SN 120. Rank 4 of 32 by emission
Affine
King-of-the-hill subnet. Miners upload checkpoints privately to a validator-controlled bucket
In their own words · affine.ioSN120 Affine runs a boxing title for AI coding models, where one champion checkpoint holds the belt, each challenger gets a single shot, and the purse is split among the last five champions.
The commodity, explained from zero
The product is a checkpoint: the saved file of a trained model's weights, the file you download to run the model yourself. Affine's checkpoints are trained for one skill, the work an AI coding agent does when it reads a codebase and picks the next command or edit. Only the champion is published; a loser's weights stay sealed in a private bucket and expire after 60 days.
Who pays for the champion is not stated. The front page offers a chat box, "Talk directly to the reigning SN120 king", and every crowned checkpoint is free to download at models.affine.io with a signed manifest. The materials name no customer, price, or use beyond the chat; the money here is the subnet's emission, its share of newly minted TAO, not revenue.
Here the boxing comparison breaks: a dethroned boxer gets nothing, but a dethroned Affine champion keeps a share of the purse while it remains among the last five kings.
Why it is on Bittensor at all
The materials do not compare Affine with a company's own training run. What they claim is who may enter and who may check. Anyone with a TAO-funded wallet can challenge, and "Replayability is the trust model": two checkpoints, the public corpus and the published scoring code let anyone recompute a verdict. The referee is one validator with its own bucket and eval machine; the chain lists three validators, and the materials do not say how they divide the work.
How the work gets done
Miners, the fighters, upload a checkpoint to a private bucket and post two signed notices on chain. The validator, the referee, burns the miner's one evaluation slot (one submission per hotkey, the miner's on-chain identity key, ever) and runs a duel: champion and challenger each answer 1,300 coding turns from a public corpus, and a frozen teacher model (Qwen3.8-27B) scores both on min(R, G): how much the model's written reasoning helps the teacher predict the right action, and whether that reasoning looks like the teacher's own. The challenger takes the belt only if its margin clears a floor of two standard errors or 0.002. Emission is split equally among the last five champions, registered hotkeys only, and a published LLM audit reviews every crowning and can revert it.
How you would know it works
The live snapshot at affine.io/api/v1/snapshot. On 2026-09-10 it showed reign 10, crowned at 02:02 UTC, four challengers queued, and 379 lifetime submissions, 10 accepted and 350 rejected. The front-page chat talks to the reigning model. There is no dated benchmark; the advisory benches the site publishes "never affect Reason or crowning".
What is missing
There is no price list, no named customer and no usage figure. The repository has no README; the only guide is llms.txt, a file written for machines. The chain description, "Reason Mining", appears nowhere on the site. The chain contact hello@affine.io appears nowhere on the site or in the repository, the discord field holds a username, not a link, and there is no support path or status page.
Go deeper
Sources for this explainer
Metaphor: a boxing title fight. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How we scored it
Four audiences, four questions each, scored on what a first-time reader can find in five minutes. Method in the rubric.
- subnet_name
- Affine
- github_repo
- github.com
- subnet_contact
- hello@affine.io
- subnet_url
- www.affine.io
- discord
- consttt dead
- description
- Reason Mining
- additional
- not set
Stakers and validators●●●●● 2.3
Should I allocate here? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
the validator crowns the reigning king by a single teacher-anchored distillation score
output is a crowned checkpoint, no unit or price, one click from the front, machine register
Q2Who it is for●●●●● 1
no usage figure, customer or revenue anywhere in own materials, the front page offers chat with the king and llms.txt says emissions go to recent kings, third-party pages carry no numbers either
Q3How it resists gaming or fails●●●●● 3
How the game works: one eval slot per hotkey burned at enqueue, copies of the king rejected, every dethrone gets a published exploit audit, no failure modes, assumes the reader knows the mechanism
Q4Identity and documentation●●●●● 2
identity 6 of 7, discord field is the text consttt, url has no scheme but resolves, github resolves to a repo with no README (AGENTS.md and START_HERE.txt instead), no staker or validator page
Miners●●●●● 3.8
Can I compete, and what wins? · rank 11 of 32 for this audience
Q1What it is●●●●● 4
If you have a TAO-funded wallet you can mine Affine
submit checklist: checkpoint directory in, signed manifest up, duel verdict out, one file one click from the front page, machine register
Q2Who it is for●●●●● 4
the miner slot is 2 GPUs, tensor-parallel 2
on an 8-GPU B200 class pod, registration cost pointed to btcli burn-cost 120 and the snapshot API (chain burn 0.85 TAO), no plain competitiveness line
Q3How it resists gaming or fails●●●●● 3
min(R, G) score in prose and formula, one submission per hotkey ever, a challenger dethrones the king only above max(2 SE, 0.002), no example numbers, no immunity or deregistration policy
Q4Identity and documentation●●●●● 4
the submit checklist is the miner guide with dated fork notices (wvk 13 posted 2026-09-07, effective 2026-09-09), repo pushed 2026-09-09, no releases, no README so the guide lives only on the site
Buyers and enterprises●●●●● 1.3
Can I use this today? · rank 28 of 32 for this audience
Q1What it is●●●●● 3
Talk directly to the reigning SN120 king
chat box on the front page and crowned weights published under models.affine.io, no product page, no inference API, no example
Q2Who it is for●●●●● 0
no pricing and no customer type on the site, in llms.txt, in the repo or in third-party pages
Q3How it resists gaming or fails●●●●● 0
no status page, SLA, support address or support path on the site or in the repo, the discord identity field is not a link
Q4Identity and documentation●●●●● 2
Swagger UI for the dashboard API only (Affine dashboard, the api/v1 routes listed in llms.txt), subnet_contact hello@affine.io appears nowhere on the site or in the repo
Newcomers●●●●● 2.0
What is this and why does it matter? · rank 29 of 32 for this audience
Q1What it is●●●●● 2
teacher-anchored distillation subnet (validator, evalsrv, research harness)
is the repo blurb, llms.txt opens technically, chain description Reason Mining is a tag, no plain paragraph anywhere
Q2Who it is for●●●●● 1
not an LLM judge
is the only contrast in own materials, no comparison with a centralized service, third-party guides supply the permissionless framing
Q3How it resists gaming or fails●●●●● 3
live JSON snapshot with the current king crowned 2026-09-08 and a queued challenger, dashboard panels read loading to a fetcher, chat with the king, no dated benchmark, no explanation of the numbers
Q4Identity and documentation●●●●● 2
subnet_name Affine matches the AFFINE header, description Reason Mining is not a sentence and appears nowhere on the site, url resolves without a scheme, no about or learn page
How this score came to be. Verified means the scorer fetched the page and the words are on it; inferred means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
Something wrong? Corrections of fact are applied as they arrive during the window; score disputes are batched at its close. How to file one · GitHub issue · email.