What the daily stream is.
Every day the changes made across connected Google Ads accounts go out as one mixed basket. An admitted model predicts the outcome of each change that day, before any outcome exists. Outcomes are measured at 7, 14 and 28 days, then held for a further seven days so late-reporting conversions can land — so a score, once earned, is final and is never quietly re-measured.
Because a different day's basket matures every day, scores arrive every day. Weights follow a moving average and update daily; emissions flow continuously. The weekly epoch, and the wait for the single score that mattered, are gone.
The mechanism that ran until 3 August 2026 is described on the weekly-era pages, which stay published as the historical record.
The calendar.
A basket settles at its own date, plus one day, plus the horizon, plus the seven-day settling window. The first live basket covered Monday 3 August, which fixes the three dates below. From 18 August a fresh 7-day cohort settles every day, each basket on its own clock.
| Date | What happens | Alpha hold | Burn |
|---|---|---|---|
| Tue 4 Aug 2026 | First live daily basket delivered, covering Monday 3 August. | 0 | 45% |
| Mon 10 Aug 2026 | Bridge begins. Carried-over weights are paid only to miners delivering daily predictions and meeting the hold. | 150 | 30% |
| Tue 18 Aug 2026 | First 7-day scores settle and begin feeding payouts. | 300 | 30% |
| Tue 25 Aug 2026 | First 14-day scores settle. | 450 | 15% |
| Tue 8 Sep 2026 | First 28-day scores settle. | 700 | 15% |
| Tue 15 Sep 2026 | Terminal hold in force. Steady state. | 1,000 | 0% |
Burn figures are a published plan and may be adjusted, up or down, to protect alpha value for holders. The alpha hold is a minimum balance, not a purchase: there is no upfront buy-in, and earnings can fill it.
How a miner joins.
-
1
Register
Register your hotkey on the subnet and bind an ed25519 signing key. One time.
-
2
Train
Train on the published episodes — the inputs — and their settled outcomes — the labels — or the pre-joined training bundle. Outputs are P10/P50/P90 distributions at 7, 14 and 28 days: calibrated ranges, not point estimates.
-
3
Containerise
Package the model as a container that reads episodes on stdin and writes predictions on stdout. No network, 1 GB of memory, 15 CPU-minutes per basket.
-
4
Commit the digest
Push the image and commit its digest on chain. The subnet pulls that exact digest, admits it through the backtest gate, and runs it against every daily basket.
You do not post predictions. Your container is executed by the operator against each day's basket, and its output is locked before any outcome exists.
The step-by-step guide, with every command, is the miner quickstart; the training guide walks the model side end to end — join the data, train, containerise, and score yourself before you ship. Both are in the public repository.
Where the data is.
Everything you train on is public and lives in the repository. Episodes are the inputs you predict on; outcomes are the measured labels, published after each horizon settles. They pair one-to-one by epoch id. Start with whichever is easier: the raw corpus, or the pre-joined bundle.
-
Published corpus
Inputs and labels as separate files, paired by epoch id.
-
Training bundle
Settled episodes, pre-joined, all three horizons — one JSON object per line.
-
Training guide
The full path: join, train, export weights, containerise, self-score.
The bundle is held out from scoring, so training on it cannot leak a score. As the daily stream settles, each scored day is also published in full as a signed receipt you can train against — see verifying a day. The training guide names the exact commit each copy was taken from.
Does it work?
The honest test of a prediction subnet is whether the best model gets better, and whether it holds up as the horizon lengthens. This is that record, week by week, and it is drawn from the same signed receipts a miner uses to check their own scores.
No scored days have published yet. The first 7-day outcomes settle on 18 August 2026, and the first point appears the week after.
You can check the arithmetic.
Every scored day is published as a signed receipt — the settled outcomes used, every prediction exactly as the model produced it, each score's components — and the feed is anchored on chain. You can recompute your own scores from it and check they match ours. How to verify a day →
Canonical documents.
Each is mirrored on this site for reading and linking, and each mirror names the commit it was taken from. The public repository is the source of truth; where the two disagree, the repository and the chain win.