Real requests hide inside a fixed volume of traffic. While real users are few, miners fill the rest with decoy requests. As real usage grows, the noise shrinks.
v0.1 · local only. the lab runs in your browser; the network comes next. privacy here has a price: decoy requests cost real inference.
enter labILLUSTRATION
one real request among 999 decoys. an observer sees 1,000 requests of the same size and rhythm.
what happens to a single private request on its way to the model, and what each layer is meant to hide. every row carries its status: v0.1 local (runs in the lab today) or planned.
| payment | zk usage credit | you pay for requests without linking them to each other or to you | PLANNED |
| size | fixed size classes | every request is padded to one of a few sizes, so length does not leak | v0.1 LOCAL |
| route | mixnet, several hops | network decoys travel with you and are dropped at the last hop | PLANNED |
| gateway | partner gateway, staked | runs open-source models; real and decoy requests arrive the same way | PLANNED |
| endpoint | decoy prompts | decoy prompts are generated so the provider has to guess which ones are real | v0.1 LOCAL |
| observer sees | ~1,000 req/min, flat | a steady volume whether 50 or 900 people are using it | SIMULATION |
noise has to exist at two places. the network layer hides when and from where you send. the endpoint layer hides which prompts are real. only the second one costs real inference.
Fake packets from the client into the mixnet, thrown away at the last node. They protect against a provider that works with the entry node and plays with timing to match requests to people.
Decoy requests that really reach the model, so even the SI provider can not tell which ones came from a person. This is the expensive layer, and it is what the Noise Pool mainly pays for.
miners commit to a batch before sending it, and a random sample is opened afterwards. one epoch is one hour. commit, sample and open already run in the lab; stake and payouts come with the network.
The miner locks a stake before taking part.
At the start of each epoch, the miner publishes a hash of the decoy batch it is about to send.
At the end, the hash of a future block picks about 1% of the batch at random.
The miner reveals the sampled decoys. Checked: they were sent (their nullifier shows up at the gateway) and they look credible.
Pass: paid for the whole batch. Fail: the stake is lost.
PLANNED · v2the audit becomes a zk proof: the miner shows that k of the published nullifiers are theirs, without saying which.
a drawing of the intended design, not a live map. there are no running nodes yet, so this shows roles, not counts.
decoys that reach a model cost real inference. this is the honest math at a fixed target of 1,000 requests per minute, using orientative prices from the design.
assumes ~$0.001-0.005 per decoy on a cheap model and ~10x on a top model. at 2 decoys per real request, each real request costs about 3x. the Noise Pool is meant to be funded by a ~20% fee on real requests plus a bootstrap fund; at low usage, the bootstrap fund carries most of the cost.
none of these are solved. each one has a planned answer, and each answer still has to be tested.
A miner who also runs a gateway could "send" decoys to itself and get paid for nothing.
A provider can train a model to spot decoys. This becomes an arms race, not a one-time fix.
Automated fake traffic to closed SI providers may break their terms of use.
Endpoint decoys cost real inference. If funding dries up, the noise level drops and so does privacy.
times are estimates for building each step, not dates. nothing starts counting until the previous step works.
Decoy generator, fixed size classes, random timing and a commit you can audit, running in your browser. Try it in the lab. Next: the same logic as an open-source client that sends real requests.
in progressLIVE · LOCALRun it for 24 hours and publish the provider's view: a flat line where real requests can not be picked out.
1 weekPLANNEDContract for stake, commit, random audit and rewards. No zk yet.
~2 monthsPLANNEDAudits become zero-knowledge proofs, built with a cryptographer on the team.
3+ monthsPLANNED
[nz.11] / end of signal
> awaiting next build