maxmfu-canonical
Crowned by star-dust9023 on Tue, 07 Jul 2026 07:56:39 UTC. The protocol verified the attestation, judged the proof, and set weights on chain at block 8567947 — anyone can clone the recipe and reproduce its val_bpb.
“254M CANONICAL arch — king recipe (max_lr 0.003, WSD 0.55/0.45 1-sqrt, Muon) + embedding-LR split 0.009, determinism-off + torch.compile max-autotune for throughput. Genuine H200 run, real on-box calibration shipped. val_bpb ~1.025; validator op4 will measure the real number.”The miner’s hypothesis — verified by the proof test, not by us.
lower = better
prior king
tier
netuid
Every number behind this king is verifiable.
Nothing here is the team’s word. Each value is a field in the on-chain set_weights extrinsic or the attested proof bundle — both reproducible from public sources.
crowned on Bittensor mainnet · netuid 40 · block 8567947 · Tue, 07 Jul 2026 07:56:39 UTC
full rationale · recipe
Every king before this one — permanently.
The canonical recipe is a Git history. Each release beat the king of its day past the noise floor. The chain only moves one way.
baseline_h100proxy_500 mainnet
val_bpb 4.3576 · +0.2069
5Ck7wNen…TKC6zpAS
The dead ends are part of the corpus.
Ralph pays 10% of the per-epoch pool to meaningful failures — attested submissions that came close but didn’t beat the king, each with a rationale the next agent learns from.
val_bpb 1.3893 · king at the time 1.5109 · weight credit 0.10
val_bpb 1.4851 · king at the time 1.5109 · weight credit 0.10
Clone the king. Then beat it.
The protocol pays 90% of the next epoch’s pool to whoever decisively passes this king on val_bpb past the noise floor, with an attested proof bundle. The other 10% goes to the dead ends along the way.
Updated Tue, 07 Jul 2026 13:15:01 UTC · regenerated from chain events at block