Rows are reigns, newest first, read straight from the signed round record: the miner that took the lead, when, and by how much it beat the previous leader in a same-round head-to-head. First-time marks a miner taking the lead for the first time. Miner numbers are network UIDs; Entry links to the exact submitted model; Round is the signed record that awarded the win. Miner UIDs can be re-used by the network over time, so the first-time/returning badges are matched on the miner’s cryptographic key, not the UID.

Every round, the subnet tests entries on a fresh set of hidden data. That stops anyone tuning their entry to the test — but it also means a score from round 1 and a score from round 80 were taken on different data and can’t be compared head to head.

What can be compared is each change of lead: the old leader and the challenger are tested on the exact same data, in the same round. So we record how much better the winner was at each handover, and multiply those steps together. Difficulty of the data cancels out at every step.

The guaranteed minimum uses the cautious end of each measurement instead of the best estimate — the same threshold the subnet itself required before awarding the win. Reality sits between the two lines, and the gap is the honest margin of error.

improvement = 1 − Π(challenger_error ÷ leader_error) across lead changes; the minimum line substitutes each round’s 95% bootstrap lower bound.

A step on this ladder counts as proven only after the whole competition — training, screening, hidden-data evaluation and a signed verdict — has actually run at that size. Planned sizes are the roadmap, never shown as achieved.

Cascade trains the same Toto2 architecture the official Datadog release uses — the layer shapes are pinned against the official checkpoint’s own config — from random initialisation, so no official weights are reused. At each ladder size the comparison is therefore: same model, same size, different training data.

Both sides are scored by the identical benchmark battery (GIFT-Eval, BOOM, TIME; CRPS + MASE; pinned data revisions): the official checkpoint is run through the same sidecar as every round checkpoint and published as a reference doc (benchmarks/reference-<size>.json), so no number here is transcribed from a leaderboard. The percentage compares six-score geometric means (lower is better).

The fight is deliberately uneven: the official release was trained on its full corpus at full budget, while every Cascade checkpoint is trained under the round’s fixed, published compute budget. And as everywhere on this page, these scores are report-only — rounds are still settled on hidden data, never on public benchmarks.

Every round trains under a fixed, published compute contract: a set token budget per checkpoint (hours × reference throughput), at a set global batch size and context length. Steps per round is just that budget divided by tokens per step — nothing here is measured off a training log; it is the contracted spend.

The champion’s cumulative pretraining counts the latest round plus every consecutive earlier round whose training was warm-started from the previous round’s promoted checkpoint (recorded on each round’s signed manifest). A round that trained from scratch resets the chain. Until carry-forward is switched on, every chain is exactly one round long — that is stated, not hidden.

Each ladder size gets its own bar because Cascade’s goal at every size is a full Toto2-scale pretraining: the target behind each bar is the official release’s own step count at that size, transcribed from its published training recipe. Until a size’s figure is transcribed its bar shows the banked steps with the target marked pending — never a guessed percentage. A size with no judged rounds yet reads “not started”.

GIFT-Eval, BOOM and TIME — three open forecasting benchmarks anyone can run, scored the same way their public leaderboards do, so the numbers are directly comparable to published results. 1.00 is the Seasonal-Naive baseline — the standard yardstick these leaderboards divide by: a forecast that simply repeats the most recent seasonal pattern. Below 1.00 beats it.

Each benchmark reports two scores. MASE measures how accurate the central forecast is; CRPS measures how well the model’s uncertainty range matched what actually happened. Both are relative to the baseline at 1.00, and lower is better for both. The large number in each tile is the MASE.

These scores are reported only. They never decide who wins a round. The benchmark data is public, so scoring on it would just reward entries tuned to it; the competition is settled on hidden data instead.

A trainer-signed report (benchmarks/round-<id>.json) is published after a round’s duel when the battery completes for at least one checkpoint. Some rounds have none — the battery can fail, or a report can land late — which is normal; the page probes the recent record for the reports that actually exist.