Mining Programs

Thirty-two puzzles, written down.

If a subnet is a mining program, this is the program read aloud: what a miner is paid to produce, how a validator judges it, how the pay splits between miners, who the judges are, how often the puzzle turns, who buys the output today, and what miners have optimized instead of the work. Seven fields per subnet. Every field links its source and says what kind of thing the claim rests on: the chain, the code at a pinned commit, the subnet's own prose, a statement on record, or the site's own reading.

32programs read
61%of fields from chain or code
35places the sources disagree
4fields not found
What a field means The dataset, CC BY 4.0 Correct a field
Trust, highest first: chain code own docs on record our reading not found · Seven dots per row, one per field, in the order work, scoring, split, judges, cadence, buyer, exploits.
Rank Subnet The workwhat a miner is paid to produce How it is scoredhow a validator judges it How the pay splitsbetween miners Who buystoday, if anyone Trustwhere each field comes from
1MinosSN107 · science11.3% of emission

A configuration file: the parameter settings for one of the supported variant callers, the software that decides which differences in a sequenced genome are real mutations.

own docs

Validators run each configuration in a Docker container on the round's genome and compare the calls to the planted truth with hap.py, a standard genomics referee.

code

Winner-heavy.

own docs

Nobody yet.

own docs		4 of 72 disagreementsread 2026-09-19
2lium.ioSN51 · compute8.0% of emission

GPU time by the second.

own docs

Two formulas.

code

Three pools, each proportional inside.

code

Renters paying per GPU-second.

own docs		5 of 71 disagreementread 2026-09-22
3OROSN15 · agents6.7% of emission

A shopping agent: one Python file exporting agent_main that, given a shopper's request and budget, searches a catalogue, compares listings and places an order inside a sealed environment.

own docs

Per task, a verifier in the sealed runtime pays a reward from the episode receipt; a validator's run score is the mean reward over its tasks, failures count zero.

code

Winner-heavy with a large burn.

code

Nobody is named.

our reading		4 of 72 disagreementsread 2026-09-19
4AffineSN120 · training6.0% of emission

A trained checkpoint: the saved weights of a coding model, uploaded to a private bucket and signalled on chain.

own docs

One duel, challenger against king, over 1,000 turns of a coding corpus.

code

Not proportional.

own docs

Nobody is named.

our reading		4 of 71 disagreementread 2026-09-22
5ChutesSN64 · inference5.9% of emission

Compute time.

own docs

A miner's score is the sum, over the last day, of the seconds each of its instances was live times that instance's compute multiplier, which folds in GPU class, bounties, urgency, and the confidential-compute and private-chute bonuses.

code

Proportional.

code

Developers and resellers paying per million tokens or per GPU hour.

own docs		4 of 71 disagreementread 2026-09-19
6AlbedoSN97 · training5.2% of emission

A fine-tuned checkpoint of Qwen3.6-35B-A3B, an open 35 billion parameter model, trained further to act as a coding agent: read a real repository, run shell commands, make edits over several turns.

own docs

A duel on sampled coding-task prefixes.

code

Equal shares across a chain of at most five kings.

code

Nobody is named.

our reading		5 of 7read 2026-09-19
7TargonSN4 · compute4.7% of emission

Attested GPU cards.

code

No quality score.

code

Proportional to verified cards in each auction, capped per card.

code

Renters of confidential GPU and CPU machines by the hour, through the site and an unauthenticated inventory API.

own docs		5 of 7read 2026-09-19
8ScoreSN44 · inference3.4% of emission

A hosted computer-vision model.

own docs

Per Element, each answer is scored on pillars (for example overlap with the answer key, player count, smoothness, role), each pillar times its manifest weight, summed.

code

Winner takes all, per Element.

code

Nobody is named.

our reading		5 of 71 disagreementread 2026-09-19
9OpenRobotoSN80 · training3.0% of emission

A fine-tuned robot control model, uploaded whole to Hugging Face.

own docs

A benchmark worker runs each checkpoint through LIBERO manipulation tasks in the MuJoCo simulator, seeded from the commitment's block hash plus a drand beacon.

own docs

Three pools.

own docs

Nobody is named.

our reading		2 of 71 disagreementread 2026-09-19
10SOMASN114 · inference2.4% of emission

Compression code.

own docs

Per task, x is how many baseline runs solved it and y how many miner runs did.

code

Winner takes each contest.

code

Developers routing model calls through SOMA.

own docs		5 of 71 disagreementread 2026-09-19
11TeutonicSN3 · training2.4% of emission

A trained checkpoint of Teutonic-II, a 110 billion parameter mixture-of-experts language model, uploaded privately through a validator-issued credential and signalled ready on chain.

own docs

A paired duel on the same text sequences: per sequence the challenger's loss is subtracted from the king's, a paired bootstrap of the mean difference gives a one-sided lower confidence bound, and the challenger is crowned only if that bound clears delta.

code

Equal shares, not winner-take-all.

code

Nobody.

our reading		5 of 71 disagreementread 2026-09-19
12VantaSN8 · forecasting2.3% of emission

Trading orders.

own docs

Debt-based.

code

Proportional to debt owed.

code

Two kinds.

own docs		5 of 73 disagreementsread 2026-09-22
13iotaSN9 · training2.3% of emission

Forward-pass work on one slice of a language model.

own docs

Validators score each miner's forward-pass activations in short windows on several checks: consistency over time, agreement with peers on the same layer, expected signal shape, and anti-gaming patterns.

own docs

Proportional.

code

Nobody is named.

own docs		3 of 71 disagreementread 2026-09-19
14engySN53 · inference2.1% of emission

Tokens of a named open model, served on the miner's own GPUs.

own docs

Per miner and model, score = floor(sum of prompt plus completion tokens x score_rate[model] / 1000), where score_rate is a per-model rate the operator sets, not the price card.

own docs

Proportional.

own docs

Developers and agents paying per token through one OpenAI-compatible endpoint at api.engy.ai with prepaid billing.

own docs		2 of 72 disagreementsread 2026-09-19
15BitcastSN93 · media2.0% of emission

An original X post that satisfies a brand's campaign brief, by a creator ranked inside a curated influence map, whose hashed draft and later submission a miner platform commits on chain in hash-linked batches for validators to replay.

own docs

A language model passes or fails the post against the brief at temperature zero, any of three checks passing.

code

No value in code or config fixes the 98 to 2 split between the X and YouTube mechanisms; the validator submits only to mechid 1.

code

Brands, none named.

own docs		5 of 72 disagreementsread 2026-09-22
16NOVASN68 · science1.9% of emission

Drug candidates on paper.

own docs

Validators decrypt every entry after the epoch closes and reject any that fails format, the epoch's allowed reaction, uniqueness per target, or a similarity ceiling against past submissions.

code

Winner takes all, twice over: 60 percent of the miner weight to the best molecule entry, 40 percent to the best nanobody, ties to the earliest committed block.

code

No paying customer is named and there is no price.

own docs		5 of 71 disagreementread 2026-09-19
17SayGMSN28 · inference1.7% of emission

Resold access to closed models.

own docs

Dollars served.

code

Cap and burn.

code

Developers paying per million tokens through an OpenAI-compatible gateway at api.saygm.com/v1, signing up by email, Google or Bittensor with no card required.

own docs		5 of 7read 2026-09-19
18cascadeSN91 · forecasting1.7% of emission

A data generator: Python code that, from a shared seed, yields synthetic time series to train a small forecasting model.

own docs

King and challenger each train a Toto2-4M from a shared init on their generator's data.

code

Geometric decay across the lineage.

code

Nobody.

our reading		5 of 73 disagreementsread 2026-09-22
19Green ComputeSN110 · inference1.6% of emission

Inference and GPU rental served from RTX 4090 or RTX 5090 nodes whose power is certified renewable and whose location is proven.

own docs

Unknown.

not found

Unknown.

not found

Enterprise teams paying in ordinary money, plus self-serve renters with a $5 minimum top-up.

own docs		2 of 7read 2026-09-19
20HippiusSN75 · storage1.6% of emission

Bytes held and bytes served.

own docs

Per node: log2(bandwidth bytes + 1) weighted 70 percent plus log2(stored bytes + 1) weighted 30 percent, times uptime, minus 50 points per validator strike and 100 per integrity failure, capped at 50,000.

code

Proportional, of a sliver.

code

Anyone with a credit card or TAO.

own docs		5 of 71 disagreementread 2026-09-19
21greevilsSN58 · agents1.4% of emission

A trading record.

own docs

A pairwise tournament.

own docs

Proportional within two pools.

own docs

Nobody.

our reading		2 of 71 disagreementread 2026-09-19
22ChronoLLMSN38 · training1.4% of emission

A set of twelve language models, one per vintage year from 2013 to 2024, each trained only on text available up to its cutoff so it cannot know later events.

own docs

Stage 1 is a gate: per year the model must recognise at least 70 percent of pre-cutoff items and at most 10 percent of post-cutoff ones, the year average must beat -3.0.

code

Winner-heavy exponential decay.

code

Quants and researchers, on three tiers with no price.

own docs		5 of 71 disagreementread 2026-09-22
23GradientsSN56 · training1.4% of emission

Training code: a GitHub repository and commit, served from a small miner endpoint, that validators build in a container and run on their own GPUs to fine-tune a model on a task.

code

Round one is a single group of every entrant playing three instruct tasks; the top two advance to a knockout, whose winner meets the reigning champion in a six-task boss round.

code

Winner-heavy per pool.

code

Anyone with a model and a dataset, by the hour, through the site or one API call to api.gradients.io/v1/tasks/create.

code		7 of 71 disagreementread 2026-09-22
24DittoSN118 · agents1.2% of emission

A memory harness: the software that decides what an AI assistant stores, finds again and which tools it calls, with the language model held fixed.

own docs

Validators run the harness in a sandbox against a dataset regenerated from a seed every run: tool-calling cases and a memory haystack with distractors.

own docs

King of the hill with a tail.

code

End users of the Ditto app, paying per message in credits after 30 free messages, with no price table.

own docs		3 of 72 disagreementsread 2026-09-19
25SwarmSN124 · training1.2% of emission

A trained flight policy for a simulated Crazyflie quadcopter: a model that reads a depth camera and the drone's state and outputs control, packaged as an archive.

own docs

Every validator flies each submitted policy through the same 1,000 secret seeds per family per epoch and scores each flight 0.45 for reaching the goal, 0.45 for time against a target with a 6 percent buffer, 0.10 for obstacle clearance, clamped to 0 to 1; a crash, tilt, timeout or infeasible run scores 0.01.

code

King of the hill per family, not winner take all.

code

Nobody is named.

our reading		4 of 7read 2026-09-19
26RedTeamSN61 · security1.2% of emission

Code that solves a security challenge.

own docs

Per challenge, the validator runs the submission in a sandbox and gives a raw score from 0 to 1 against the previous best; below the challenge minimum, 0.85 for bot_virus_v1 and 0.5 for ada_detection_v3, it is not accepted.

own docs

Proportional within each challenge, then weighted across challenges: ada_detection_v3 carries 0.6 of the miner share and bot_virus_v1 0.4.

code

Nobody today.

our reading		4 of 71 disagreementread 2026-09-19
27CliqueAISN83 · other1.1% of emission

A clique: a set of vertices in a validator-supplied graph in which every pair is connected, returned inside a time limit of 6 to 30 seconds.

code

Validity first: an invalid or non-maximal clique scores zero.

code

Top-heavy, not proportional.

code

Unknown.

not found		6 of 71 disagreementread 2026-09-19
28HoneSN5 · data1.1% of emission

A program that solves a coding problem.

own docs

All or nothing, then speed.

code

Proportional.

code

Looked for a price, a customer, a product page or an endpoint on honedashboard.com, a miner-performance dashboard that renders Waiting for aggregate data to a fetcher, and in the README, DESIGN.md and DEMO_MINER.md at the pinned commit.

not found		5 of 7read 2026-09-19
29KubeTEESN90 · compute1.1% of emission

Confidential GPU capacity.

own docs

A dollar target, not a benchmark.

code

Proportional to dollar target, capped by the pool.

code

Not an enterprise yet.

own docs		4 of 71 disagreementread 2026-09-19
30BitMindSN34 · media1.0% of emission

A media-provenance classifier: a model that takes an image, video or audio clip and returns class scores for real against synthetic.

own docs

Per modality a detector is benchmarked on public, private holdout and generator-made samples and scored by sn34_score, which combines classification performance, the Matthews correlation coefficient, with probability calibration, the Brier score, under round-specific sample weights.

own docs

Fixed lanes: image 0.40, video 0.40, audio 0.04, generators 0.16.

code

Developers and enterprises.

own docs		3 of 7read 2026-09-19
31RidgesSN62 · agents1.0% of emission

A software engineering agent: one Python file exporting agent_main that takes a problem statement and returns a unified diff.

own docs

Deterministic: apply the patch, run the hidden tests, score the fraction that pass.

own docs

Proportional to decaying reward scores across every approved agent, summed per hotkey.

code

Developers with a GitHub repository, through Ridgeline: one credit turns one issue into one pull request from an isolated machine.

own docs		3 of 71 disagreementread 2026-09-19
32blockmachineSN19 · other0.9% of emission

Answered JSON-RPC requests.

own docs

Work times price.

code

Proportional to dollars consumed.

code

Apps that call blockchains, paying per request unit: as low as $1 per million requests at scale, with a calculator putting 100 million requests a month at $165 against Alchemy at $979, and failed requests not billed.

own docs		5 of 72 disagreementsread 2026-09-19

A program is the site's reading of the subnet's mechanism, not a review of it. The first sentence of each field is shown here; the full field, its quote, and any place the sources disagree are on the subnet's page. The index measures whether a first-time reader can find the puzzle. This records what the puzzle is.

Three programs were read by hand and the rest by reading agents against one brief, each checked by a script before publication and reviewed after. The method page has the field definitions, the trust scale, and the sources in order.