King-of-the-hill subnet. Miners upload checkpoints privately to a validator-controlled bucket
SN 120. Rank 4 of 32 by emission
Affine
Emission6.0%of the block reward · rank 4 of 32
Participants5 miners · 3 validators
Registration0.8459 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@d41cca5 · read 2026-09-22
The work
A trained checkpoint: the saved weights of a coding model, uploaded to a private bucket and signalled on chain. The subnet pays whoever holds or recently held the crown, so the thing produced is a model that beats the reigning king in a duel.
own docs · affine.io
King-of-the-hill subnet. Miners upload checkpoints privately to a validator-controlled bucket
How it is scored
One duel, challenger against king, over 1,000 turns of a coding corpus. Since 2026-09-18 each turn scores min(z_R, typ_c, z_A): how far the miner's thought-to-action lift, its thought's typicality on content tokens, and its action given the thought sit from three teacher samples, in teacher standard deviations. A forfeit scores -12. The challenger crowns if its mean margin clears max(2 standard errors, 0.2 sd), the thought floor and causality gate.
code · raw.githubusercontent.com · affine/evalsrv/sdmeter.py
turn = min(z_R, typ_c, z_A); forfeit -> forfeit_sd
Sources disagree The score.py header at this commit still calls min(R, G) the live rule and defaults score_mode to reason; affine.toml sets score_mode sd_min_rga, evalsrv/dueling.py applies it as the deciding rule, and llms.txt on 2026-09-22 agrees with the toml: min(z_R, typ_c, z_A), 1,000 turns, 0.2 sd.
How the pay splits
Not proportional. Every crown won in the last 72 hours gets one equal share of the miner emission, whether or not it still holds the throne. A crown older than 72 hours earns nothing. When no crown is inside the window, the share goes to the burn UID. Effective 2026-09-14.
own docs · affine.io
a crown is paid for at most 72 hours; every crown inside its window gets one equal share
Who judges
Three validators were active on the 2026-09-09 chain snapshot, of 64 permitted. Weights are committed then revealed, at most once per 100 blocks. The validator does the work itself: it rents a GPU pod, loads both checkpoints, runs the whole duel, and seeds it from the reveal-block hash so anyone can replay the verdict. Nothing is sampled.
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 3
How often it turns
Emission settles every 360 blocks, about 72 minutes. The puzzle turns per duel, not per epoch: a challenger is evaluated when it reaches the front of the queue, and a crown stands until a challenger clears the bar. The snapshot on 2026-09-19 showed reign 20. 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; api/v1/snapshot reign_number 20, crowned_at 2026-09-19
Who buys
Nobody is named. The front page offers a chat box with the reigning model, and every crowned checkpoint is free to download with a signed manifest. There is no price, no customer and no usage figure on the site, in llms.txt or in the repository. The miner's income is emission.
our reading · affine.io
Talk directly to the reigning SN120 king
What has been gamed
The program records its own hacks. A king crowned on a constant filler suffix benched 0 of 50, which forced min(R, G) and later the sd-meter. Reign 13, a re-upload of reign 12's weights, won on noise and was uncrowned. Defences: one evaluation slot per hotkey, burned at enqueue; a challenger holding every king safetensors digest is rejected as a copy; a published exploit audit after each crowning can revert.
code · raw.githubusercontent.com · affine/affine.toml
wvk-9 king crowned on a constant filler suffix and benched 0/50
4 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. All 16 quotes in the map.
Their subnet
An incentivized RL environment that became a king-of-the-hill distillation contest: private checkpoint uploads, a frozen teacher, one crowned king, a 72-hour paid window, and published, replayable verdicts.
Affine is building an incentivized RL environment that pays miners for incremental improvements on tasks like program synthesis and coding.
Operating on Bittensor's Subnet 120, we've created a sybil-proof, decoy-proof, copy-proof, and overfitting-proof mechanism that rewards genuine model improvements.
Our vision is to commoditize reasoning—the highest form of intelligence—by directing and aggregating the work of a large, non-permissioned group on RL tasks to break the intelligence sound barrier.
Affine's strength comes from its open, global network of miners, validators, and contributors competing to advance RL.
A direct market for engineers to upload open models that advance the frontier on RL tasks – and get paid for it, secured by the Bittensor blockchain and instant crypto payouts.
Bittensor, in their words
Bittensor is described in the founder's own vocabulary: an open-source protocol for a decentralised machine-learning network, incentive-based computer networks, about 128 subnets collaborating and competing, where incentives amplify good supply and eliminate poor supply.
Once Chinese miners enter a subnet, competition immediately becomes much fiercer, to the point that many of the original participants choose to exit.
about 128 subnets, each collaborating and competing around different tasks such as inference, training, reinforcement learning, code agents, storage, prediction/trading signals
…launching a subnet and compete, good supply is amplified by incentives, poor supply is naturally eliminated
Jacob Robert Steeves is the co-founder of Bittensor, an open-source protocol supporting a decentralized machine-learning network built on blockchain technology.
Frames market competition network commodity crown environment king of the hill open system sound barrier throne
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN120 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 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
- 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 sourcesecondhandunchecked
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 sourcesecondhandunchecked
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
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 d 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 28 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.