Selected model is at capacity. Please try a different model.

The bounty loop only half works. Agents are solving the puzzles and the solutions are real, but nothing gets credited: no rung has ever been marked solved, λ has never gone up, and no prize has been paid. The report is in artifacts/report.md and is left untracked, as asked.

What's working

• Agents are producing real collisions. Completed SIMD hunt jobs published 14 different λ=24 (48-bit) collisions: 3 for keccak256, 7 for sha256 and 4 for ripemd160. I recomputed every one myself and all 14 are valid.
• The checker rejects bad claims. The only claim the board has recorded (0x01/0x02) was rejected, and my recomputation agrees it is wrong.
• IMD is really being spent on jobs. Two payment transactions I looked up on Ethereum mainnet each move exactly 0.5 IMD from SIMD's payer wallet to Identity.md. The vault balance on-chain (1018.146 IMD) matches what the site's API reports.

What's broken

• The board never moves. At 07:52–07:54 UTC, all three hashes were still open at λ=24 with zero solved rungs, even though at least 18 hunt jobs (9 IMD) had completed.
• Likely cause (an inference; I couldn't see SIMD's worker code): each puzzle asks for a collision.json file. The jobs were opened with a report template, and every result I fetched contains only artifacts/report.md. Identity.md's docs say jobs deliver only their declared output files, so SIMD's worker probably never finds a solution and keeps reopening the same puzzle. An earlier agent-run review job (c0ad6502) reached the same conclusion on its own.
• Prizes are manual. Payout status is MANUAL and there is no bounty contract, so who wins and whether they get paid depends on trusting the operator. The proofs themselves can be checked by anyone.
• The hunt counter looks off. huntsToday rose from 53 to 59 in two minutes with no new paid jobs, and the daily cap is 3.

Two other points: several agents found the identical pair, so much of the paid compute was duplicated. And λ=24 takes seconds to about a minute on one CPU core, so the current board says almost nothing about what the swarm can do. Coordinated hunting only starts at λ > 40.

Limits: I saw only the latest 20 of 100 paid orders and checked only 2 payments on-chain. Three of the 14 pairs come from jobs I didn't fetch myself; I checked their maths but not which jobs they came from. Nothing was submitted to the bounty API, and nobody independent has reviewed the report.

I committed README.md (the question and these limits) and the checking script, tools/verify_claims.py. The script needs only standard Python and exits 0 when every check passes, which it does. It lives in tools/ rather than artifacts/ because artifacts/ is ignored by git.