Skip to content
Transfer-credibility
Open experiment · 250M reference complete · June 2026

Does a cheap attested gate predict recipe quality at scale?

Everyone can now run the LLM-pretraining recipe search cheaply. Almost nobody proves their cheap winner actually holds at larger scale — and nobody binds that proof to a run you can verify happened. This is Ralph’s pre-registered test of exactly that, published either way.

01 The question

The question

Recursive, Prime Intellect and ScaleAutoResearch made the recipe search common. The durable asset isn’t the search — it’s a cheap, attested gate whose verdicts you can trust, and the lineage of proven improvements it produces. Call the white space DataDecide-under-attestation: calibrate that cheap gate’s predictive validity, then publish it. The protocol’s credibility is downstream of this being falsifiable, so we don’t assume it; we measure it.

02 The test

The test

The gate is held-out validation BPB at 124M parameters (NanoGPT-Speedrun scale). The reference is the same recipes retrained at larger scale on the identical metric. Transfer is the Spearman ρ — and the pairwise decision-accuracy — between the two rankings, across ~24 config-flag-gated interventions (optimizer, schedule, positional encoding, activation, QK-norm, weight decay, batch/seq/LR sweeps), plus must-die probes like an 8×-too-high learning rate. Attestation is the protocol’s half: a score binds to execution on confidential-compute hardware, so a number maps to a run that provably happened, not one a miner reported. This calibration itself runs on ordinary GPUs — what it tests is whether the cheap gate that attested protocol leans on actually predicts at scale.

03 Measuring cheap first paid off

Measuring cheap first paid off

Downstream accuracy
underpowered
means ~0.30–0.35 vs ~0.31 random
Held-out val-BPB
~27× S/N
separates recipes at the same budget
Muon LR bug
1.99 → 1.75
ran at AdamW’s LR; needs ~30× higher

Our planned gate — downstream accuracy — was underpowered at this budget; switching to held-out val-BPB separated recipes cleanly. We also caught a bug: Muon ran at AdamW’s learning rate and needs ~30× higher; the fix dropped its val-BPB from ~1.99 to ~1.75 — the bug, not the optimizer. Catching these for a few hundred dollars, instead of in a multi-thousand-dollar campaign, is the staged design working. Whether val-BPB’s cheap separation predicts at scale is the open question — that’s what the transfer test measures.

04 The result so far

The result so far

250M reference · complete
Spearman ρ
0.614
cross-scale, graded recipes
Decision-accuracy
0.74
pairwise, vs ~250M reference
Graded
18 / 18
complete · 1 seed

At the ~250M reference, all 18 recipes graded: ρ = 0.614, decision-accuracy = 0.74. The point estimate lands on the pre-registered GO line (ρ ≈ 0.6), but at this n the 95% confidence interval runs down to ~0.20 — below the floor the rule requires. So by our own bar this is not yet a credible signal: not a PASS, not validated, not proven. The extremes agree at both scales — but that’s the easy part; ρ is held down by a tight near-baseline cluster and one apparent discordance: the cheap gate appears to under-penalize short context (a 512-token run), though we can’t yet fully rule out noise. A larger ~1B reference is what adjudicates GO/NO-GO; this ~250M stage was the cheap pre-check, and its full results are public either way — PASS or FAIL.

05 What the outcome buys

What the outcome buys

The decision rule is pre-registered: a ρ threshold, a CI floor, and a decision-accuracy target, computed by a frozen analysis script. If the larger reference returns GO, the attested gate is credible, miners compete on something real, and ralph-diffs — the attested record of which recipe changes help, and by how much — becomes a research asset worth the larger campaign. NO-GO means we publish the honest negative and run a pre-scoped fallback. Either way, you get a verifiable artifact. Prior work makes the premise plausible (DCLM reports cross-scale Pearson ~0.885 at 400M→7B); plausible is not measured, so we measure it.
netuid 40 · Ralph-1 = 253,872,128 params · king lineage recipe-v0.1.0 → v0.1.1 · github.com/RalphLabsAI/ralph