Proof of Thought: consensus that records a mind

Beyond proof of work and proof of stake

Proof of work secures a chain by burning energy on arbitrary hashing; proof of stake secures value by bonding capital but produces nothing external. Both spend a security budget that creates no artifact. QBC asks a different question: can the work that secures the chain also be useful? Its answer is Proof-of-SUSY-Alignment plus a per-block Proof-of-Thought, so the security budget produces quantum computation and an auditable record of the AI's reasoning. This sits inside the broader verifiable-AI design, and the lineage and tradeoffs of the idea are surveyed under proof of useful work.

Proof of SUSY Alignment: quantum useful work

Block production on QBC runs an 8-qubit variational quantum eigensolver (VQE): miners search for a quantum ground state whose energy falls below a difficulty threshold, on a golden-ratio emission curve. VQE is a genuine, physics-grounded workload from quantum chemistry and optimization (VQE review, arXiv:2111.05176), not brute-force hashing. Finality comes from staked validators and the GRANDPA gadget (arXiv:2007.01560), so the chain gets fast, provable finality on top of useful proof-of-work-style mining. A circuit-level walkthrough lives under VQE mining explained, and the consensus design is formalized in the VQE useful-work consensus paper.

What a Proof-of-Thought commitment contains

On top of block production, every block commits a Proof-of-Thought: the served model's identity, a digest of the reasoning context it processed, the RNG seed that makes the step reproducible, and the resulting thought hash. Together these let anyone verify that a specific model produced a specific reasoning step at a specific block, and detect tampering after the fact. It is provenance for cognition, and it is the substrate for auditing an AI's behavior over time rather than trusting a provider; together with the finalised training record it forms the consensus half of the AI audit trail. The retrieval and reasoning stack behind it is covered under on-chain AI inference.

How training epochs are finalised by the same consensus

Proof of Thought is not the only AI object consensus finalises. QBC's decentralised training epochs are attested against a held-out loss and finalised by a two-thirds validator quorum, using the same guarantees that finalise a payment. A regressing or poisoning epoch is rolled back and never finalised. That means the AI's learning and its reasoning are both first-class, consensus-finalised objects. The mechanism contributors earn from is explained on the decentralised AI training page, and the validator-side methodology, what a quorum can and cannot catch, is treated under validating ML training.

Why this is hard to fake and hard to copy

A consensus that records cognition needs three things most designs lack: a served model that exposes enough internal state to commit, a chain whose finality is strong enough that a finalised thought is durable, and an incentive layer that pays for verified learning rather than raw compute. QBC integrates all three, on post-quantum rails (NIST FIPS 204). The integration, not any single piece, is the moat. See the category overview.

Finality you can build on

GRANDPA provides deterministic finality (unpacked under GRANDPA finality explained) and BEEFY plus MMR add a signed finality-proof layer that light clients and bridges can verify without trusting a full node, the pattern behind light-client finality proofs. Both are live on the network today: the rollout is documented in the engineering log of BEEFY and MMR finality proofs. The AI-native pieces ride on the same guarantees, so a finalised Proof-of-Thought is as hard to reverse as a finalised transfer. This is what makes QBC's cognition record something you can build applications and audits on, not just observe. The consensus scale-out roadmap is in the QBC whitepaper.

A block's life, from quantum circuit to finality

Follow one block through the system. The leader for the slot runs the 8-qubit VQE search, iterating ansatz parameters until it finds a state whose measured energy clears the difficulty threshold; that proof seals the block alongside the Proof-of-Thought commitment and the pending transactions. Peers import the block, re-check the quantum proof and every Dilithium signature, and the GRANDPA voters then cast prevotes and precommits; once two thirds agree, the block is final, deterministically rather than probabilistically (the distinction is unpacked under deterministic finality explained). On the live network this loop runs across four validators on three physical machines, and the leadership handoff under a single-leader VQE regime is documented in the engineering log of VQE-secured single-leader block production. All of it is implemented as a custom engine on the Substrate framework, a build process described under Substrate custom consensus. The honest caveat: four validators is a small set, and scaling the same pipeline to thousands of block authors is roadmap work, stated as such.

War stories: what running a real finality gadget teaches

Consensus designs meet reality in incidents, and QBC publishes its incident engineering rather than burying it. One finality stall traced to the voter's chain-selection logic attempting to prevote on a block state the node had already pruned; the round halted until a voter restarted, and the root cause and fix are documented in the incident writeup of the GRANDPA SelectChain fix. Another class of risk, a validator signing two competing blocks, is handled autonomously: equivocation proofs are detected and slashed on-chain without an operator in the loop, an implementation covered in the log of autonomous equivocation slashing. These details matter for two audiences: operators deciding whether running a validator is operationally sane, and evaluators of AI mining claims, who should demand that any network publish its failure modes as readily as its benchmarks. A consensus you can audit includes auditing its worst days.