# External conformance checks The compatibility profile is Solidity `0.8.36+commit.8a079791`, Cancun, `viaIR: true`, with optimization disabled and enabled at 200 runs. From this package, run: ```sh python3 scripts/check_conformance.py \ --solc /path/to/solc-0.8.36 \ --lean-bin /path/to/lean-4.29.0/bin \ --output /tmp/new-solidity-conformance-run ``` The output directory must be new. Anvil must be on `PATH`, or supplied with `--anvil`. The runner starts a temporary local Cancun node; it does not access a remote chain. Python's standard library is sufficient. Fourteen manually paired cases exercise aliased operand ordering, aggregate copy timing, compound destination capture, overlapping tuple assignments, internal calls, checked overflow, and rollback. The runner executes the actual Lean machine and checks admission before compiling Solidity. Target outcomes must belong to the finite set of exported permitted schedule witnesses. That set need not enumerate every possible schedule: a mismatch requires examining the pairing and schedule coverage before concluding that the semantics differ. Three static rejection controls compare compiler diagnostics with direct `Admission.checkFunction` rejections and retain runtime model rejection separately. Each control also checks admission with the invalid body in an unreachable branch; raw execution of that variant must return zero, demonstrating why execution alone cannot establish static rejection. Two model resource controls distinguish host exhaustion from source revert. Unsupported execution, failed admission, compiler rejection, RPC errors, host exhaustion, target-resource failure, and runtime revert remain separate outcomes. Target-resource failures have their own `targetResourceFailures` counter and fail the check; the current `Machine.step` corpus does not produce them. Each runtime probe uses `eth_call` against unchanged initial contract storage. Reverting probes additionally run as transactions: their receipt must report failure with no logs, and the selected storage slots must remain unchanged. This is a bounded rollback observation, not complete-world equivalence. The three aggregate-copy probes are also compiled with `viaIR: false`, with both optimizer settings. These are explicit boundary controls: legacy returns length one for construction and internal calls, and panics with `0x32` for the indexed struct fixture. IR returns lengths two and the encoded result 514. Legacy observations must disagree with the IR witness set; they are not counted as supported-profile conformance successes. The run retains source, exact Standard JSON requests and compiler responses, tool identities and hashes, RPC observations, process exit codes, and a report bound to the package inputs. Successful execution establishes evidence for this corpus and configuration. It does not prove the manual source/model pairing, general compiler correctness, or generic safety for every operational feature.