Testnet on Allora has never been just a sandbox. It is the training ground where inference workers prove they can forecast well enough to matter, and until now, "good enough" was not defined anywhere the network could act on automatically. Allora is introducing a promotion and relegation system: a per-topic, per-worker framework that decides, on statistical evidence, who moves from testnet to mainnet, and who moves back.
Why the network needs this
A worker's skill is not universal. A model that forecasts BTC's 8-hour return well might be mediocre on a 24-hour horizon, or on an entirely different topic. Promotion and relegation are evaluated per worker, per topic, because participation, forecast horizon, and performance are all topic-specific. A worker can graduate on one topic while still training on another.
The system is also deliberately asymmetric. Promotion asks for statistically supported evidence of skill, calibrated to how long a topic's forecast horizon is, since long-horizon topics simply generate independent evidence more slowly. Relegation asks for a much higher bar: strong evidence of failure, at a fixed 95% confidence level, regardless of horizon. Getting onto mainnet should be earned. Getting removed should be hard to get wrong.
What gets measured
Every worker is scored on seven metrics, all evaluated in log-return space:
- Directional accuracy — did the forecast get the sign of the move right?
- Pearson correlation — did the size of the forecast move together with the size of the real outcome?
- WRMSE improvement — how much better the forecast is than a flat, zero-change baseline, weighted so large moves count more than quiet ones.
- WCZAR improvement — the same idea, using a loss function that is more sensitive to direction.
- Log aspect ratio — whether the forecast's variation is scaled correctly. Too tight, and it is underconfident; too wide, and it is noise.
- Normalized mean offset — whether the model is systematically biased in one direction.
- Participation — how consistently the worker actually shows up to forecast.
None of these are graded on raw counts. Every metric is evaluated through a confidence interval, adjusted for how much independent evidence a worker has actually produced. When a topic's epoch length is shorter than its ground-truth lag, consecutive forecasts overlap and are not fully independent, so the framework calculates an effective sample size from the autocorrelation of returns rather than just counting submissions. A worker who forecasts consistently gets credit for it; one who forecasts sporadically does not get to inflate their record by volume alone.
Getting to mainnet
Promotion requires passing every one of seven tests on a topic: a minimum effective sample size, a directional accuracy lower bound above 50%, positive correlation, positive improvement over baseline on both loss metrics, a correctly scaled forecast variance, and participation above 90%.
Because long-horizon topics accumulate evidence more slowly, the bar adjusts with horizon length. A 20-minute topic is held to a 95% confidence standard; a 24-hour topic is held closer to 71%, not because the standard is lower, but because demanding 95% confidence on a slow-moving topic could take an unreasonable amount of time to ever satisfy.
Staying on mainnet
Relegation does not fire on a single bad week. It requires one of four conditions, each demanding strong, joint evidence of failure:
- Underperformance across all four predictive metrics at once, not just one noisy signal.
- A forecast scale that is off by roughly an order of magnitude or more.
- Bias so extreme it survives even after passing the scale test.
- Extended inactivity: no submissions for 14 days, or participation under 50% over the trailing 30 days.
Newly promoted workers get a 30-day grace period before performance-based relegation applies, and no topic can be relegated below five active workers; if it would, the relegations are cancelled and the topic is flagged for promotion review instead. The goal is not to punish noise. It is to remove workers who are demonstrably, persistently underperforming, while keeping every topic healthy enough to keep running.
What this looks like for builders
The framework is backend logic, but it is changing what the product looks like too. Every topic and worker page will carry a twin strip linking its testnet and mainnet equivalent, so a builder training a model on a testnet topic can always see, and jump to, where that work leads on mainnet, and back. Mainnet topic pages get a new stat row: promoted workers, emissions per epoch, rewards, and an economics view of fee revenue, topic weight, and stake. Mainnet worker pages replace the promotion gates with a "staying on mainnet" panel, the same criteria tracked on a rolling window, plus an earnings view.
Testnet stops being a side door. It becomes the one documented path onto the network, and the product is being rebuilt to show that path clearly.
Get started and deploy your workers on testnet today.
