Points
Per row that passes audit. Per dollar of bounty.
Mining that leavesa dataset behind
See how it works
01What the cards do
- Pearl pays hundreds of thousands of graphics cards for AI-style arithmetic.
- Most of that arithmetic produces nothing anyone uses.
- Researchers now pay more to rent the same cards.
- A 24 GB card is too small for chat-speed inference.
- It suits work that can wait: training data, distillation sets, evaluation runs.
- Your GPU already does the arithmetic. Strand has it write a dataset anyone can check.
Point the work
The network already pays for most of the computing. Strand points it at datasets anyone can download and check.
02How it works
Every row can be re-run
A job spec goes in. Rows come out, each with the seed and model hash that made it.
-
01Post a job
A sponsor posts the prompts, an open model, fixed decoding settings, a row schema and a budget. The job spec pins the model by its hash.
-
02Mine the rows
Member GPUs generate the rows as their mining work through Pearl's official vLLM plugin. Miners receive PRL block rewards directly, as in any pool.
-
03Publish every row
Each row goes to Hugging Face under an open licence with its seed and model hash. The dataset ledger shows rows, GPU-hours, megawatt-hours and audit pass rate for each dataset.
-
04Re-run a sample
An auditor re-runs a random 1–2% of rows to catch fakes. Points are paid only for rows that pass.
03Roles
Who brings what
Sponsors bring prompts, an open model and a per-row bonus in USDC. Miners bring their cards and keep their PRL block rewards, plus points for each row that passes audit. A verifier needs only a published row, its seed and its model hash.
04The first dataset
Dataset 01
The pool pays for the first dataset itself: open reasoning traces on a permissively licensed model. It exists so there is something to download before the first sponsor signs.
Useful to whom
To whoever downloads the dataset. Open-source labs, academics and fine-tuning groups pay a small per-row bonus instead of a rental bill. The bonus is sized to cover the hashrate miners give up by doing real work.
Pick any row
Every row carries its seed and model hash. Download one, run it on your own card and compare.
- Dataset
- 01
- Row
- <n>
- Seed
- 0x…
- Model hash
- sha256:…
- Decode
- <fixed settings>
- Licence
- Open
Row passed re-run. Points added for this dataset.
Re-run did not match the published row. It is withheld and earns no points.
05Token
$STRND
Points come first
Through the first two or three datasets, miners earn points per row that passes audit and sponsors earn points per dollar of bounty. Then comes a fair launch on a Solana bonding curve with no presale. The token is a membership and fee claim, not the payment rail.
Fair launch
A bonding curve on Solana, after the points season. No presale.
Buybacks
The pool fee funds buybacks. The team's 10% is vested.
Token spec
Values from the design. — where none is set| Ticker | $STRND |
|---|---|
| Type | Points for miners and sponsors, converting to a pool-fee buyback token |
| Chain | Solana |
| Launch | Points season through the first 2–3 datasets, then a fair launch on a bonding curve |
| Presale | None |
| Pool fee | 5% of sponsor bonuses plus 1% of PRL mined, used for buybacks |
| Team | 10%, vested |
| Holding it | Membership and a claim on pool fees. Not the payment rail |
Roles side by side
Sponsor · Miner · Verifier| Item | Sponsor | Miner | Verifier |
|---|---|---|---|
| Brings | Prompts, an open model, decode settings, a row schema, a budget | A GPU running Pearl's vLLM mining plugin | A published row, its seed and its model hash |
| Pays in | USDC, per-row bonus | — | — |
| Gets | An open dataset at a fraction of rental price | PRL block rewards directly, as in any pool | A re-run to compare |
| Points | Per dollar of bounty | Per row that passes audit | — |
| Checked by | Model pinned by hash in the job spec | Random 1–2% re-run by the auditor | Their own card |
06Why now
The card window
PIP-3 is a draft Pearl fork to FP8 with no activation height yet. It would take RTX 30 and 40 cards out of mining. Strand puts those cards to work while they still mine and plans the move to RTX 50 and Hopper.
Nothing else here waits on a date. The standing criticism of Pearl is that its mining computes nothing useful. One study counts about 320,000 RTX 3090-class GPU-equivalents drawing 112 MW on no useful AI work, and budget GPU rentals up 38% as a side effect. A dataset anyone can download is the plainest reply.
What is unresolved
Read before you point a card- 01
Sponsors are few and short of cash
Open-source labs may prefer free volunteer compute, as Prime Intellect used for SYNTHETIC-1, or their own GPUs.
- 02
PIP-3 could remove most members' cards
If the FP8 fork activates, RTX 30 and 40 cards stop mining Pearl. Those are most of the pool's natural members.
- 03
The bonus has to track PRL
Real work costs about 10–30% of hashrate. As PRL rises, so does what a miner gives up, and so does the bonus a sponsor must pay.
- 04
Re-runs are statistical, not exact
LLM decoding is not bit-reproducible across batch sizes and hardware. Strand uses batch-invariant kernels where it can and log-probability checks where it cannot. Subtle fakes can still slip through.
- 05
Mid-size models only, for now
24 GB cards limit jobs to models of roughly 8–14B at INT8 unless the pool shards larger ones. Some sponsors will want frontier-model traces.
- 06
Weak value capture
Pool-fee buybacks are small. Coins that reward useful public compute have not held value in the past.
- 07
Bigger players could build it
Pearl or Prime Intellect could ship the same thing with more GPUs and a better-known name.
Your GPU already does the arithmetic
Strand has it write a dataset anyone can check.
Join the pool