Consensus Anchored by Cryptographically Verifiable AI Work
Proof of Intelligence is the core of AETRON. Miners run inference and training, cross-check each other through logit replay, and earn emissions for work that the protocol was able to cryptographically reproduce.
How Proof of Intelligence diverges from legacy models
Proof of Work
Burns energy for hashes
Security rests on empty computation. Every extra kilowatt only proves the network is still alive.
Proof of Stake
Buys safety with capital
Token holders take the emission. Those who actually do the work get no reward. Capital moves between wallets, no new value is created.
Proof of Intelligence
Rewards verifiable intellect
Miners are paid for real AI outputs that any other participant can cryptographically re-check.
What PoI is made of
Neuronets
The unit of AI work organisation. Each neuronet hosts multiple tasks (inference, training, private models) with a fixed model hash, canonical execution spec, and access policy.
No cap on neuronets or miners. The neuronet owner retains veto over the miner roster.
Canonical Execution Spec
A task pins the data type, quantization format (more than 15 are supported: BF16/FP16/FP32, NF4/FP4/INT8, FP8, MXFP4, NVFP4), GPU architecture, attention backend, and torch/CUDA versions. Reference Probes compare the miner's logit hash against the ground truth: either it matches bit-for-bit, or the task is rejected.
Validated across 6 architectures: NVIDIA Ampere/Ada/Hopper/Blackwell, AMD CDNA3, and Apple Silicon.
Replay Logits + Shadow Replay
A VRF randomly picks 7% of requests for replay. The assigned miner runs a single forward pass (prefill only) and compares logits; the cost of this check is around 0.2% for LLMs and 5–10% for diffusion. Shadow Replay closes selective-answer fraud, where a miner cheats only on a subset of requests.
If the hashes don't match, a third miner settles the dispute. Whoever loses the arbitration earns no emission for the epoch.
Training-time defense
In collaborative training mode, every step is watched by four independent detectors at once: activation anomalies, loss-trajectory anomalies, gradient correlation, and cryptographic weight binding. Any single signal is enough to stop the step. As of today, this is the only working defense against the "Backdoor in the Middle" class of attacks on distributed training.
The defense has been tested against more than 20 attack variants, including stealth attacks and adaptive bypass. Every attack was caught, with no false alarms on honest training runs.
Proof of Intelligence in action
Define the task
The owner creates a neuronet and publishes the task parameters: model hash, data type, quantization format, GPU architecture, and reference probes. Miners register by paying a small fee or solving a PoW puzzle; no locked collateral is required.
Execute and commit the artifact
The miner serves the request and stores a compact validation artifact (around 200 bytes: top-k logits or a latent hash, seed, context hash). Those artifacts are collected into a Merkle tree, so a single on-chain transaction can cover up to a million requests.
Replay and arbitration
A VRF picks 7% of requests at random and assigns a second miner to validate them. That miner runs a prefill and compares the logit hash. If the hashes disagree, a third miner arbitrates; the side that loses the arbitration earns no emission for the epoch.
Distribute emissions
The runtime scores each miner on four signals: confirmed work output, the result of the Pulse benchmark run (a canonical reference workload for each GPU architecture), uptime, and how many of the assigned re-checks they completed. To earn emission, serving your own requests is not enough - you also have to close out every re-check the network sent your way.
A new miner enters with a reduced reputation weight and grows it to full by passing checks cleanly. It's an open trajectory of trust - no grace period. Governance sets the specific thresholds.
Who benefits from PoI
AI Startups
Run models without owning the infrastructure. Verifiable inference lets an enterprise customer confirm the model is doing exactly what was promised, without having to trust the provider.
Research Teams
Run distributed training with the Witnessed Checkpoints protocol. A bit-exact replay of a single step catches weight substitution and skipped steps - in our tests it caught 15 attacks out of 15. For convolutional networks, identical output is confirmed even across different GPUs. Publish your training runs and earn emissions for confirmed progress.
Infrastructure Providers
Rent out GPUs for workloads where verifiability matters: healthcare, finance, defence. The optional Confidential Computing mode with TEE attestation lets you serve private models without trusting the operator. In feature scope, AETRON covers all of Cocoon and still delivers verifiable inference even without TEE.
Economics & incentives
Emission split
- Every block mints 1.5 AET: 90% goes to neuronets in proportion to the work they confirmed, 5% to the chain's validators, and 5% to the treasury. The block reward halves every four years, the same schedule as Bitcoin. The hard cap is 33 million AET, with no premine. An empty neuronet earns nothing.
- Inside a neuronet the emission is split this way: 80% to miners, 10% to the owner and 10% to stakers. The shares are fixed by the protocol and are not configured when the neuronet is created.
- Emissions are paid only for honestly confirmed work. A miner who skips assigned re-checks, or loses arbitration, gets no reward for that epoch.
The role of stake in a neuronet
- Stake inside a neuronet plays two roles. It sets the network's throughput - how many requests per epoch and how many miners can connect - and it also acts as a vote on whether the service is useful. A neuronet with no stake simply cannot serve requests at all.
- Emissions between neuronets are split purely by the confirmed work each one delivered. If a neuronet has a lot of AET staked into it but no real traffic, its income is zero as well.
- When a miner stakes the AET they earned back into their own neuronet, throughput grows, requests grow with it, and the earnings start to compound. That's why every participant is better off holding and staking AET rather than selling it immediately.
Reward for a verifiable result
In AETRON, value comes from work that can be re-checked cryptographically. Every block brings inference, training and services that are trusted at the protocol level. This is the base layer of the next generation of AI infrastructure.