Retrospective LLM-Based Complexity Evaluations

Amsterdam Comparison

Human and LLM Complexity Evaluations

A descriptive comparison of published human assessments and blinded LLM assessments for 12 Amsterdam Execution Layer EIPs, evaluated under the same checklist revision so criteria and totals are directly comparable.

Why Amsterdam? It is currently the only fork with a completed human complexity evaluation; the separate manual evaluation for Hegotá is still underway (25 of 46 Hegotá candidates have a scored human checklist, and each Hegotá EIP page shows its Human status and a same-rubric comparison where one exists). This page compares the Amsterdam LLM-based and human assessments to examine how closely they align and whether their disagreements reflect systematic differences in how the complexity criteria were interpreted. It is a structured sanity check, not evidence that the LLM is calibrated to human judgment.
The two LLM evaluations differ only in the template. Both used the same pinned EIP and supporting inputs under the same blinded LLM procedure. LLM v1 used the historical complexity-assessment template v1, the same 24-criterion checklist the human reviewers used, and is the evaluator compared here; LLM v2 used complexity-assessment template v2 and is the study's primary score. No proposal revision or implementation evidence changed between them.
Same-rubric comparisons
12
2 clean, 10 confounded
Mean Δ total (LLM − Human)
+2.83
mean |Δ| 5.33 · median |Δ| 5
Who scored higher
8 LLM · 4 Human
0 equal totals
Same tier
2 of 12
revision-1 tiers: Low <10 · Medium 10–19 · High ≥20

Complexity profiles per EIP

Human and LLM revision-1 profiles for each EIP on one shared scale, ranked by the Human total. Δ is LLM minus Human. Open an EIP for its per-criterion comparison with the rationale from both evaluators.

  1. HumanChecklist v1
    LLMChecklist v1
  2. HumanChecklist v1
    LLMChecklist v1
  3. HumanChecklist v1
    LLMChecklist v1
  4. HumanChecklist v1
    LLMChecklist v1
  5. HumanChecklist v1
    LLMChecklist v1
  6. HumanChecklist v1
    LLMChecklist v1
  7. HumanChecklist v1
    LLMChecklist v1
  8. HumanChecklist v1
    LLMChecklist v1
  9. HumanChecklist v1
    LLMChecklist v1
  10. HumanChecklist v1
    LLMChecklist v1
  11. HumanChecklist v1
    LLMChecklist v1
  12. HumanChecklist v1
    LLMChecklist v1

Where the Evaluators Differ by Criterion

Mean per-criterion difference across the 12 EIPs, LLM minus Human, under the shared revision-1 checklist. Positive values mean the LLM scored the criterion higher on average.

Per-criterion differences across 12 Amsterdam EIPs (checklist revision 1)
Security risks+1.501.83921
Cross-EIP interactions+1.171.17804
Performance risks+1.001.00705
EVM Gas rule changes−0.921.25255
Encoding changes (RLP/SSZ)+0.500.502010
Edge/boundary conditions+0.500.67516
New or modified transaction validity mechanisms−0.420.58147
New fork activation mechanism+0.420.422010
Patterns affecting pre-existing tests−0.331.17543
Modified system contracts−0.250.250210
New EVM gas refund−0.250.250210
Engine API changes−0.250.250111
Modified opcodes+0.170.67219
Block syncing changes+0.170.172010
Added system contracts−0.080.080111
Blob gas accounting changes−0.080.080111
Engine API encoding changes+0.080.081011
Transition-tool interface changes−0.080.42129
Added opcodes00.000012
Added precompiles00.000012
Modified precompiles00.000012
New transaction types00.000012
New block / header fields00.000012
Cryptography00.000012

Totals per EIP

Amsterdam totals per EIP: Human and LLM under checklist revision 1, and the study's primary revision-2 LLM score
EIP-7928Block-Level Access Lists2926−3High40No
EIP-8037State Creation Gas Cost Increase2821−7High35No
EIP-8038State-access gas cost update2017−3High vs Medium17No
EIP-2780Resource-based intrinsic transaction gas1320+7Medium vs High25No
EIP-7778Block Gas Accounting without Refunds108−2Medium vs Low13No
EIP-7708ETH transfers emit a log913+4Low vs Medium19Yes
EIP-7610Revert creation in case of non-empty storage710+3Low vs Medium9No
EIP-7843SLOTNUM opcode717+10Low vs Medium23Yes
EIP-7981Increase Access List Cost611+5Low vs Medium11No
EIP-8024Backward compatible SWAPN, DUPN, EXCHANGE611+5Low vs Medium13No
EIP-7976Increase Calldata Floor Cost512+7Low vs Medium10No
EIP-7997Deterministic Factory Contract513+8Low vs Medium14No

What Differed Systematically?

The strongest supported use is as a structured second opinion: retain the LLM’s criterion-level flags and ranking signal, then have a human approve the evidence, aggregation, and final score.

Criterion legend for the revision-1 checklist

Every stacked bar, comparison matrix, and criterion table on this site uses the same criterion colours, abbreviations, and order. Colour marks the criterion group; the abbreviation and name identify the criterion. Scores are 0–3 per criterion (4 is exceptional; cross-EIP interactions is uncapped).

EVM surface

Opcodes, precompiles, and system contracts that are added or modified.

  • Added opcodes
    Introduces new opcodes
  • Modified opcodes
    Modifies pre-existing opcodes
  • Added precompiles
    Introduces new precompiles
  • Modified precompiles
    Modifies pre-existing precompiles logic or gas-accounting
  • Added system contracts
    Introduces new system contract, stateful or not
  • Modified system contracts
    Modifies pre-existing system contracts

Gas and accounting

Execution, blob, and state gas rules, refunds, and where charges happen inside opcodes.

  • EVM Gas rule changes
    New EVM gas accounting rules
  • Blob gas accounting changes
    New Blob gas accounting rules which potentially affect pre-existing tests
  • New EVM gas refund
    New gas-refund mechanism

Blocks, transactions, and encoding

Transaction types and validity, block and header fields, encodings, syncing, and activation-time changes.

  • New transaction types
    Introduces a new transaction type
  • New or modified transaction validity mechanisms
    Creates new or modifies pre-existing transaction types' validation mechanisms
  • New block / header fields
    Introduces new block or block header fields
  • Encoding changes (RLP/SSZ)
    Introduces encoding changes at the transaction/block/interfaces level
  • Block syncing changes
    Modifies block RLP validation mechanisms that require test client syncing.
  • New fork activation mechanism
    Modifies state, internal variables, or similar, at the fork activation block

Client interfaces

Engine API and transition-tool interface changes.

  • Engine API changes
    Introduces new fields to the Engine API directives
  • Engine API encoding changes · Checklist revision 1 only
    Engine API encoding changes (the revision-1 template defines no anchor text for this row).
  • Transition-tool interface changes
    Modifies or adds new fields to the transition tool interface.

Testing impact

Rework, new invariants, and new primitives required in the test framework.

  • Patterns affecting pre-existing tests
    Implements a new validation mechanism or rule that translates in reworking pre-existing tests

Risk and validation

Security, performance, boundary conditions, and cryptography that need validation.

  • Security risks
    Introduces or modifies mechanisms that could compromise the security of the chain, users, validators, or other stakeholders, if not implemented properly.
  • Performance risks
    Introduces or modifies mechanisms and requires performance validation.
  • Edge/boundary conditions
    Feature contains edge/boundary conditions.
  • Cryptography
    Introduces new cryptography mechanisms or modifies existing functionality that involves cryptography

Coordination

Cross-EIP interactions and behavior that clients must agree on before tests exist.

  • Cross-EIP interactions
    Introduces or modifies mechanisms that affect other EIPs in either the same or past forks.
Download criterion scores (CSV)Open a Human vs LLM pair in the comparison viewRead the methodology and limitations