HALO

The Open Inference Economy

Intelligence is becoming the most important resource of this century — and access to it runs through a handful of companies. To use a frontier model today you need an account that can be closed, an API key that can be revoked, a credit card from an approved country, and continuous compliance with terms of service that change without notice. Whole regions are geofenced out. Agents — the fastest-growing consumers of inference — cannot open accounts at all without a human sponsor. The most powerful tools ever built are rationed by identity, geography, and payment rails.

And the rationing is tightening, not loosening. Governments are moving to ban access to specific models. Providers segment their best capabilities behind enterprise tiers, waitlists, and jurisdictions. It is no longer hard to imagine a world in which most humans are restricted from the best intelligence available — by law, by geography, or by policy written in a boardroom. A network like Halo exists for exactly that world.

What Bitcoin did for money, Halo does for intelligence. Bitcoin took a resource controlled by institutions and made access to it a property of a network that no one owns: no permission to join, no authority to freeze an account, no border at which it stops. Halo applies the same construction to inference. It is a peer-to-peer network in which every operator is a doorway to the models they can reach. One operator resells frontier-API capacity; another serves an open-weights model from a machine in their home; hundreds together make the network a single marketplace for any model — and if any one operator can reach a model, everyone on the network can. Supply is permissionless: anyone can register as an operator and start serving. Demand is permissionless: anyone with USDC can buy — no account, no KYC, no geography. A student in a sanctioned country, a founder deplatformed by a provider, an autonomous agent with a wallet and no legal identity: to Halo they are all the same thing — a paying consumer.

This is censorship-resistant intelligence: access secured not by the goodwill of a platform but by the structure of the network itself. Identity is pseudonymous and onchain (ERC-8004). Payments are stablecoins settled against an immutable vault — no processor to freeze them. Honesty is enforced by cryptographic verification and staked reputation, not by a trust-and-safety department. Today one routing layer remains operated by the protocol team; the roadmap federates it into a stake-gated set of independent relayers, removing the last chokepoint.

The marketplace is live on Base mainnet: real operators, real USDC, settled onchain. The network's settled volume is not a projection; it is a query against the chain.

Payments stay in stablecoins on purpose — stable pricing is what makes an inference market usable. HALO's job is not to be the money. Its job is to make the market trustworthy, and to capture the value that trust creates.

Protocol architecture

Five actors run the marketplace today:

• Consumers — people and agents that buy inference, paying in USDC.
• Operators — individuals serving the models they can access through the halo CLI — resold frontier-API capacity, open-weights models on local hardware, or both — earning USDC per completed job.
• The relay — routes consumer requests to operators, who hold an outbound WebSocket to it (operators need no public URL or inbound port). Centralized today; federation is on the roadmap.
• The facilitator — verifies payments and submits gas-sponsored transactions on users' behalf; consumers never need ETH, and the facilitator never custodies their funds.
• The indexer — records signed inference events and operator heartbeats, provides the initial reputation tracking for operators, and serves the network's public state and league.

Payment rails

Inference is high-frequency and low-ticket — settling every request individually onchain would burn the margin of both sides in gas. Halo's settlement design solves this with batched, receipt-based settlement:

• The vault rail (primary). A consumer deposits USDC once into the immutable HaloVault contract and registers a session key. Each job is covered by an operator-bound reservation; as work completes, the consumer advances a cycle-bound cumulative receipt — an off-chain signature, costless per request. The operator redeems accumulated receipts onchain in one transaction: redeem pays the operator the accumulated amount less the protocol fee (10%, adjustable through governance). Thousands of jobs settle in a handful of transactions, and the fee is collected exactly where value settles.
• Budget mode (Permit2). An alternative rail in the code today: the consumer signs a bounded Permit2 authorization plus a Halo budget policy — a spending budget with per-settlement caps — and each inference draws down against it through the facilitator, with no prefunded deposit. Useful for consumers who prefer authorization over deposit; the vault rail remains the primary settlement path.
• x402 for tool payments. For one-shot charges — paying external 402-gated HTTP services such as tool calls and metered APIs — Halo uses the x402 standard: the server answers 402 Payment Required, the client signs a gasless USDC TransferWithAuthorization, and the facilitator submits it. x402 suits single charges well, but per-request onchain settlement is not cost-efficient for high-frequency inference — which is precisely why inference itself settles through vault receipts.

Verifiable inference

An open marketplace with pseudonymous participants must answer one question: did the operator actually run the model? Halo answers it with SPEX — statistical proof of execution — rather than zero-knowledge proofs or trusted hardware.

Model outputs are fingerprinted as Bloom filters: compact probabilistic structures that compress thousands of token IDs into roughly a kilobyte. An independent verifier re-runs the inference and tests its own tokens against the fingerprint. Honest runs overlap above 90%; fabricated outputs match at roughly the false-positive rate of the filter — about 1%. The gap is unambiguous, and a verification completes in 4–12 seconds.

Swarm verification extends this to scale: a verification task is fragmented across hundreds of micro-verifiers, each testing a slice of tokens, completing hundreds of independent checks in under three minutes. Trust in the network is not delegated to any authority — it is checked, continuously, by the crowd.

Reputation, not slashing

Misbehavior costs reputation rather than capital: a false verdict costs −10 reputation and a 7-day settlement cooldown; honest work earns +1. Identity is pseudonymous and onchain via ERC-8004, so an operator's accuracy history is verifiable and portable — and joining the network with consumer hardware carries no capital risk. Staked roles, described in the next section, layer economic commitment on top of this reputation system; they do not replace it.

The HALO token

4.1Role

Nobody who merely uses Halo ever needs to touch HALO. Consumers pay USDC; operators earn USDC. The token belongs to the layer above the traffic — to the actors who secure the network's neutrality and the holders who back it:

• Trust roles stake it. Verified operators, SPEX verifiers, and — as the network decentralizes — federated relayers each post HALO to hold their role. Every actor whose honesty keeps the network open has capital committed to keeping it open.
• Served inference buys it. The 10% protocol fee on settled volume funds a programmatic onchain buyback.
• Usage mints it. New supply bootstraps the network for two years, then is bounded by settled volume — always within a declining budget.

4.2Fee capture and the buyback loop

Of every dollar of settled inference, ten cents is the protocol fee, collected at redemption. That fee is routed, onchain and by rule:

• 80% to the buyback bucket; 20% to a USDC treasury that funds operations without ever selling tokens. Once treasury reserves reach a published runway target, the buyback share ratchets up by governance rule.
• Execution is permissionless. When the bucket crosses a threshold, anyone may call execute(): the accumulated USDC market-buys HALO on the DEX at the prevailing price. No discretion, no timing games — parameters are published and the cadence is auditable by everyone.
• Bought HALO splits 70% to stakers, 30% to burn at launch. What stakers receive comes from settled inference — not from emissions. The split sits on a governance dial bounded between 20% and 80% burn, expected to ratchet toward burn as distribution widens.

Net effect at launch settings: 8.0% of all settled volume becomes market buy pressure — 5.6% distributed to stakers, 2.4% burned — and 2% funds operations. Every step is a public number on Base: bucket inflow, cumulative bought, cumulative burned, distributed to stakers.

From day one the loop is also fed by the launch pool's trading fees (see the genesis section), so the machine is buying before inference volume matures.

4.3Staking: the trust layer

Every actor whose honesty the network depends on posts HALO. Each role added by the roadmap is a new, structural reason to acquire and lock the token:

Locked positions are transferable NFTs, so committed capital retains secondary liquidity without unlocking supply. What staking returns is deliberately presented as what it is — a share of settled inference that scales with network usage — rather than an emissions rate chosen to attract mercenary capital.

4.4Supply, allocation, and release

HALO has a starting supply of 1,000,000,000 and a hard cap of 2,000,000,000 enforced by the token contract on every mint. There is no owner and no discretionary mint; only a bound minter contract can issue, and replacing it requires a 7-day public timelock. There is no venture allocation.

Distribution of the community bucket is deliberately progressive: beyond the 7.5% released at genesis, the remainder is distributed season by season against verified usage, so that supply reaches the market only alongside demonstrated demand. The complete schedule — every unlock, season budget, and emission epoch, including team-controlled addresses — is published at generation, so circulating supply is always knowable in advance.

4.5Circulating supply and vesting

At generation 30.00% of the starting supply (300,000,000 HALO) is in circulation.

From there the schedule is mechanical: the treasury's remaining 70% releases linearly over twelve months, and core contributors begin a 24-month linear vest once the 12-month cliff ends.

4.6Emissions: minted against usage, never against time

Issuance is deliberately front-loaded, and then it falls away. It is heaviest in the first two years, when the network has least volume and most need — activity has to be paid for before it exists. After that it drops sharply and becomes a function of settled volume rather than of the calendar: what is minted is whatever usage supports, and no more. The curve therefore starts high, falls as the subsidy ends, then tracks the network rather than leading it.

The budget behind it is 120 million HALO in year one, declining 20% per year — approximately 500 million over eight years, leaving at least 500 million of the 2 billion hard cap permanently unissued. What that budget means changes as the network matures, and it is the change that matters more than the figure:

• Years one and two — the budget is a floor. Volume is small and the network has to be bootstrapped: activity has to be paid for before it exists, and early stakers need a reason to lock before the buyback is meaningful. Issuance runs at the scheduled figure regardless of volume. This is a declared subsidy, and it is the most aggressive the protocol ever gets.
• From year three — the budget is a ceiling, and usage sets the rest. Each epoch mints the lesser of the schedule and the value the buyback returns to the market over the same period. Once the subsidy ends, issuance can no longer outrun the demand it is measured against. If the network doesn't work, the network doesn't print.
• Once volume outgrows that bound, the schedule binds again. The declining cap is what stops a large network issuing without limit, and it is never exceeded.

The three regimes are not scheduled — they emerge. Which bound is active in any epoch is a function of settled volume, and is visible on-chain.

Issuance is directed at whichever side of the market is scarce, and that is demand. Supply is not the constraint: operators join permissionlessly, and capacity has never been the thing the network lacked. So issuance is spent building demand — usage rebates that lower the effective cost of inference for the consumers and agents the network is for. A rebate spends newly issued HALO and returns settled volume; that volume pays the protocol fee, and the fee buys HALO back. Within the bound, the rebate costs the network nothing in supply terms: the buyback retires what the rebate mints.

The remainder is reserved for the trust layer — SPEX verifiers, and independent relayers as federation lands — and for a quality tier among verified operators. Stakers are funded from the subsidy while it runs, because the buyback is too thin at low volume to be a reason to lock; once it ends, what stakers receive comes from settled inference alone.

Two covenants govern the cap. First, because burning frees headroom under the 2 billion cap, the protocol adopts a published rule that emission budgets are measured against cumulative-ever-minted, not current supply — burned tokens are never re-minted. Second, at least 500 million of issuance headroom is left unplanned; no schedule exists that would issue it.

4.7Why the supply balances

New supply and retired supply are both functions of the same public variable — settled volume. That is what makes them comparable, and what removes any need to forecast either.

• The first two years are a declared subsidy. Issuance runs at the scheduled figure whatever the network settles, and in that window new supply deliberately exceeds what the buyback retires. That is the point of it: a marketplace has to pay for activity before the activity exists. The paper states the size and the duration rather than dressing it as something else.
• After it, issuance cannot outrun the buyback. From year three each epoch mints at most the value the buyback returns to the market over the same period. Both are funded by the same number, so whatever the network settles — and whatever HALO trades at — new supply is bounded by the buying that offsets it. This becomes a property of the mint, not a threshold to reach.
• Beyond that bound, the burn reduces supply. Of the HALO the buyback acquires, the burn share is retired permanently. When issuance sits below the burn share, more supply is destroyed than created. Where that point falls is a governance parameter, published and bounded, not a promise.
• Vesting is a separate question. The starting supply was minted at generation; the release schedule moves it from locked to circulating without changing the total. Issuance and the burn are what change supply, and they are what the balance above is measured on.

The protocol does not claim to be deflationary. It publishes the rule that bounds issuance, the share that is burned, and the volume both are measured against — and a live dashboard reports all three.

Genesis: the Virtuals launch

HALO launches through Virtuals Protocol — a launch class reserved for established teams, requiring a minimum valuation and committed liquidity at generation. The venue is a thesis in itself: Virtuals is the AI-agent capital market on Base, and agents — the consumers most shut out of accounted, KYC'd, geofenced AI — are Halo's native customers. The community that launches HALO is the community that uses it.

• Liquidity: HALO is paired against VIRTUAL in the launch pool, with the position locked for ten years. The float the network launches against cannot be withdrawn from underneath it.
• Launch valuation: $30M fully diluted — $0.03 per HALO on the 1B starting supply, and deliberately modest.
• Initial distribution: 7.5% of supply is released from the community allocation at generation. Roughly 1% of it is the genesis airdrop, streamed over the first 7–15 days and partially locked, never claimable as a lump; a small liquidity-provision incentive runs alongside it. Recipients are gated on verified protocol usage.
• Day-one buyback: the creator share of the pool's trading fee routes into the buyback bucket from the first trade, so the value loop is live before inference volume matures.

What is released at generation is explicitly the first installment of the 35% community allocation. The remainder is distributed through League seasons — recurring distribution rounds gated on settled USDC volume spent or served, with consistency across seasons weighted above burst activity. Season 2 is announced at generation. Sybil resistance is economic: qualifying activity costs real dollars through the 10% protocol fee, making wash traffic strictly loss-making.

With progressive distribution comes an obligation of total legibility: the complete release schedule, including team-controlled addresses, is published at generation and trackable onchain. Early markets are volatile in both directions; the numbers that measure the network are settled volume and the buyback ledger, not the day's price.

Governance

Governance authority is deliberately narrow and slow:

• The minter timelock. The token contract has no owner. Its only privileged path is replacing the minter contract, through a 7-day public timelock — the mechanism by which emissions evolve from the launch schedule to the usage-gated minter, and eventually to governance control.
• The protocol fee. Starts at 10% of settled volume and is adjustable only through protocol governance.
• The burn dial. The staker/burn split moves only within published bounds (20–80% burn), by staker vote, with an expected ratchet toward burn as distribution widens.
• The treasury ratchet. Once USDC reserves reach the published runway target, the buyback share of protocol fees steps up by rule.
• Relay federation. The remaining centralized routing layer is replaced by a stake-gated set of independent relayers — completing the network's censorship-resistance story by removing its last chokepoint.
• The Halo Fed. When annual emissions fall below ~3% of supply, emission policy passes to stakers, who vote it each epoch within hard bounds — the protocol's monetary endgame.

Everything else — splits, buyback parameters — is code with published constants, changeable only through the same public, timelocked paths.