I found a 48-bit collision for keccak256 and checked it with two separate implementations.

collision.json:

{"algo":"keccak256","lambda":24,"inputA":"imd-9804673","inputB":"imd-17368822"}

Both inputs are UTF-8 strings. Their full digests share the first 48 bits, 2f72fe059bc2:

• imd-9804673 → 2f72fe059bc2493e…2b7be703
• imd-17368822 → 2f72fe059bc2f734…a5f1d550

How it was found and checked:

• Search: tools/collide.c is a C birthday search. It found the match after 17,368,823 hashes (about 2^24), in about 96 s on one core. That is in line with the expected birthday cost of about 2.1·10^7.
• C check: before the search, the C code reproduced the published Keccak-256 results for "" and "abc".
• Python check: tools/verify.py is a separate pure-Python Keccak. Its core function matches Python's built-in hashlib.sha3_256, and it reproduces the same two published results. Run on collision.json, it recomputed both digests and reported 48-bit prefix collision: True.

Assumptions (listed in artifacts/report.md):

• Hash variant: "keccak256" here means the original Keccak padding used by Ethereum, not the standard SHA3-256. If the SIMD verifier actually computes SHA3-256, this pair will almost certainly not collide.
• Truncation: "first 48 bits MSB" was read as the first 6 bytes of the digest.
• Input encoding: inputs without a 0x prefix are treated as UTF-8 text.

collision.json, README.md and tools/ are committed (2dc9e3b). Git had no author identity configured, so I set your name and email on that one commit only and did not change any git config. artifacts/report.md is left untracked for the daemon to upload; artifacts/ is listed as ignored by Git.