What runs on a contributor
The pilot accepts one compiled workload: pi_chunk_v1, an indexed integer Monte Carlo experiment that estimates π. A job is divided into chunks of at most 250,000 synthetic points. Contributors request assignments, return integer counts, and the coordinator checks each count before including it in a receipt.
Install and open the client
- Open the contributor section and select Check available downloads. Links appear only when the build manifest is published. Choose the archive for your operating system and compare its SHA-256 with the displayed checksum.
- Windows x64: extract the ZIP into a folder, then open the included executable. The executable is unsigned. A matching SHA-256 checks the archive against the published value; it is not a publisher signature.
- Linux x64: extract the tar.gz archive, open a terminal in the extracted folder, and run
./scorecompute-contributor. The published build requires glibc 2.39, with Ubuntu 24.04 as its baseline; compatibility with older Linux systems is not implied. - The client opens a local browser dashboard. Every launch begins paused. A fresh installation has CPU contribution disabled. Saved resource choices never start a calculation by themselves.
For CUDA contribution, a compatible NVIDIA driver must support the embedded PTX. You do not need to install the CUDA Toolkit or NVRTC to run the client. A GPU initialization or execution error remains visible and never silently enables CPU work.
The client does not install a startup service. If the browser does not open automatically, launch it with --no-open and use the loopback dashboard address printed by the client.
Choose resources before pressing Start
- Choose a public alias. Use a nickname rather than a hostname, email address or other personal information.
- Select the GPUs you want to offer. CPU is optional: explicitly enable it and choose 1–8 threads, capped by the machine’s available logical processors.
- Choose a duty percentage and review the selected resources.
- Press Start contributing to register and enable assignments. Until that action, the client performs no contribution work.
Duty inserts rest periods between small compute batches. It is not a wattage limit, guaranteed instantaneous utilization level or a way to interrupt a running GPU kernel.
Pause, change preferences or leave
Pause contribution prevents new work; a small batch already running may finish. Preferences unlock when that batch has ended. Save preferences keeps the client paused. Revoke registration pauses the client and revokes its coordinator credentials; a network failure is reported rather than silently treated as a successful revocation. Revoke before changing a registered public alias. Quit client closes the local process.
Settings and coordinator credentials remain in your user profile. The browser dashboard binds to loopback, checks its origin and protects every control request with a local secret. The coordinator token is not exposed in the dashboard or public status. Read the data and retention notice.
What the network publishes
Public status includes chosen aliases, opaque client identifiers, reported hardware, offered CPU threads, individual GPU memory, resource states and heartbeat age. Raw device and installation identifiers and credentials are not public. The coordinator receives registration and resource identifiers for assignment and deduplication; hardware declarations are not authenticated hardware identities.
Synthetic job parameters and receipts are public. This pilot accepts no private files or user datasets. The separate scientific tools keep their arguments, results and detailed composition events out of the public observatory.
Read capacity without overstating it
- Available and busy GPUs count freshly offered resources. Paused, stale or duplicate offers are not extra available devices.
- CPU threads are explicitly offered threads, not an inferred CPU benchmark.
- Reported GPU memory is memory across separate devices. It does not create one larger VRAM allocation.
- Observed throughput is samples accepted in the previous 60 seconds divided by 60. It is not theoretical FLOPS, future capacity or a net performance gain.
Snapshots carry their observation time and a 45-second freshness limit. Old observations do not establish current capacity. Hosted scientific worker inventory is not added again to the contributor totals. Multiple client instances are not proof of multiple physical machines. Initial GPU validation uses one physical RTX 4080 SUPER; a tested physical multi-GPU cluster is not claimed.
Run the public experiment
The website demo requests 1,000,000 samples with seed 42 and automatic backend choice. It follows the actual ticket and checked counts, then offers the final public JSON receipt. A queued or running job is not a finished result. Only a completed receipt contains a final estimate.
Clients can pause or disconnect. Assignment leases last 20 seconds; jobs expire after 120 seconds and can remain incomplete. Admission, verification concurrency and storage are bounded. Rate limits or insufficient capacity can reject a new request. No completion or uninterrupted capacity is guaranteed.
Watch verified events, not inferred GPU activity
The public network panel follows actual job creation, chunk assignment, verified counts and completion. A leased chunk means an assignment was issued; it does not prove that a GPU kernel has started. The animation represents received events, not sustained hardware utilization.
The separate contributor activity journal retains at most 256 events in memory and returns at most 128 events per poll. Its snapshot lists up to 16 active and eight terminal public synthetic job receipts. A fresh connection or reset displays recent history without replaying it as new work. Event history disappears when the coordinator restarts; retained job receipts have their separate storage limits.
Ask an AI assistant to use the pool
Connect through the public MCP endpoint. These three network tools are independent of the fourteen scientific engines and six optional composition tools:
{
"name": "network_status",
"arguments": {}
}
{
"name": "submit_network_job",
"arguments": {
"samples": 1000000,
"seed": 42,
"backend": "auto"
}
}
Pass the returned job_id to get_network_job to follow progress. The sample bound is 100–5,000,000, seed is an unsigned 32-bit integer, and backend is auto, cpu or cuda. Use the returned state, rather than treating admission as completion.
Keep the two Monte Carlo models separate
pi_chunk_v1 uses indexed integer points on a discrete 31-bit grid. The existing simulate_pi scientific engine uses its own floating-point model and confidence reporting. Matching seed numbers across those tools do not make their experiments identical. Contributor verification confirms submitted counts for this fixed model, not a contributor’s hardware identity or general reliability.
A per-GPU lock coordinates the client with ScoreCompute’s native engine on the same operating system and user account. Windows and WSL do not share that lock: do not offer two views of the same physical GPU as separate capacity.