Reference

Security and limits

What Mason protects, how it catches cheating workers, and what it does not do.

Anti-cheat

  • Hidden checks. Before its first paid job, and then after every 8 jobs and every 5 idle minutes, a worker receives a prompt with one checkable answer (arithmetic, exact echo, capitals, uppercase), phrased like ordinary traffic. Operators can load a private pool of checks so the templates are not public.
  • Throughput check. Each price class has a ceiling on believable speed. A check answered faster than that fails, which catches workers that return canned text instantly.
  • Billing. Tokens are counted by the orchestrator; a worker's own count is only used when it is lower. max_tokens is enforced on the stream.

A failed check forfeits the earnings since the last passed check. Repeated failures pause the worker for 30 minutes.

What the tests prove

tests/network.test.mjs (15 tests) runs the real server with scripted workers: exact debits and 70/30 splits, the OpenAI stream format, the free allowance, a cheater paused without pay, a speed faker caught, transparent failover, the no-charge rule on a mid-answer drop, max_tokens enforcement, and the Ollama node.

tests/robinhood.fork.mjs (30 checks) runs on a local fork of Robinhood Chain with the real Pons contracts: a coin launched with the treasury as fee recipient, ETH deposits at the on-chain price and their replay and data protections, exact ETH payouts, real Pons fees claimed into the treasury, staking signatures and snapshots, the buyback on the Pons curve straight to the burn address, a graduated coin bought in its Uniswap v4 pool and burned, staker payouts, and the operator withdrawal.

Privacy

Prompts are streamed to one worker and not stored by the network. A worker sees the content of the job it serves, not who sent it. Do not send secrets you would not show to a stranger's computer: that is, by design, where the job runs.

What Mason does not do

  • No cryptographic proof that a given model produced a given answer.
  • No content filtering on top of the open models.
  • No guaranteed latency: speed depends on which workers are online.
  • No closed models, only open-weight ones.