Why We Built Proof of Feed

The history of blockchain consensus is a history of tradeoffs. Satoshi chose Proof of Work because it was simple, auditable, and resistant to Sybil attacks. But he chose it in 2008, when the idea of running a useful computation inside a consensus protocol was not on the table.

We are building in 2025. The table is different now.

What is wrong with existing consensus

Proof of Work asks miners to find a hash below a target. The computation is deliberately useless. The waste is the point — it makes attack expensive. But it also means that every joule spent securing a PoW network is a joule that did nothing for anyone except increment a difficulty counter.

Proof of Stake replaced wasted electricity with locked capital. This is cleaner, but it shifts the power to those who already have tokens. Work is not required. The validator with 100,000 staked tokens earns 10,000 times more than the validator with 10 tokens, regardless of what either of them does.

Neither mechanism has any concept of useful output.

What AI inference is

A transformer model evaluating a prompt is a fixed, deterministic computation. Given the same input, the same model, and the same configuration, the output is predictable within a defined tolerance. This is important. It means inference work is verifiable.

When Blazil runs a data chunk through Tract ONNX, it produces an output and commits a TransferId to TigerBeetle. That TransferId is a cryptographic anchor. Any validator can query TigerBeetle and confirm that the computation happened on a specific machine, at a specific time, for a specific input.

This is not possible with Bitcoin mining. There is no way to verify that a miner actually ran a computation that was useful to anyone. The hash either meets the target or it does not.

How Proof of Feed works

Proof of Feed is simple in structure. A miner receives a data chunk from the network, feeds it through a co-located Blazil engine, and gets back a TransferId from TigerBeetle. That TransferId, combined with the chunk hash and a timestamp, forms a FeedReceipt. The miner signs the receipt and includes it in a block candidate. Validators confirm that the TransferId exists in TigerBeetle. If it does, the proof is valid.

The key insight is that TigerBeetle is not owned by any single party. It is a distributed, Byzantine fault-tolerant ledger. The TransferId commitment cannot be faked without compromising the TigerBeetle cluster, which requires controlling a Byzantine majority of its nodes.

What this means in practice

Every block mined on Feedable Network represents real inference work. The miner earned their reward by serving intelligence to the world, not by burning electricity on a puzzle or locking capital in a contract.

The economic alignment is different from any prior consensus mechanism. Miners are rewarded for throughput. The more data they feed through Blazil, the higher their probability of producing a valid block. This creates a direct incentive to run Blazil at full capacity, which is exactly what the network needs.

The challenge protocol ensures that 5% of inference jobs are randomly re-verified by independent miners. Since the selection is seeded by a future block hash, miners cannot predict which jobs will be checked. This makes gaming the system statistically unfeasible.

What comes next

Proof of Feed is live in devnet form. The genesis block contains the message that started this project:

On this block, intelligence became a public utility. No permission required.

We are working toward testnet launch. The full specification is open source at github.com/feedable-network-lab/feedable-blockchain-coin.