What a full-output reader sees in t8n.

A t8n tool takes a pre-state, a block environment, and signed transactions, then emits the post-state and the artifacts needed to assemble a block. Jaune implements that interface over the same Lean definitions used by its fixture runner and proof stack. Its gate — nine cases when this study was written, thirty-nine now — compares result, alloc, and body against goldens generated by a pinned EELS checkout, with every normalization and deviation explicit.

That gate asks whether Jaune reproduces its declared target. A separate acceptance run asks a different question: what happens when Jaune, the target, and an independent client receive identical signed inputs? On the eight comparable blockchain-mode cases, three were unanimous and five had registered divergences. Jaune agreed with its target in all five.

The root comparison below is part of that automated three-way run. The body is not one of the comparator’s result fields; it was decoded and compared separately on the same signed three-transaction input.

EELS’s consensus transaction processor inserts a transaction into the shared transaction trie before admission checks. That is harmless in consensus execution: an invalid transaction invalidates the whole block, and no mutated output is used. The transition tool adds a non-consensus behavior around it—catch the admission exception and continue—without restoring that trie.

With input 1 rejected — its nonce is too high — the transaction trie contains original input keys {0, 1, 2}. Receipts exist only for the accepted transactions but retain their original keys {0, 2}. Geth instead emits both accepted transactions at contiguous keys {0, 1}.

EELS builds the body as rlp.encode([tx.rlp() for tx in txs]). A legacy transaction’s rlp() is already encoded, so the outer list treats those bytes as a byte string instead of placing the transaction’s RLP list inline. Typed EIP-2718 envelopes really are byte strings in a block body, which is why typed transactions happen to take the right shape.

The execution-spec-tests framework is not trusting these outputs naïvely. Its block-fixture path either replaces the value, constrains the input away from the failing shape, or does not consume the output. Those are sensible defenses. They also mean a general reader of the full transition-tool response exercises a surface the ordinary path does not.

The exhibit is reproduced at the commits it was found on. Jaune df58966 carries the case, its goldens from EELS 9d6e6f83, and the three-way script. Its ordinary gate needs only a built Jaune binary: it reruns every case twice (nine at that commit), checks determinism, verifies every golden digest, and compares result, alloc, and body with all normalizations and deviations explicit. The three-way script additionally needs an EELS checkout at the same pin and a Geth evm binary. It makes the root divergence observable but, deliberately, does not compare body.

The broader lesson is simple: an interface may have many callers while parts of its output still have only one effective reader. Recomputing, constraining, or ignoring fields can protect a downstream pipeline, but it does not validate what the producer emits. Adding another full-output reader makes that otherwise quiet surface observable.