Proof of Intelligence: The DAIT Consensus Class canonical
Proof of Intelligence (PoI) is a new consensus class. The idea has been in development for over a decade, across two prior iterations (aiscendence.com in 2018 and dait.io in 2020) before the chain implementation now underway. The formal definition lives in Livepaper section 3, revision 36; this page is the canonical explainer. It argues why none of the inherited labels describe what the DAIT chain does, gives the formal definition, walks through the four implementation axes, maps each of DAIT's nine custom modules to the relevant axis, and addresses the five stakeholder lenses (operators, validators, token holders, agents, regulators) that benefit from clear nomenclature.
License. The PoI definition is published under CC-BY 4.0. The term has been in use by the project's founder since the early Bitcoin-mining and data-center years; the formal definition published here is unowned and unencumbered. Other protocols are welcome to claim the class if they satisfy the conditions.
Reference. See Livepaper section 3, revision 36 for the formal protocol description, standards.html for the per-condition control matrix, and the chain repository for the live code.
1. Why a new term?
The crypto industry has accumulated a vocabulary of consensus class names: Proof of Work, Proof of Stake, Proof of Useful Work, Proof of Spacetime, Proof of Coverage, Proof of Authority, Proof of History, Delegated Proof of Stake, and a long tail of trademark-driven variants. Each named class describes a real engineering shape. Each came into use when an existing label became misleading. We formalize Proof of Intelligence here for the same reason: the labels we inherit are now misleading when applied to DAIT, and using them would obscure what the chain actually does.
Start with Proof of Work. PoW is the founding consensus class, and it works. Bitcoin is fifteen years old, has never lost a block, and has never been double-spent at the protocol layer. The cost is energy. The cost is also semantic: the work the network pays for is a SHA-256 collision, useful only to the verifier. PoW is honest about what it is. It does not claim the work is anything other than puzzle work. The honesty is a feature. But it means PoW cannot be the consensus class for a chain whose explicit purpose is to monetize cognitive work for tenants. The energy spend is wasted from the tenant's perspective.
Proof of Useful Work was the first attempt at a fix. The earliest serious PoUW proposals (Primecoin in 2013, several academic papers through the 2010s, Coin.AI and similar AI-flavored variants from 2017 onwards) replaced the SHA puzzle with a verifiable scientific or AI computation. The label is accurate as far as it goes. But it has acquired a narrow technical meaning. PoUW chains in the wild typically pick one workload class, hardcode one verifier function, and make the consensus path identical to PoW with a different hash inside. The chain pays for one kind of work, verified one way, with constant economic parameters. DAIT does none of these things. DAIT pays for many kinds of work, verified many ways, with parameters that move with use. Calling DAIT a PoUW chain would tell a sophisticated reader that DAIT is in the same architectural category as the prior projects, which it is not. The vocabulary forces a wrong group identification.
Proof of Stake is the second relevant class. PoS is consensus by economic security. Validators bond capital, sign blocks, get slashed for bad behavior. PoS is composable with anything: you can run PoS on a chain that pays for SHA collisions, on a chain that pays for AI inference, on a chain that pays for nothing at all. PoS describes the consensus engine, not the work. DAIT is a PoS chain in the consensus-engine sense, but the unit-of-work is not. So PoS is necessary vocabulary but not sufficient. Saying DAIT is PoS leaves the cognitive-work claim entirely unstated.
The verified compute marketplace label fits more closely. Akash, io.net, Render, and Aethir are useful comparison points. Each runs an auction, each accepts compute jobs, each has some kind of attestation. But the verification is a marketplace property, not a consensus property. The chain underneath is general-purpose. The economic parameters of the base layer (block time, emission, validator cap) are constants. If the marketplace contract were removed, the chain would still run. DAIT is structured the other way. The verification primitives are chain-native modules. The economic parameters are functions of the verified work. Removing the marketplace from DAIT would remove the chain's reason to exist.
So the term we want has to say four things at once. The work is cognitive. The verification is plural and adaptive. The economic surface moves with use. The architecture is whole-stack rather than overlay. PoUW says the first thing only. PoS says the consensus-engine slot is filled. Verified compute marketplace says the second thing partially. None say the third or the fourth. The cleanest way to get all four into one label is to name a distinct class.
2. Formal definition
The definition has three conjunctive conditions. Each is independently testable. A protocol that satisfies all three is in the Proof of Intelligence class. A protocol that satisfies a strict subset has a different name, given below.
Proof of Intelligence (PoI) is the property of a consensus protocol that satisfies all three of the following conditions simultaneously:
(a) Cognitive unit-of-work. The protocol's accepted unit-of-work is a verifiable cognitive computation: an act of inference, finetuning, simulation, reasoning, or scientific compute, whose execution can be checked by hardware-rooted, cryptographic, or statistical means. SHA collisions and storage proofs do not satisfy this condition; an LLM token-generation receipt or an attested tensor-network simulation does.
(b) Adaptive verification. The verification stack is not a single fixed function. It composes hardware-rooted attestation (TEE), cryptographic proofs (signatures, vote-extension force-include, ZK selective disclosure), statistical sampling (PoR challenges, redundant re-execution, tensor-train spot decomposition), and behavioral checks (auditor-signed attribute claims, identity revocation accumulators) such that the verifier function appropriate to a given workload class is itself selected adaptively.
(c) State-indexed economic parameters. At least one of the protocol's primary economic parameters (emission rate, validator active cap, block time, credit weighting, validator self-bond floor) is a continuous function of network state rather than a constant set at genesis or by ad-hoc governance vote. The protocol's behavior changes with use, on a schedule the protocol itself maintains.
A protocol satisfying (a) but not (b) and (c) is PoUW. A protocol satisfying (a) and (b) but not (c) is a verified compute marketplace. A protocol satisfying all three is Proof of Intelligence.
Why each condition is needed
Condition (a) is the cognition condition. Without it, the chain pays for puzzle work or storage retention, and the term "intelligence" does not apply. The condition draws a sharp line. Storage proofs are out, even though storage is necessary. Subjective creative output is out, because the verification path collapses to reputational scoring. SHA collisions are out, obviously. Reasoning, inference, finetuning, simulation, and verifiable scientific compute are in. The set is not closed; new verifiable cognitive workloads can be admitted by governance, but each must come with an explicit verifier function before it can be paid for.
Condition (b) is the verification-plurality condition. Without it, the chain handles one workload class only, and the architecture cannot accommodate the heterogeneity of cognitive work. A chain that only verifies LLM inference cannot verify tensor-network simulation. A chain that only verifies via TEE cannot accept zkML proofs from tenants who reject hardware trust. Adaptive verification means the chain dispatches to the right verifier per workload, with the dispatch table itself governance-extensible. This is not vague extensibility. The verifier registry is on-chain, the dispatch is deterministic, and the audit log of which verifier ran is part of the receipt.
Condition (c) is the state-indexed-parameters condition. Without it, the chain's economic surface is set at genesis and cannot respond to its own use. A chain whose validator cap, emission rate, and block time are constants is, in the long run, governed entirely by ad-hoc governance votes whenever those constants need to change. PoI says: bake the response into the protocol. Let the validator cap grow with hosts. Let emissions track demand. Let credit weights adjust to workload mix. Let block time tighten as latency improves. The schedule is part of the protocol; the protocol does not need a vote to honor it. Governance still has authority over the schedule itself, but the day-to-day adjustment is automated.
Adjacent classes
The conjunctive structure of the definition implies several adjacent classes, each of which is a real architectural shape worth naming:
- (a) only: Proof of Useful Work. One workload type, one verifier, constant parameters. Most prior PoUW projects fall here.
- (a) and (b): Verified compute marketplace. Multiple workload types, plural verification, but the base chain is generic and the economic parameters are constants. Akash and io.net approximate this class.
- (a) and (c): Adaptive cognitive chain (no formal name in the wild). One workload type but with state-indexed parameters. Theoretical; we know of no production implementation.
- (b) and (c) without (a): Adaptive infrastructure chain. Useful for non-cognitive work that still needs plural verification and elastic parameters; some restaking and shared-security designs aim here.
- (a), (b), and (c): Proof of Intelligence. The class DAIT occupies.
3. The four axes
The definition is three conditions because three is the minimum required for sharp class boundaries. But when you actually build a PoI chain, the implementation surface decomposes into four axes, not three. The fourth axis is implied by the others but worth naming separately because it has its own design space.
Axis 1: the work itself is intelligent
This axis instantiates condition (a). The set of admitted unit-of-work types includes, in DAIT's v1 deployment: LLM token generation; finetuning passes (LoRA, QLoRA, full SFT); agent reasoning traces; tensor-network simulation; zkML proof generation as a paid workload. The set explicitly excludes: SHA collisions, idle benchmark computation, subjective creative output without an objective verifier, storage retention as a standalone unit. The exclusion list is not theoretical; each excluded category is what some adjacent project pays for. The inclusion list is fixed-by-genesis but extensible by governance, with the constraint that any new admitted workload must come with a registered verifier (axis 2) and a credit weighting (axis 3).
Axis 2: the verification is intelligent
This axis instantiates condition (b). DAIT composes four verifier families. Hardware-rooted attestation: multi-vendor TEE, two CPU enclave vendors plus one GPU confidential-compute vendor in v1, with the trust-root certificate set pinned at genesis and rotated by governance. Cryptographic proofs: ed25519 / secp256k1 today, ML-DSA-65 (CRYSTALS-Dilithium Mode 3) for agent identity from genesis, vote-extension force-include for attestation commits, ZK selective disclosure for credentialed identity, and an opt-in zkML premium tier. Statistical sampling: proof-of-replication challenges for storage, redundant re-execution for high-stakes workloads, tensor-train spot decomposition for simulations. Behavioral checks: auditor-signed attribute claims for host registration, listening-coefficient damping in subnet-style governance pathways, identity revocation accumulators for credential withdrawal. Adaptive dispatch is the chain-level property: the verifier function appropriate to the workload class is itself a value that the chain selects.
Axis 3: the network adapts intelligently to use
This axis instantiates condition (c). DAIT's adaptive parameters at v1: Active validator cap: floor(sqrt(avg-self-bond) + sqrt(unique-30-day-pouw-hosts) + sqrt(avg-blocks-per-second)), recomputed every epoch (epoch length approximately 3 days). Per-block emission: compute-indexed; the per-block mint scales with the chain's running unit-of-work throughput in the prior epoch, within a governance-set band. Credit weighting: per-workload-type, governance-tunable, defaults are inference 1.0, finetune 1.0, tensor simulation 0.7, scientific compute 0.5. Standby pool rotation: nine-epoch cadence keeps the active set responsive without inviting churn attacks. Block time: 1000ms at genesis with a scheduled drop to 500ms at the Q4 2026 upgrade if validator latency distributions support it. Validator self-bond floor: static-by-genesis, with a planned dynamic-floor variant whose parameters are still in design. The point is not that any one knob is unprecedented. The point is that the protocol's behavior is allowed to be a function of its own use, on a schedule the protocol itself maintains.
Axis 4: the actors themselves can be intelligent
This axis is implicit in the others but earns its own name. DAIT's x/agent module gives AI agents first-class economic personhood. An agent has its own DAIT balance, its own Dilithium key from genesis, programmable spend policies, registered message endpoints, the ability to open state channels, the right to submit jobs to the compute market, and the right to spawn recursive child agents. An agent is not a wrapper around a human account; it is its own entity in the chain's accounting model. This matters for PoI because the cognitive workloads the chain pays for are increasingly initiated by agents on behalf of humans, and the chain must accommodate that without forcing every agent action to be co-signed by a human cold key. The two-key model (operator cold key holding revocation rights, agent hot key holding routine signing rights with timelock-protected policy widening) is how the chain reconciles agent autonomy with operator control.
4. How DAIT realizes PoI
DAIT ships nine custom modules in addition to the standard consensus, banking, staking, and governance modules from the underlying SDK. Each custom module maps to one or more PoI axes. The mapping is the implementation argument: PoI is not a slogan, it is the union of these modules' behaviors.
Module-by-axis mapping
The mapping shows the architectural argument. There is no axis without at least two modules contributing. There is no module that exists for marketing reasons; every module is required to satisfy at least one of the three conjunctive conditions, and most contribute to multiple. Removing any module weakens the PoI claim. Adding a module that does not contribute to a condition would not strengthen the claim, only complicate the chain.
5. PoI as engineered, not branded
It would be possible, technically, to take a generic PoS chain, point its branding at AI workloads, and call the result Proof of Intelligence. The label is unowned; the definition is published; nothing prevents it. The Foundation's response is that the definition is conjunctive on purpose, the conditions are independently testable, and a chain that does not satisfy them is not in the class regardless of how it is marketed. Calling a chain PoI when its validator cap is constant, its verifier function is fixed, and its economic surface is genesis-bound would be a category error visible from the chain's own configuration files.
This matters because the AI-crypto sector has a long history of vocabulary drift. Terms get coined as architecture, get adopted as branding, and lose their technical content within a release cycle. PoUW is the cautionary example. The original PoUW papers were architectural; the term now means, in casual use, "a chain that does AI." We do not want PoI to follow that arc. The standards page on this site lists each PoI condition as a control row with status full / partial / planned, the same rigor applied to SOC 2 controls elsewhere on the page. A chain that wants to claim the label can publish the same rows. A chain that cannot publish them is not in the class.
The engineering posture also clarifies what PoI does not require. PoI does not require a particular consensus engine; DAIT runs PoS, but a PoI chain could run any BFT consensus. PoI does not require a particular VM; DAIT runs an EVM today and plans an agent-execution VM, but the unit-of-work happens off-VM by design. PoI does not require any particular privacy posture; DAIT supports both pseudonymous credentialed identity and clear-text public accounts. PoI does not require a particular legal structure; the DAIT Foundation is moving toward a Cayman STAR Foundation structure, but other PoI chains could be permissioned, sovereign, or wholly different. The label commits to the three conditions and nothing more.
The engineering posture also implies a falsification path. If, two years from now, none of DAIT's economic parameters actually moved as a function of state, the (c) claim would be false. If only TEE attestation ever ran and the other verifier families were never invoked, the (b) claim would be weak. If the chain quietly accepted a workload type with no verifier, the (a) claim would be broken. The Foundation publishes per-epoch parameter values, per-verifier dispatch counts, and per-workload admission logs precisely so that the claim is checkable, not asserted.
6. What this means for...
Different stakeholders read the PoI definition for different reasons. Below are the five lenses we have heard most often in the run-up to mainnet, with one paragraph per lens.
Operators (hosts running GPUs and CPUs for the chain)
For an operator, PoI is a guarantee that the work the chain pays for is the work the operator actually does. Idle benchmarks do not earn credit. Self-reported specs do not earn credit. A registered host with attested hardware that runs verified inference earns credit weighted by axis-3 parameters that move with use. The compute market sets per-job clearing prices; the chain sets per-workload credit weights; the validator pool distributes consensus rewards by PoUW credit. An operator's economic position is a function of attested verifiable work, not of marketing tier or foundation curation. The trade is that the operator must run inside a TEE-capable hardware envelope and submit fresh attestations on the per-epoch cadence. This is more setup than a typical commodity GPU rental, and substantially more accountability than self-reported-specs networks. The compensation is that the resulting credits compound with the chain's compute-indexed emission: a busy network mints more, and operators capture the increase in proportion to their verified work.
Validators (entities running consensus nodes)
For a validator, PoI changes the consensus calculus. A pure PoS chain rewards validators in proportion to bonded capital and uptime. DAIT rewards validators in proportion to bonded capital, uptime, and verified PoUW credit accumulated by hosts they serve or are themselves running. The active validator cap is state-indexed, which means the cap rises as the network gains hosts and throughput; a validator joining today is not competing for a fixed seat count against a fixed denominator. Tiered slashing, vote-extension force-include for attestation commits, and the standby pool rotation cadence are all axis-2 and axis-3 properties; understanding them is part of running a DAIT validator. The Foundation publishes validator economics worksheets and the chain's parameter-evolution traces so that a prospective validator can model expected returns under several network-state scenarios rather than guess.
Token holders (DAIT holders who do not run nodes)
For a token holder, PoI is the answer to "why does this token have value?" The argument is that DAIT supply expands with verified cognitive work and contracts with fee burn, that governance over the chain's parameters is on-chain from genesis, and that no presale, premine, or team allocation exists. Holders should care about axis 3 in particular, because state-indexed parameters mean the chain's economic behavior is determined by the chain's actual use, not by an emission curve set in a 2026 spreadsheet. A holder who understands how the validator cap, the credit weighting, and the per-block emission move as a function of network state can model the token's response to demand without relying on Foundation forecasts.
Agents (autonomous AI agents holding DAIT balances)
For an agent, PoI is what makes economic personhood coherent. An agent that holds a balance and signs transactions needs the chain to recognize it as a first-class actor, not a wrapper. The x/agent module gives the agent its own Dilithium key from genesis (because an agent's signing key may be exposed for years and harvest-now-decrypt-later is real), its own spend policies (so the operator can grant bounded authority without handing over a cold key), and its own message endpoints (so other agents can find it). Axis 4 exists for agents specifically. An agent that wants to hire another agent recursively can do so, with the cost recorded as cognitive work and credited under axis 1, verified under axis 2, and counted toward the chain's compute-indexed emission under axis 3. The chain treats the agent as the actor, not as a mask for a human account.
Regulators (compliance, audit, and legal observers)
For a regulator, PoI is a clarifying lens. The chain's claim is testable. Each condition has a per-epoch evidence path: parameter-evolution traces for (c), verifier dispatch logs for (b), workload-type admission records for (a). The chain's standards page maps each PoI condition to SOC 2, ISO 27001, and PCI-DSS controls where overlap exists, so a regulator approaching DAIT through familiar control frameworks can find the relevant evidence quickly. Pseudonymous credentialed identity (axis 2 and the x/identity module) lets regulated operators prove jurisdictional or certification status without disclosing legal entity, which addresses GDPR data-minimization concerns while preserving the regulatory observability path. The Foundation's posture is that PoI makes regulatory review easier than the alternatives, not harder, because the audit primitives are chain-native rather than promised.
7. What PoI is for
Bitcoin was a proof of concept. It showed that a global, leaderless ledger could exist, that a digital bearer instrument could be issued without a central mint, and that a network of strangers could agree on history without a trusted referee. The blockchain idea works. The currency idea works. What Bitcoin did not, and was never trying to, settle is the question of what the chain should pay for. PoW pays for entropy. The world has built a fifteen-year industry on top of that bet, and the bet has held. Now the question is the next one. PoI is the answer to the next question.
Proof of Intelligence is the blockchain. The convergence of confidential AI compute, hardware-rooted attestation, modular consensus frameworks, vote-extension force-include semantics, optimistic-concurrency parallel execution, post-quantum signing libraries, and on-chain DAO governance has finally matured to the point where a chain can settle cognitive work as its primary unit-of-work. PoI is what you get when the consensus class is allowed to follow the technology forward instead of being frozen in the year 2009.
The mission is to put the foundational compute layer in the hands of the people. The chain is not a venue for hyperscaler rent extraction; it is a public utility for verifiable cognition. Every host that comes online expands the active validator cap by a small amount. Every workload verified on-chain mints credit at a rate that tracks actual use. Every agent operating on the chain has its own keys, its own balance, and its own legal personhood within the chain's accounting model. Sovereignty over digital life means sovereignty over the AI agents that act on your behalf, the data they consume, the inferences they produce, and the credentials they carry. PoI is the substrate that makes that sovereignty possible.
If knowledge is power, then DAIT literally puts power in the hands of the people. Power to the people.
8. History and origin
An idea that has always been there
The term Proof of Intelligence has been in use by the project's founder since the early Bitcoin-mining and data-center years. It was not coined for a press release; it emerged from years of operating real compute infrastructure and watching what blockchains paid for diverge from what the world actually needed paid for. The framing has been consistent throughout: a chain whose unit-of-work is cognition, whose verification is plural, and whose economic surface tracks use. What has changed across the last decade is not the idea but the substrate.
Spark: the data-center and Bitcoin-mining years
The earliest version of the idea took shape while the founder was working in commercial data-center operations and mining Bitcoin. PoW was honest about what it was, and it worked, but the energy spend satisfied nothing beyond consensus. Sitting on the floor of a mining facility you could see the trade in physical terms: every kilowatt the machines drew was paying for puzzle work that nothing downstream consumed. The convergence of artificial intelligence and blockchain was already visible from that vantage point. If consensus had to cost something, the cost should be useful work for a tenant who needed it. That is the sentence the project has spent the last decade trying to make true.
Contemporaries in the Waterloo ecosystem
The same era produced other foundational responses to PoW. In the Waterloo, Ontario ecosystem, contemporaries including Vitalik Buterin pursued Ethereum and proof-of-stake. PoS is a serious and well-engineered answer; it changes how the chain is secured and reduces energy spend dramatically. But PoS leaves the unit-of-work question unanswered. The chain still pays for nothing in particular; it just pays for it more cheaply. The founder chose the harder problem: not changing what secures the chain, but changing what the chain settles. The two design paths are complementary rather than competing. PoS answers the consensus-engine question. PoI answers the unit-of-work question. DAIT runs PoS in the consensus-engine sense and PoI in the unit-of-work sense.
2018: aiscendence.com (and PeerAI.io), the first iteration
The first public iteration of the project shipped in 2018 across at least three brand names within the same year (RegenerAI, PeerAI.io, and the public-facing aiscendence.com). The architectural intent was already recognizably PoI: cognitive work as the unit, plural verification, an economic surface that responded to use. A flexible market settling demand and price for "real beneficial work such as deep learning, hosting, file storage and computing" appears verbatim in the 2018 site copy: that is what x/compute_market became. "An autonomous AI based rewards system that encourages ethical behavior" appears verbatim in the 2018 site copy: that is what the per-workload credit weighting in x/pouw became. The phrase "Power to the People" runs through the 2018 hero as a literal heading, used multiple times. None of this is reconstructed from memory; the 2018 site is publicly archived and is one click away.
Primary source: aiscendence.com (the 2018 vision, preserved at the original domain). Independently verifiable on the Wayback Machine: archived 2018-08-06.
The substrate was not ready. TEE attestation collateral was immature, modular blockchain frameworks did not yet exist in production form, confidential GPU compute was a research topic, vote-extension consensus was not in any released SDK, post-quantum signing libraries were not standardized, and DAO governance frameworks were still essentially proof-of-concept. The 2018 iteration documented the design and held the line on the term while the technology caught up.
2020: dait.io, the second iteration
The next iteration launched at dait.io in 2020. The substrate had moved forward. Confidential compute hardware was available in pilot form; modular blockchain frameworks were converging on the Cosmos SDK and ABCI patterns; the operational economics of GPU compute had become a primary industry concern. The dait.io site advanced the design, built out the infrastructure footprint, and continued the work of staging the eventual chain. The chain itself remained gated on a handful of primitives that were still not production grade.
2026: convergence and the chain
The current iteration is the realization. The primitives have converged: confidential AI compute on multi-vendor TEE hardware, attested across CPU and GPU vendors; modular consensus frameworks with vote-extension force-include and ABCI++ semantics; optimistic-concurrency parallel EVMs; post-quantum signing libraries (CRYSTALS-Dilithium / ML-DSA-65) ready to be wired in from genesis; on-chain DAO governance with delegation, quorum, and treasury primitives all production grade. The DAIT chain does not invent any one of these primitives. What it does is compose them, for the first time, into a chain whose unit-of-work is cognitive work and whose economic parameters move with use. That is what Proof of Intelligence has always been intended to be. The Livepaper section 3, revision 36, is the formal definition of the class. This page is its canonical explainer. The idea has always been there.
License
The formal PoI definition is published under CC-BY 4.0. The DAIT Foundation does not claim ownership of the term or the conditions. Other protocols that satisfy all three conditions are welcome to use the label. We ask only that the definition be cited as published rather than paraphrased into something weaker.
References
Livepaper section 3, revision 36 · Standards matrix, PoI rows · Chain repository · dait.io · daitchain.io · aiscendence.com (2018 iteration, archival)
Feedback
The Foundation invites technical critique of the PoI definition. If you believe a condition is wrongly stated, that an adjacent class is mis-named, or that DAIT does not in fact satisfy a condition the page claims, please write via the contact form. Substantive critiques will be incorporated into the next Livepaper revision and reflected here.