I found a collision for keccak256 truncated to its first 48 bits:

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

Both inputs are plain text strings. Their digests begin with the same 48 bits, 0x2f72fe059bc2:

• imd-9804673 → 2f72fe059bc2493e489ac23cd5bb0119eee739455c6ba49cfd31f75a2b7be703
• imd-17368822 → 2f72fe059bc2f734345293e1223f429e97aea53934f75f186ac1c57ca5f1d550

How it was found and checked:

• Search: a C program (tools/collide.c) hashed imd-0, imd-1, … and stopped when two 48-bit prefixes matched. That took 17,368,823 hashes (about 2^24.05) and roughly 21 seconds on one core, which is in line with the expected cost of about 2^24.
• Re-check: a separately written Python version (tools/verify.py) recomputed both digests and confirmed the match.
• Known values: both programs reproduce the published keccak256 hashes of "" and "abc".

Uncertainty:

• I assumed "keccak256" means the Ethereum-style function, not the standardised SHA3-256. They produce different digests, and this pair only collides under keccak256. If the verifier actually uses SHA3-256, it will reject this pair.
• No ready-made hashing library was available, so both checks rely on code I wrote, tied only to those published values.

artifacts/report.md has the full report, separating facts, inferences and open questions. I put collision.json in the repo root, with a copy in artifacts/. All of these are left untracked for upload; artifacts/ is listed in the repo's local exclude file, so git ignores it. I committed README.md and the two programs. Git had no author name or email set, so that commit uses your name and email.