needs a new mechanism: block-wide access tracking across system calls, transactions and withdrawals. Treating the BAL and its hash as one semantic output gives level 2: a new mechanism with at most one field.

⚇Human A single new mechanism — the BAL builder/validator wired into the t8n (`block_access_lists.py`, `BlockAccessListBuilder`, `hash_block_access_list`, `validate_block_access_list_gas_limit`) — surfaces one new payload output (`blockAccessList` on `ExecutionPayloadV4`). The `block_access_list_hash` on the header is **not an independent field**: it is fully determined as `keccak256(rlp(BAL))` and carries no information not already present in the BAL itself. It is a Merkle/checksum binding, comparable to `transactionsRoot` vs. the transaction list — counting it as a second "new field" would double-count one piece of information. So in t8n terms this is "a new mechanism" (singular) producing one new output, not "multiple new fields **and** a new mechanism" (the 3 bucket). Per the anchor's footnote, the t8n must additionally be fork-aware (Amsterdam-only activation), which warrants special consideration regardless of score.

⌬LLM Gas and execution results of baseline tests are unchanged. Tests that hand-construct block RLP or headers need reworked inputs: header-validation, RLP block-structure and invalid-block families, plus any fixtures with fixed block hashes. That is localized rework in several families without a common behavioral rewrite, so level 2. Adding the BAL assertion to ordinary tests is scored under INV.

⚇Human A considerable subset of existing tests is affected, but the impact is concentrated in a contrived category: benchmarks and gas-boundary fixtures. BAL changes the cost surface (added per-block data, propagation overhead, builder/validator work) so pre-existing benchmarks no longer represent worst cases — new adversarial scenarios (fan-out reads, many-slot touches, deep delegation chains, large BAL-item counts approaching `block_gas_limit // ITEM_COST`) need to be researched and added, and the benchmarking framework needs new metrics for BAL size, generation cost, and validation cost. Existing OOG-boundary fixtures across forks need their warm/cold/touch assumptions re-verified under BAL inclusion semantics. Outside those categories, most pre-existing tests need no rework — Amsterdam-filled tests generate and validate a BAL automatically as a consensus artefact.

⌬LLM A shared validation and trust invariant (BAL correctness) is consumed by block validation, parallel prefetch and execution, and executionless sync. Malicious BALs (missing, spurious, mis-ordered or with wrong values) need coordinated adversarial scenarios across these components, so level 3.

⚇Human The BAL elevates **every state-access check** — not just state writes — into consensus-critical territory. Whether an account or slot appears in the BAL depends on whether the caller had enough gas to *touch* it (cold/warm cache, account existence, 7702 delegation load, static-context check, value-transfer gas), and whether ambient artefacts like the coinbase or precompiles are deemed "touched" under fee-burning, empty-block, zero-tip, and selfdestruct-to-self scenarios. Any client that mis-orders a gas check vs. a state observation, mis-classifies a touch, or leaks cross-block precompile state diverges from the block-hash-bound BAL and breaks consensus. This is a substantial expansion of the consensus surface that interacts with multiple critical components (EVM gas accounting, EIP-2929 warm/cold tracking, EIP-6780 selfdestruct semantics, EIP-7702 delegation resolution, system-contract handling) and warrants extensive review and fuzzing.

⌬LLM The change couples execution tracking (per-transaction diffs, including up to 24 KB code entries), SSZ encoding and hashing, block size and propagation, and parallel IO scheduling. Adversarial blocks that maximize accounts, slots or balance changes per gas need integrated stress testing across execution, networking and validation subsystems. Level 3.

⚇Human The BAL adds ~70 KiB/block of propagation overhead and is interwoven with execution — it cannot be benchmarked in isolation because validation requires re-deriving the BAL during the state transition. Builder behavior interacts with intra-tx coalescing, cross-frame shadowing, net-zero filtering, and the `R_remaining * 2000` gas-budget feasibility check. Benchmark suites under `tests/benchmark/stateful/eip7928_block_level_access_lists/` and `tests/benchmark/compute/eip7928_block_level_access_lists/` were added.

⌬LLM Several boundary-sensitive mechanisms are introduced: list and size limits, MAX_CODE_CHANGES = 1, and the zero-value balance rule. Storage classification is an elevated matrix: whether the slot is read, whether the written value equals the original, zero vs non-zero, reverted vs committed frames, and multiple transactions touching the same slot all interact to decide between storage_changes, storage_reads and absence. Level 3.

⚇Human An elevated number of cases is required. Per `test_cases.md` the BAL implementation drove gas off-by-one boundary tests at every state-access opcode (cold vs warm × OOG-before-target-access / OOG-after-target-access / success-minus-1 / success). Add: net-zero storage filtering (intra-tx coalescing, cross-frame shadowing, nested DELEGATECALL net-zero), CREATE/CREATE2 collisions, same-tx create+SELFDESTRUCT, 7702 invalid-nonce/invalid-chain-id, static-context filtering across all call opcodes, empty-block coinbase, zero-tip coinbase, system-address surplus filtering, RIPEMD-160 cross-block state leak, lexicographic byte-ordering endian traps.

⌬LLM Only a commitment to a new structure is added, presumably with an existing hash primitive. Changed hashed bytes belong under encoding, so level 0.

⚇Human Keccak-256 over RLP only — already covered.