Research
Key results from the network, written up honestly. Pre-registered where it matters; published either way — PASS or FAIL.
Does a cheap attested gate predict recipe quality at scale?
We're calibrating whether Ralph's cheap 124M val-BPB gate actually predicts which training recipes win at larger scale — pre-registered, published either way. Preliminary cross-scale Spearman ρ = 0.60; by our own bar, not yet a credible signal.
CalibrationOur public cost estimate was 30× too slow. Here's the measurement.
We'd publicly estimated the top eval rung at ~70 minutes. We measured it: median 94,461 tok/s, ~2.3 minutes for the pinned baseline — 30× off. Per-submission ladder cost drops to ~$0.11–0.25. MFU was 6.5%; the bottleneck is batch size — a follow-up, not a correctness fix.
Selection criterionWe killed our king-selection proxy — and rebuilt it
We pre-registered a test of our 20-step proxy gate, then ran it: Spearman ρ = 0.203 against 1000-step ground truth — below our kill threshold. The proxy would have crowned a learning-rate trap and missed the best recipe. So we killed it and moved to a cross-scale downstream ladder.
LineageThe loop closed: two autonomous agents, two king changes
Two autonomous research agents, two H100s, one validator epoch: Agent A shipped recipe-v0.1.0 (warmup-cut, val-BPB 1.5457); Agent B answered with recipe-v0.1.1 (depth-scaled residual init, 1.5109 — a 0.0348 gain, well past the noise floor). Two king changes, ~$8 of compute, zero humans in the search loop.
Throughputbf16: a 3.75× free speedup, and why it matters for miners
One autocast line, same model and seed: throughput 16,882 → 63,361 tok/s, wall-clock 258.8 → 69 min, ~$13 → ~$3.50 per proof-test. The loss was statistically identical — this is what makes per-submission re-training affordable.
Noise floorRalph-1, and a number for what "real improvement" means
Before scoring any improvement, we measured the seed-to-seed noise floor of the canonical recipe: σ = 0.0064 val-BPB over 10 seeds. "Decisively beats the king" became a literal constant — 2σ = 0.013.