cascade

SN91 cascade is a school where contestants compete to write the textbook, a single fixed student model studies each one, and the textbook whose student forecasts real data best takes the throne.

The commodity, explained from zero

A time-series model forecasts the next values in a stream of numbers: electricity demand, sales, prices. A foundation model is one trained to do that on any stream, not just one. cascade's output is such a model, a small one called Toto2-4M with 4 million parameters, saved as a checkpoint: the file you would download to run it.

The model is fixed and every training run is identical except for one input: the synthetic data it learns from. Miners write programs that generate practice data; the README's bet is that "high-quality synthetic data is critical for training a time-series foundation model." So the textbook is the entry, and the student is shared.

Who would pay, the materials do not say. The README describes a first phase and a roadmap to larger models, but names no buyer, no price, no hosted model, and says nothing about adoption. This is where the school comparison breaks: in a school the textbook is the product, but here the product is the trained student, and nobody has said who wants to hire it.

Why it is on Bittensor at all

On the materials, decentralization adds two things: anyone can enter without a GPU ("Miners need no GPU"), and outsiders can check every round: the audit doc says the machinery exists "to let anyone re-derive the owner's published work instead of taking it on faith." The training itself is not decentralized; the owner runs the one trainer on fixed hardware. Cost, censorship, and coverage are not discussed.

How the work gets done

Miners write a data generator in Python, prove it produces identical output from the same seed, and commit an encrypted pointer on chain that unlocks minutes before the deadline, so rivals cannot copy it. The owner's trainer gives every generator a one-hour "heat" run, keeps the top finishers, then trains on the king's and best challenger's data for three hours each under identical seeds. Validators, the judges who decide pay, score both on private held-out data with two error measures and dethrone the king only when a statistical lower bound on the challenger's improvement clears a margin that shrinks from 2 to 0.5 percent as a reign lengthens. Emission, the subnet's share of new tokens, goes to the current king and up to four prior kings; the README and llms.txt say the split halves each step (about 52, 26, 13, 6, 3 percent) while the architecture doc says an equal share. Each round takes about twelve hours.

How you would know it works

No artifact a reader can check exists on the subnet's own pages, in those words. The chain URL lands on testnet.cascadesub.net, which shows only a title. Signed round receipts are published to a file store and a cascade-audit tool can replay them, but that needs the command line.

What is missing

The site is empty. The chain description is a tag, "SOTA Time Series Foundation Models", and the contact address appears on no page. There is no product, price, customer type, or support path. The miner guide points at the testnet registration command. No page is written for stakers. The two reward-split accounts disagree.

Go deeper

• README: thesis, roles, roadmap
• Architecture: rounds and scoring
• Audit: replaying a round

Metaphor: a school where the contest is over who writes the textbook. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.