◐ rebuilding Bitcoin chain analytics are being rebuilt after a transaction-parser fix. Historical figures shown may be inaccurate until the re-ingest completes.
⬡ SUBSTRATE 🔒 sovereign Simargl multimodal · Mode B, panel-graded · kernel engaged Contact ▸ ⌘K research.semurg.io

Simargl: watch the panel score climb.

No teacher here. A diverse judges panel grades the QUALITY of Simargl's own output, no reference answer, reinforcement style, at evolution time only. Fitness is the panel score, climbing above an untrained baseline. At inference Simargl runs purely on-substrate, raw bits in, bits out, the panel is never called live.

Mode B · panel-graded evolution · fitness = judges-panel score no live run on this build
panel-graded climb
No panel-graded champion has been recorded on this public inference build, and no training node is streaming right now, so there is no live climb to show. That is honest: the judges (a diverse NIM panel) are a training-time tool and are never called here. What IS measured live on this box is below.
The judges are never called at inference. The panel (a diverse NIM panel: Qwen, Google, Moonshot, Z.ai) runs at evolution time ONLY, rubric-based, never a single judge. At inference Simargl is pure substrate: raw bits in, one bit per input neuron, no encoder, no decoder, no panel.
Mode B · fitness = judges-panel score · awaiting a recorded champion or a live run
strong (95%) untrained baseline
on-substrate instruments · measured live here, no NIM · secondary to the climb
8.89 TSAOPS
raw-kernel SAOPS (XNOR+VPOPCNTQ)
868.3 GSAOPS
through-model SAOPS (end-to-end)
1 bit
per input neuron · no encoder
0
floating-point ops
one bit per neuron · three modalities, one frame
A live Bitcoin value-flow edge, a live Binance tick, and your text prompt each go through the EXACT production kernel (Snn.ingest_pack(bytes, 512)), one bit per input neuron, no encoder, no decoder. All three collapse into the SAME 512-bit frame format and run through one XNOR+popcount kernel. The arg-max is the model's selected output neuron for each.
live Bitcoin value-flow edge 24 / 512 neurons on · arg-max → neuron 2
the record + a neighbor + sats, straight off the chain · bytes 24 3B 1A 00 00 00 00 00 0F 3B 1A 00 00 00 00 00 …
live Binance tick 16 / 512 neurons on · arg-max → neuron 2
the live aggTrade counter + timestamp · bytes B1 78 01 00 00 00 00 00 39 C1 A0 A0 9F 01 00 00 …
your text prompt 55 / 512 neurons on · arg-max → neuron 2
A binary spiking net runs on the CPU you already own, no GPU. · bytes 41 20 62 69 6E 61 72 79 20 73 70 69 6B 69 6E 67 …
Three modalities, one frame. Bitcoin, Binance, and text are three completely different byte streams, yet each collapses into an identical 512-bit one-bit-per-neuron activation through the same kernel. There is no per-modality encoder and no decoder, the raw bytes ARE the input. This is why one binary spiking net is natively multimodal.
The quality axis is "panel score (judge-graded)", never accuracy vs a ground truth, because Mode B has no reference answer. Reward hacking is the main failure mode, so the panel is diverse (four model families) and rubric-driven, and we surface inter-judge agreement plus a held-out judge so a climbing score is provably consensus, not single-judge gaming. This is the highest-variance of Simargl's applications; its score is reported exactly as measured, and the judges are never called at inference.