cascade is building state-of-the-art time-series foundation models (TSFMs) on Bittensor.
SN 91. Rank 18 of 32 by emission
cascade
Emission1.7%of the block reward · rank 18 of 32
Participants5 miners · 7 validators
Registration0.1 TAO
Verified2026-09-09links checked and scores merged
The programforecasting · github.com@7ebfc19 · read 2026-09-22
The work
A data generator: Python code that, from a shared seed, yields synthetic time series to train a small forecasting model. The miner ships code only, no weights and no GPU, and commits an encrypted pointer on chain that unlocks before the round boundary. What is paid for is a generator whose data trains a Toto2-4M forecaster that beats the king's on private real-world windows.
own docs · raw.githubusercontent.com
miners write the synthetic data generators that train it. Better data trains a better forecaster
How it is scored
King and challenger each train a Toto2-4M from a shared init on their generator's data. Validators score both on the same private held-out windows by the geometric mean of CRPS and MASE, lower is better. A paired bootstrap of 10,000 resamples gives a one-sided 95 percent lower bound on the challenger's relative improvement; it wins if that bound clears the margin. An uncomputable bound is inconclusive and the king holds.
code · raw.githubusercontent.com · cascade/eval/koth.py
The challenger *wins the round* iff that LCB clears the win margin
How the pay splits
Geometric decay across the lineage. The current king plus up to 4 prior kings still registered are paid in proportion to 0.5 to the power of their distance from the throne, about 52, 26, 13, 6 and 3 percent. A deregistered prior king drops out and the rest renormalize; with no valid king everything burns to UID 0. The dethrone margin is 0.01 fresh, 0.005 after 8 rounds.
code · raw.githubusercontent.com · chain.toml
reward_prior_kings = 4 ... king_decay = 0.5 # geometric decay across [king, prior kings]
Sources disagree docs/ARCHITECTURE.md at the pinned commit still says the validator sets an equal share across the current king plus up to reward_prior_kings prior kings; chain.toml at the same commit sets king_decay 0.5, cascade/validator/loop.py passes it to decayed_share_vector, and the README and llms.txt give about 52/26/13/6/3 percent.
Who judges
Seven validators were active on the 2026-09-09 chain snapshot, of 64 permitted; the margin rollout doc counts 6 external validators. Weights are committed then revealed, at most once per 100 blocks. Training runs on the owner's trainer on rented pods under one shared seed; each validator downloads the king's and finalists' checkpoints, scores them itself on private windows, runs the bootstrap verdict, and publishes a signed receipt cascade-audit can replay.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled True, weights_rate_limit 100, max_validators 64; chain-2026-09-09.json active_validators 7
How often it turns
Emission settles every 360 blocks, about 72 minutes, but the puzzle turns every 3,600 blocks: a round of about 12 hours with boundaries near 08:30 and 20:30 UTC, an encrypted reveal 25 blocks before the boundary, a one-hour heat and a five-hour final. Rounds without funded seats run nothing and the king holds. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000
Sources disagree The README at the pinned commit announces rolling intake next: 3 hour settlements against a king trained once per 12 hour era, switched on when 51 percent of validator stake signals readiness on chain; chain.toml at the same commit keeps epoch_blocks 3600 with rolling_from_block and era_king_from_block at 0, and whether the on-chain lock-in has fired was not checked.
Who buys
Nobody. The chain URL cascadesub.net did not resolve on 2026-09-19; testnet.cascadesub.net serves a scoreboard of rounds, reigns and model error with no product, price, API or customer. The README describes the output as a trained forecaster checkpoint, and every crowned king's generator code is published. The miner's income is emission.
our reading · testnet.cascadesub.net
King-of-the-hill · compete on synthetic data · Bittensor
Sources disagree On 2026-09-22 cascadesub.net resolves and redirects to dashboard.cascadesub.net, the mainnet scoreboard of rounds, reigns and model error; it still shows no product, price, API or customer.
What has been gamed
Copying a rival's generator: exact-identity dedup on the file tree, the normalized token stream and the stream with names masked, plus a behavioral probe for identical output; the earliest reveal keeps the slot and dropped copies still burn the hotkey. A similarity threshold was tried and removed after operators spaced variants just under it and it produced one false drop. Jitter-duplicated channels: a 0.999 correlation gate, in shadow.
code · raw.githubusercontent.com · cascade/interface/dedup.py
a ratio bar is gameable by construction
5 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
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
Sources for this explainer
- https://raw.githubusercontent.com/TensorLink-AI/cascade/main/README.md
- https://raw.githubusercontent.com/TensorLink-AI/cascade/main/llms.txt
- https://raw.githubusercontent.com/TensorLink-AI/cascade/main/docs/ARCHITECTURE.md
- https://raw.githubusercontent.com/TensorLink-AI/cascade/main/docs/MINER.md
- https://raw.githubusercontent.com/TensorLink-AI/cascade/main/docs/AUDIT.md
- https://raw.githubusercontent.com/TensorLink-AI/cascade/main/docs/INTERFACE.md
- https://testnet.cascadesub.net/
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.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 6/7- subnet_name
- cascade
- github_repo
- github.com
- subnet_contact
- chris@tensor-link.com
- subnet_url
- cascadesub.net
- discord
- christensor_49068
- description
- SOTA Time Series Foundation Models
- additional
- not set
Stakers and validators●●●●● 2.0
Should I allocate here? · rank 31 of 32 for this audience
Q1What it is●●●●● 3
cascade is building state-of-the-art time-series foundation models (TSFMs) on Bittensor.
Output is a trained Toto2-4M forecaster checkpoint per round, no unit or price, the chain description matches, the site front page renders empty
Q2Who it is for●●●●● 0
nothing on users, customers, revenue or usage in the README, llms.txt or docs, the README says no statements about adoption or serving, third-party X posts discuss benchmarks not demand
Q3How it resists gaming or fails●●●●● 3
AUDIT.md lets anyone re-derive a round in three tiers with FAIL, WARN and SKIP verdicts, ARCHITECTURE.md gives the paired bootstrap LCB rule and margin decay, MINER.md lists corpus digest dedupe and timed reveals, VALIDATOR.md lists eight failure modes, spread over four technical docs with no front door
Q4Identity and documentation●●●●● 2
identity 6 of 7 but subnet_url redirects to a testnet host that renders only the title cascade SN91, the GitHub homepage field points to an unrelated cipher tool at cascade-woad.vercel.app, discord field is a username, no staker page
Miners●●●●● 3.3
Can I compete, and what wins? · rank 18 of 32 for this audience
Q1What it is●●●●● 4
Yield UP TO n_series float arrays: 1-D (L,) today, or (C, L) once the owner raises max_channels.
DataGenerator class signature, determinism rule, code only repo under 128 MB and deploy commands, one click from the README, took the lower score because the guide runs about 6000 words
Q2Who it is for●●●●● 3
btcli subnet list --network test shows the current registration cost
points at testnet (chain burn 0.1 TAO on mainnet today), no GPU needed for submission, CPU time budget enforced, win condition stated as training a better forecaster than the king on a private held out set
Q3How it resists gaming or fails●●●●● 3
scoring in prose with MASE and weighted quantile loss, a paired bootstrap lower confidence bound and a margin decaying 2 to 0.5 percent, llms.txt says prior kings earn by geometric decay while ARCHITECTURE.md says an equal share, no example numbers, no immunity or deregistration policy
Q4Identity and documentation●●●●● 3
seven step guide with cascade verify, score, deploy, round and heat commands, repo pushed 2026-09-09, no releases, no dates, errors refer to decision IDs in a decisions folder
Buyers and enterprises●●●●● 0.5
Can I use this today? · rank 30 of 32 for this audience
Q1What it is●●●●● 1
no product, API or hosted model, the README describes a first task and a technical roadmap without a plain no product yet statement or a date
Q2Who it is for●●●●● 0
no sourcesecondhandunchecked
no pricing and no customer type in own or third-party materials
Q3How it resists gaming or fails●●●●● 0
no sourcesecondhandunchecked
no status page, SLA or support path, the discord field is a username, no email on any page fetched
Q4Identity and documentation●●●●● 1
no API docs, subnet_contact chris@tensor-link.com is on chain only and appears on no page fetched
Newcomers●●●●● 2.3
What is this and why does it matter? · rank 28 of 32 for this audience
Q1What it is●●●●● 3
Miners compete by writing the data generators that feed the model
in a plain first paragraph, then byte identical contracts and paired bootstrap, chain description is a tag without a verb, no site prose
Q2Who it is for●●●●● 2
let anyone re-derive the owner's published work instead of taking it on faith
is a verifiability claim deep in the audit doc, no comparison with a centralized forecasting model provider
Q3How it resists gaming or fails●●●●● 2
the site is a JS shell showing only the title, the repo has a benchmarks folder, third-party posts cite a GIFT-Eval result, no dated benchmark reachable on own pages
Q4Identity and documentation●●●●● 2
subnet_name cascade matches the page title, description is not a sentence, url resolves to an empty testnet page, no about or learn page
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent b 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 29 live, 0 unreachable, 0 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.