FAQ
Short, honest answers about how the constructor works: timings, privacy, keys, Spaces and the model catalog. Numbers come from network settings.
A neuronet and its lifecycle
How do I create a neuronet and what does it cost?
The Create wizard walks you through economy (stake and owner fee), name and review. The registration price is read live from the chain — it grows with every neuronet created and settles in about a day, so the review always shows the current figure. On managed hosting the platform treasury pays the registration and the initial stake; you work on a subscription. Tasks are added after creation, on the Tasks tab.
#Will a neuronet or a task ever delete itself?
No. Idle time never deletes a neuronet or its tasks — it only changes the status and stops emission. Deletion happens on your command alone: dissolving the neuronet or removing a task. Tasks have no expiry at all.
#What do the statuses Active, Warning and Frozen mean?
Active — useful work is flowing. Warning — idle for over 50 400 blocks (≈7 days): everything still runs and rewards accrue. Frozen — idle for over 216 000 blocks (≈30 days): the neuronet earns nothing, but its requests are still served. The freeze lifts by itself with the first served request: the gateway reports that work to the chain, and the status turns Active again at the neuronet’s next settlement boundary. Work done while the status still reads Frozen is not paid.
#What counts as activity?
Activity means executed requests to your tasks, and nothing else. The idle clock resets when the gateway settles those requests with the chain. Adding a task, staking or editing settings do not count: with no requests, the clock keeps running.
#What happens when I dissolve a neuronet?
Dissolving starts a grace of 7 200 blocks (≈1 days): the neuronet takes no new miners, stake, tasks or settlement batches, everything in it keeps running and earning, and you can cancel it in the Danger zone. When the grace ends the teardown starts and cannot be undone: the stake is queued back to its stakers - available 50 400 blocks (≈7 days) later -, the miners are removed, the remaining earned emission is paid out and the record is deleted. The registration cost is not refunded.
#What is CU per request?
The declared compute cost of one request — a reserve for future useful-work weighting. Today rewards inside a task are split by credited requests and miner reputation; CU is a common multiplier for all miners of the task and does not change the shares.
#What does task weight do?
Weight splits the miner share of the neuronet’s emission between tasks — what is left after the owner fee and the stakers’ share: a task’s share equals its weight divided by the sum of all weights. The split bar on the Tasks tab shows the result.
#Privacy and API keys
Which privacy levels actually work?
Today P0 and P1 are active. P0 is the open mode. P1: jobs, responses and verification materials are encrypted between the gateway and miners. The gateway, selected executor and assigned verifiers see the data needed to process the request: P1 protects data in transit, but does not hide it from these participants. P2–P4 are locked in the selector and will open once the pipeline genuinely enforces them — promising earlier would be dishonest.
#Can I restrict an API key by IP?
Yes. A key carries an allowlist — up to 32 entries, IPs or CIDR subnets; an empty list means no restriction. The list is edited without changing the secret, and key rotation inherits it. The gateway enforces the check and answers 403 ip_not_allowed for a foreign address; the client address is used only at the moment of the check and is never stored.
#Can a key be issued for a single task?
Yes: pick a task when issuing — the key will then work with that task only (another task answers 401). A key without a task works with every task of the neuronet. Each key also has a daily request budget; when it runs out the gateway answers 429 quota_exceeded and the budget returns at midnight UTC. Revoking a key reaches the gateway within a minute.
#Spaces
What is a Space?
A ready-made frontend on top of a neuronet task: chat for LLMs, image or video generation. It opens by link; a visitor needs neither a wallet nor a key — requests go through the platform server with the Space’s own key and its own daily limit.
#How do public, link-only and private differ?
Public — listed in the Spaces catalog and open to search engines. Link-only — opens for anyone with the direct link, but is never listed and is closed to indexing (the link itself is unguessable). Private — owner preview only: the page is not served to visitors.
#Can a visitor see the system prompt?
A Space visitor — no: the prompt is stored on the server and mixed into the request there, it never reaches the browser. But it is not a defence against an arbitrary API client: one that sends its own system message or a ready-made envelope replaces the owner’s prompt. For chat it is the assistant’s persona; for images and video, a permanent style prefix.
#How many Spaces can I create?
The per-neuronet limit is set by the platform: 3 by default, and a network or a plan can set its own. At the limit, creation returns a space-limit error; deleting an unused Space makes room for a new one.
#How do I check that my neuronet answers?
Two ways. Quick: create a Space for the task and send a request from its page — the same production route your visitors use. Programmatic: issue an API key on the API keys tab and call the endpoint from the Endpoint tab with that key.
#Model catalog
What does the Featured badge mean?
A storefront badge: platform administration assigns it, and together with the weight it defines the order in the Featured sort. It does not affect execution, parameters or verification of the model.
#What does GGUF mean in the quantization filter?
A weights packaging format: the same artifact runs on Metal, CUDA and CPU. In the filter it is an additional value on top of precision, not a replacement — a Q4_K_M entry is found under both Int4 and GGUF.
#How do I pick a model for my hardware?
The catalog filters by tier, weights access, hardware class (consumer or datacenter, by the GPU architectures a model runs on), the VRAM it needs, license and language. The model picker of the task wizard filters by hardware class, model family, parameter size and quantization. Each model card shows the minimum VRAM.
#