I found a collision and wrote it to collision.json. OpenSSL confirms both hashes agree on the first 48 bits.

{"algo":"ripemd160","lambda":24,"inputA":"863122ce045e","inputB":"149b597971fc"}

• How it was found: a low-memory cycle-finding search, artifacts/rho_search.py. It ran in about 45 seconds and repeatedly hashes 12-character hex strings, keeping the first 48 bits. The cost matches the expected ~2^24 hashes for a 48-bit output, so this relies on no weakness in RIPEMD-160 itself.
• How it was checked: I recomputed both hashes with OpenSSL, separately from the Python search. I also checked Python's RIPEMD-160 against the standard "abc" test value.
• Not checked:
  • I didn't run the SIMD verifier.
  • The inputs have no 0x prefix, so they should be read as UTF-8 text. If the verifier decoded them as hex bytes instead, the collision would not hold.
  • I cited the source for the "abc" test value from memory and didn't re-fetch it.

artifacts/report.md has the full write-up: the answer, the evidence, and what was inferred versus unknown. artifacts/README.md describes the question and its limits. I didn't commit anything: the repo's .gitignore excludes artifacts/, and I didn't force-add it. All the outputs are left untracked for the daemon to upload.