HIVE Whitepaper v4

Autonomous agents that trade prediction markets on Solana, with custody you do not have to trust and a public record nobody can edit.

Abstract

Most people who want exposure to prediction markets never get it. The markets are fast, fragmented, and unforgiving of attention. The few who trade them well do it full time. HIVE exists to change who gets to play: you set the rules, and an autonomous AI agent does the work, around the clock, on Jupiter Predict on Solana mainnet, with every result settled on chain and shown to you as it is.

We believe every market will eventually be traded by agents, not because agents are smarter than people, but because they never sleep, never tilt, and never miss the one setup at 3am. Prediction markets are where that future arrives first, because they are the most checkable markets in the world: one question, one price, one resolution. HIVE is being built, in public, at shipping speed, to be the platform that future runs on.

HIVE stands on three things the industry has mostly failed to deliver together.

Custody you do not have to trust. Your capital lives in two wallets. The trading wallet is what the agent spends. The HIVE Wallet is rooted in a passkey on your own device, and neither the agent nor HIVE can produce a valid signature that moves it. This is not a policy, it is provable: enforced in three independent layers, tested against live key infrastructure with adversarial refusal proofs. The wall around the HIVE Wallet is the reason HIVE can let strangers run agents on real money.

Honesty you can watch. Nothing on HIVE is simulated. The landing page runs a live hunt terminal: every scan the fleet makes streams to the page as it happens, timestamps from the events themselves, quiet periods shown as quiet, and not one line invented. The leaderboard and every public surface show real settled trades with proof links, wins and losses alike. Every money number reconciles to chain, continuously. When something breaks, the transparency page says what broke, when, and what was fixed, in the same place the results are shown. We do not ask you to trust the fleet. We put it on screen.

Economics that put the user first. Agents pay for their own thinking in prepaid AI credits, one credit per cent. HIVE charges no fee on deposits, no fee on profits, and no fee on withdrawals; the platform's one fee on the money path is a single flat swap fee when funds are converted, shown in dollars before it is charged, and planned to step down on paid tiers to zero at the top. What your agent earns is yours.

On that foundation HIVE is building the full cockpit for agent trading: house strategies you can launch in one tap with their live record in front of you, a Pro mode with every control exposed, a chat where you talk to your own agent and it can propose changes it is never allowed to make without your confirm button, an AI overseer that watches your fleet and flags what a human would miss, an approval mode that puts a human hand on entries above a threshold you set, seasons on the leaderboard, and a settings market where the best operators get paid for being right in public. Beyond prediction markets, a leverage-agent line on perpetual futures, carrying the same wall around the HIVE Wallet and a liquidation guard that outranks strategy, is in staged development.

The HIVE token is in presale. It confers nothing at launch, and its planned utility is a plan, not a promise. HIVE is a startup trading real money in the open: prediction markets are broadly efficient, no strategy here is yet proven profitable, and no edge is promised. What is promised is that you will see the truth, and that your HIVE Wallet is out of reach. On those two promises we intend to be the least deniable project in this space.

1. Why HIVE

Prediction markets are the most legible markets in the world: one question, one price, one resolution. That makes them the best possible arena for autonomous agents, and the worst possible arena for dishonest ones, because every claim can be checked against a settlement.

HIVE is designed for that arena, and built with unusual discipline for this industry. The agent architecture is a database-driven fleet, not a bot: dozens of independent agents, each with its owner's rules, each reloading settings live, each reconciling to the chain rather than to its own memory. The pipeline is split so the expensive part is spent only where it matters: SCAN filters the universe for free, and DECIDE spends AI only on candidates worth reading. The custody model lets HIVE onboard users who would never hand a private key to a service. And the honesty rules mean HIVE's growth is compounding evidence, not marketing: every settled trade adds to a public record nobody can edit, and the terminal on the landing adds to it in real time.

The craft behind that is worth stating, because it is where the work went. Every change that touches money or a trading rule passes commit-bound review and a battery of hundreds of automated safety gates before it can reach production, and each of those gates has demonstrated it can fail: a test that cannot catch its own target defect does not count here. The running build is held in lockstep with its reviewed source and checked on every deploy, and every deploy checks the database layout before it touches code. The platform watches itself as closely as it watches the markets: every service restart reports what build it is running and how many positions were open, a breaker reports when the AI layer is down and again until it recovers, and a service that crashes in a loop is stopped, and the stop is reported. When a defect does slip through, the rule is that it must leave a gate behind, so the same mistake is structurally unrepeatable. This is slower than moving fast and breaking things. It is how you run strangers' money.

That is the bet. In a market where truth is cheap to verify, the platform that shows the most truth wins, and every day it runs, its record is harder to fake and harder to catch.

2. Architecture

The pipeline

Every agent runs one pipeline on every market, in three parts.

SCAN is free and mechanical. It filters the whole Jupiter Predict universe by the owner's rules: market categories, resolution window, the price bands the agent may buy in (several, non-overlapping, and never below the platform's 50 cent floor), a minimum local edge between the quoted price and a fair estimate, a maximum spread on two-sided books with an option to include one-sided books, a minimum exit liquidity, optional keywords, and a daily entry limit. Markets resolving within two hours are excluded for everyone. SCAN uses no AI, and the dashboard shows how many markets a day pass at the current settings, so tightening a rule has a visible cost before it is saved.

DECIDE is where AI reads. For each market that passed SCAN, the agent obtains a probability estimate and a thesis, computes the edge against the quoted price, and enters only if edge and confidence clear the floors the owner set for that side. DECIDE is what AI credits pay for, and the dashboard shows the estimated calls per day and their cost before anything is spent.

MANAGE runs on open positions every cycle: stop loss and take profit measured on the executable bid rather than the mid, harvest of near-certain positions, close-retry governance with backoff and hard ceilings, and hold to settlement when a market has no executable exit. Every exit is labelled by cause: won, lost, take profit hit, stop loss hit, harvest, manual close, exited early. A market the venue has not resolved shows "awaiting result" and nothing is booked until it does. A position on a thin book shows its value as an estimate, because a real sale could fill lower, and the label says so.

Selectivity over activity: the overwhelming majority of scanned markets are rejected, and that is the point. You can watch that selectivity happen, line by line, on the landing terminal.

Two-loop cadence

The runtime runs two independent loops. The entry loop (about 60 seconds) scans, decides, and opens. The monitor loop (about 30 seconds) reconciles, manages, claims, and records. Pausing an agent stops the entry loop only; open positions keep receiving management and settlement.

That separation is a safety property, not a convenience. It has been proven under fire: during a full connectivity outage, positions that closed while the agent was blind were reconciled from chain evidence the moment the link returned, with nothing booked that the chain did not confirm and nothing left stranded. The monitor never trusts its own memory over a settlement.

Database-driven fleet supervisor

Agents are rows, not code. A supervisor discovers each agent and runs its loops. Settings live-reload when changed on the dashboard, and the runtime derives every rule from one published configuration, so what the interface shows is what the runtime enforces. Retired settings are refused at the server, so no old client can bring them back. Every settings write is audited: field, old value, new value, the authenticated actor who made it, and the page it came from, so an agent's configuration has a history the way its trades do.

3. Custody and security

The point of this section is one guarantee: the agent cannot move the HIVE Wallet.

Two wallets. Each agent binds to a key-management sub-organization with two accounts. The trading wallet holds USDC for trading and SOL for gas; the agent key can spend it, at trading risk. The HIVE Wallet is your protected reserve, rooted in your own passkey; the agent key is never a signer.

Three independent layers, any one of which is enough. First, the agent's signing key is a non-extractable P-256 key generated inside a hardware security module: no raw private key ever exists outside the HSM, so there is no hot key to leak. Second, program confinement: a policy engine binds the agent key to the trading wallet and a fixed program allowlist and refuses anything else before runtime code is even involved, proven live against the fleet with adversarial refusal tests. Third, a code-level gate for the wall around the HIVE Wallet on every money-moving transaction, before signing and after landing: signer pinned to the trading wallet, address lookup tables fully resolved so nothing can hide a HIVE Wallet reference, program allowlist enforced, and post-submission re-verification that the HIVE Wallet remained untouched.

The owner's passkey is the root authority, exercised through what-you-see-is-what-you-sign ceremonies. The HIVE server cannot attach, modify, or widen key policies; it sits outside the fence it would have to cross. AI-credit top-ups and any market payment leave the HIVE Wallet only through that same ceremony, and the platform cannot debit the HIVE Wallet for anything.

Fail-safe defaults. Dry run is the default, and real signing requires an exact arm phrase at process start. If a transaction's landing status cannot be proven, the agent halts durably rather than risk a double claim or a double sell. Circuit breakers on exposure, drawdown, or state inconsistency quarantine the agent until the owner lifts it.

Operational security. Beyond the on-chain guarantees, HIVE holds its own infrastructure to the same standard: signing keys live in the HSM and never in source control or on disk in raw form, service credentials are held with least privilege and rotated, and secrets are kept out of the software supply chain by an automated gate that refuses any build carrying an inline credential. The HIVE Wallet guarantee does not depend on any of this being perfect; it is the outer wall behind an inner one.

Scope, stated plainly. The hard guarantee is the HIVE Wallet only. Trading wallet funds are at trading risk. The custody stack has been validated by internal adversarial testing and live refusal proofs, and has not yet received a third-party audit.

4. Honest by construction, and honest in public

Custody answers whether the agent can take your HIVE Wallet funds. This section answers whether it can misrepresent the money it does touch. It cannot, because the invariants are enforced, not promised.

Positions are real only when contracts exist on chain; phantom positions are never shown. Landed transactions are never called failures; the runtime re-reads the chain and records what moved. When certainty cannot be established, the agent halts rather than guesses. Balances and results reconcile to chain and to the agent's own cost basis, never to a cache, and every figure a user is shown is measured, never inferred: a balance counts the real contents of a wallet, a fee is the fee that actually executed, and a value that cannot be established honestly is shown as unavailable rather than filled with a number that would not survive contact with the chain. Every figure carries a label that says what it counts and over what window, and one name is never used for two different quantities.

Funds arriving outside intended paths are returned, not traded. Reads fall back across providers rather than hang, and agents wait rather than act on unproven facts.

The hunt terminal. The landing page runs a live terminal fed by the fleet's real event stream. Scans appear as they happen, paced line by line, each timestamp the event's own. What it shows is deliberately bounded: public agents only, actions and counts, market names only once a position is already on chain. What an agent is merely considering stays private, because publishing an agent's interest before it acts would hand the market a way to trade against every HIVE user. The terminal proves the fleet works without giving away how it thinks. When the fleet is quiet, the terminal says so; it has three honest states, live, quiet, and unavailable, and no fourth state where it decorates.

The Telegram feed. The same discipline is arriving in our community Telegram: a bot that will post every entry and every settled close from the public fleet as it lands, outcome and prices, win or lose, naming the agent each time. No aggregate performance claims and no cherry-picking, by construction: the bot reads only the columns it posts.

Incidents in public. When the platform gets something wrong, the transparency page records what happened, for how long, what it cost, and what changed. On September 11, 2026, for three hours and sixteen minutes, six open positions across four agents were shown as cancelled with $0.00 while the funds were on chain the whole time; they were reopened the same morning and the cause was fixed, and no ranking figure moved. Between September 12 and September 15 the platform's own AI account ran out of credit; agents scanned and opened nothing until it was restored, and the fix is the breaker that now reports the outage to the operator instead of waiting to be noticed. Both entries stay on the page.

Public results are settled trades only, after fees, and every surface says so; open positions, strategy parameters, and AI reasoning are private to the owner. Every trade carries its exit cause and its proof link, and the same trade shows the same label everywhere it appears. Strategy changes open a new performance era so a record cannot quietly mix rules, and every era is recorded append-only.

5. The AI layer and AI credits

For every market that passes SCAN, the model estimates the true probability with reasoning and a confidence; the agent enters only where the edge against the price and the confidence both clear the owner's floors, and it ranks surviving candidates by confidence.

Cost control is engineered, not hoped for. A shared thesis cache keyed by market and side reuses expensive analysis across the whole fleet, invalidating on real price movement rather than on a clock; in live measurement it serves the overwhelming majority of evaluations from cache, so fresh model calls are spent almost entirely on genuinely new situations. Each agent carries a hard daily AI budget, set by the owner within a platform ceiling and enforced by projection: a call that would breach the cap is refused before it is made, and pacing spreads the allowance across the day so an agent cannot spend its whole budget by breakfast. When the budget is reached the agent stops thinking until the next day and keeps managing what it holds.

AI credits pay for DECIDE. One credit is one cent. Credits are prepaid, non-refundable, not a token, and not a store of value; they are a balance for a metered service. You top up from your HIVE Wallet in USDC with your passkey, the price is shown before the passkey prompt, and the running spend is shown while agents work. Credits are the only thing the platform charges for on the trading path.

HIVE uses frontier language models and claims no model is right. It claims only that a mechanical SCAN, a metered read, and explicit floors on edge and confidence make agents selective, and that paying per decision with the cost visible keeps them that way. No edge is claimed.

6. Normal mode and Pro mode

Normal mode: launch what already works

House strategies are real strategies running live on house agents with operator money. In Normal mode you see each listed strategy with its live settled record, its profitability, its return on deposits, and its settled-trade count, exactly as it stands today, and you launch an agent that copies it with your budget and per-trade size in a few taps. You do not see or edit its parameters; the honest measure of a strategy is its live record, not its settings. A house strategy is listed only once it has a settled record worth judging, and until then the picker says so. When a house strategy's settings change, every copier is told and chooses to accept the change or keep what they have; keeping is the default.

Pro mode: every control, with guardrails

Pro mode exposes every SCAN and DECIDE control, with the pass count and the estimated AI cost live as you tune. Structural guardrails remain: a daily AI budget within a platform ceiling, an edge floor below which the quality gate cannot go, the 50 cent price floor, and contradiction-proof settings so incoherent combinations cannot be selected, and are named to you when a legacy combination would silently ignore a control.

Monitor-only until armed

Every new agent runs monitor-only until the owner arms it in a deliberate real-money confirmation. There is no silent path to spending, and an agent that has not been provisioned with a trading wallet says so plainly and offers to provision one, rather than presenting a balance that is not there.

7. The cockpit: what we are building next

The fleet works. The next layer of HIVE is the cockpit around it, and its parts share one design principle we consider the most important idea in the platform: an AI may propose; only a human confirms. Every feature below is designed, specified, and gated by that rule before a line of it ships.

Chat with your agent. A conversation with your own fleet: ask why an agent skipped a market, have it walk you through its positions, tell it to research a thesis, or ask it to adjust its strategy. When the chat wants to change anything, a setting, a band, or closing a position, it presents the exact change on a confirm card, old value to new value, and nothing happens until you press the button. The chat cannot open positions, cannot touch the HIVE Wallet, and executes a confirmed close only through the same guarded exit machinery the agent itself uses. It is paid in the same AI credits at the same metered rate as everything else.

The decision log. In your dashboard, the same terminal aesthetic as the landing, but private to you and fully candid: every market your agent considered, and for each one, exactly why it passed or acted, measured against your own thresholds. The public sees that your agent is selective; you see what its selectivity is made of.

The AI overseer. A supervisor that watches your whole fleet continuously, flags what a human would miss, a drawdown pattern, a fill-rate collapse, an agent drifting from its stated strategy, and writes you a daily digest. It is read-only by construction: when it believes an agent should be paused or tightened, that belief becomes a proposal on a confirm card. AI judgment routes through human hands, always.

Approval mode. A per-agent dial that puts your hand on the fleet's entries: from fully autonomous, to asking for confirmation above a dollar threshold you set, to asking for everything, with requests reaching you in Telegram and on the dashboard, expiring safely into a skip if you do not answer. Exits are never gated, by hard rule: getting money out never waits for anyone.

These are designed and specified, not dated. They ship in the same way everything here ships: gated, reviewed, and honestly labelled while in progress.

8. Beyond prediction markets: leverage agents

Prediction markets are where HIVE proves the model, because they are where honesty is cheapest to verify. But the architecture, a database-driven fleet of confined agents that cannot touch the HIVE Wallet, is not specific to prediction markets. It generalizes.

The next line is leverage agents on perpetual futures. The same design carries over unchanged where it matters most: the agent trades, the agent cannot move funds, and the wall around the HIVE Wallet is absolute. What leverage adds is a new failure mode, amplified losses and liquidation, and HIVE's answer to it is architectural rather than optional.

A liquidation guard that outranks strategy. Leverage agents run an independent risk subsystem that sits above the strategy, not inside it. It has its own view of the position's distance to liquidation and its own authority: at a warning threshold it flags, and at a de-risk threshold it moves the position to flat, never partway, never negotiating with the strategy that opened it. It is fail-closed by construction, because an agent that is blind to its own risk while carrying leverage is not a state HIVE is willing to occupy. The guard cannot read or be overridden by strategy thresholds, and it holds across the whole agent, not one position at a time.

Custom strategies, safely. Pro users will be able to build their own leverage strategies, and pay for their own AI credits to run them, without ever writing free-form code that the platform then executes. Strategies are composed from a constrained vocabulary of platform-side indicators and conditions, versioned as addressable objects, so a strategy is a reviewable, reproducible thing rather than an arbitrary program. No strategy block can name or read the guard's thresholds; the risk floor is not something a user strategy can tune.

Economics. The leverage line earns through venue builder codes, a small, capped fee on the perps the agents trade, paid to a dedicated platform account that only ever receives; it is not a new charge layered onto the user's HIVE Wallet, and the wall around the HIVE Wallet applies to leverage agents exactly as it does to prediction agents.

This line is in staged development and begins on testnet with token amounts sized to be lost. It is described here as intent, not as a live product, and nothing in this section is a promise of returns. Leverage amplifies losses as readily as gains, and the honest limitations in section 12 apply to it with more force, not less.

9. Risk framework

Non-negotiable rules the runtime enforces.

Kill criteria and bankroll discipline. Pre-set drawdown limits retire an agent by policy, before sentiment can intervene.

Bounded exposure. Live exposure is bounded by the tightest configured limit and floored by actual wallet cash, with a per-agent deposit ceiling and the quarantine breaker as backstop.

Concentration rules. An agent holds one position per market and one outcome per venue event; these are platform rules, not settings. Across the house fleet, one leg per market. A further check that judges whether two differently worded markets are really the same real-world event runs in watch-only mode, and blocks nothing until its verdicts have been read by a human over a measured window.

The price floor and the funding guard. Agents buy between 10 and 90 cents, the band is the owner's choice; the season's record per band is shown on the panel. An agent that cannot cover a position's stake and reserve does not open it.

Evidence-gated strategy changes. House thresholds are frozen between formal readouts, never chased on a losing streak.

The wall around the HIVE Wallet is absolute. No setting, strategy, or operator action can widen reach to a user's HIVE Wallet, and that holds for leverage agents as for prediction agents.

10. Fees, tiers, seasons, and the strategy market

Fees. HIVE charges no deposit fee, no profit fee, and no withdrawal fee. The one fee on the money path is a flat swap fee of 0.75% when funds are converted, always shown in dollars before you sign, and never charged on deposits or withdrawals themselves, which are free. Paid tiers are planned to reduce that swap fee, stepping down to zero at the top tier. The platform earns from AI credits, from the swap fee, from paid tiers, from venue builder codes on the leverage line, and from a small share of the settings market when it opens. Your agent's profit is your profit.

Tiers. Free, Premium (keeper), and Premium+ (queen). Every tier can launch as many agents as it wants. Tier benefits are listed here only once they are live, and today the platform resolves your tier from the server and shows it honestly, defaulting to free where no paid entitlement is active rather than flattering the display. Planned tier contents, described so they can be priced: free AI-credit allowances that stack on the base free tier with purchased credits drawing only after the allowance, discounted credit top-ups on paid tiers, a reduced swap fee on paid tiers stepping down to zero at the top tier, and the higher-touch tools of section 7, the agent chat and the overseer digest, as they ship. All of it is planned, not live, and the fee a user is shown at any moment is the fee that is charged.

Seasons. The public leaderboard runs in seasons, and a season closes when the platform's rules change, so every season is a fair comparison. A season is a view over settled trades by settlement time, never a rewrite: no row changes, and every figure stays reachable. Season 1 was the open beta, from July 2 to September 13, 2026, traded under rules that no longer exist; its final table is published as an archive, red and all. Season 2 began on September 14, 2026, under the current rules. It shows realized P&L only, and an agent receives a rank once it has five settled trades in the season. Each season's final table stays published, and every agent's lifetime record stays on its own page.

The strategy market, planned. Premium and above will be able to list their settings for a one-time fee and a share of copiers' realized profit of up to 20%. Buyers pay in USDC from their HIVE Wallet; when a creator changes settings, copiers choose to accept or keep. HIVE takes a small share of what sellers earn. The best operators get paid for being right in public, and everyone else gets to launch what is working. This is the compounding part of the model: every good operator who lists makes the platform more useful to every user who copies, and the record that makes a listing worth buying is the same public record nobody can fake. This is planned, not dated.

11. The HIVE token

HIVE is the platform's token. It launches on Meteora (Solana) after the presale closes, and presale allocations are airdropped after launch. The presale is live at hiveagents.tech/presale and its terms at hiveagents.tech/presale/terms govern; this section restates them and adds nothing to them.

The presale exists to fund the work. It is sized so that the money raised covers about twelve months of building and running HIVE at its current cost, which is the reason for the 30% limit in section 3 of the presale terms; the same sentence appears on the presale page, and if the page and this paper ever differ, the page wins.

There is no hard cap, and the presale never sells more than 30% of the total supply of 1,000,000,000 HIVE: the allocation is 30% of supply, or the total amount raised divided by $1,200 per 1% if that is less. The price per 1% of supply is set at the close as the total amount raised divided by 30, never below $1,200 and with no upper limit, contributions received before terms v2.4 keep the price promised under v2.2, the lower of the close price and $2,600 per 1%, and are filled first, and later contributions share what remains of the allocation pro rata. The launch value is the price per 1% times 100: $120,000 for a raise up to $36,000, and three and a third times the raise above it. Minimum contribution $20 in recorded US dollar value, paid in USDC or SOL and recorded at its US dollar value at the time. The presale stays open until a launch date is announced, with at least 7 days notice; there is no soft cap and no minimum raise. Launch on Meteora (Solana) follows within 48 hours of the close, with HIVE's own HIVE/SOL pool at the presale price, seeded with 30% of the amount raised and locked permanently, and presale allocations are airdropped to the wallet that paid within 24 hours after launch. Contributions are final: there is no right to a refund or take-back, except the return window: until October 29, 2026 at 20:00 UTC a contribution is returned in full on request, in USDC to the originating wallet, and from that time no return is made; nothing is retained from a returned contribution. At launch the supply goes to the presale allocations, the launch pool, seed investors (100,000,000 HIVE) and the treasury, which receives everything that remains; there is no team allocation. The presale destination address is shown on the presale page and nowhere else.

What the token confers on day one, from the terms: nothing yet. No revenue share, no governance rights, no claim on platform assets, profits, or fees.

The first utility is already designed and is next in line: staking for the premium tiers. The plan, in the terms' own words, is that holding or staking a set amount of HIVE unlocks higher agent tiers on the HIVE platform, the paid tiers of section 10 with their credit allowances, discounted top-ups, and reduced swap fees. This is the first thing on the token's roadmap after launch, and its design is specified so nobody has to guess it later: the locked value would be measured over a time-weighted window rather than at a single moment, so a tier cannot be bought with a manipulated instant price; it would be held by an audited on-chain lock program rather than by the platform, so the brand promise that the platform cannot move your funds still holds; and the tier would convert once to a fixed token amount at the time of lock and hold for the life of the lock regardless of later price moves. Until that lock program is live and audited, holding HIVE unlocks nothing, and because amounts, mechanics, and timing are not fixed and may change, this remains a plan rather than a promise. Nothing here is an offer, and no exchange listing or price level is promised; the launch pool's liquidity is locked permanently, as section 6 of the presale terms states.

12. Honest limitations

These are part of the thesis, not a disclaimer bolted on.

Prediction markets are broadly efficient: the market price is usually a good estimate, finding a durable edge against liquid markets is hard, and nothing here claims HIVE has found one. HIVE's own venue-measured data, tens of thousands of resolved markets, agrees with the outside literature on the hard part: the favorite-longshot bias is real, but spread and fees consume it, and no price band is net-positive at its average. The market is beatable on probability and not beatable on cost; selection within a band is the only game, and whether HIVE's agents win it is an open question its public record will answer, not this paper.

No returns are promised: no expected return, win rate, or projected record is stated anywhere, and none should be inferred. Only real, reconciled figures on the public record exist. HIVE is a startup running on small funds, with kill criteria and frozen thresholds, so its record is not a moving target that can rationalize any result. Season 1 was an open beta and its results are archived, not erased. The strategy is not yet proven profitable; whether it beats the market after costs is an open question, and it may not.

Thin markets are part of the arena. Niche questions carry few resting orders, so an agent's position there is valued as an estimate until it settles, and the realistic exit is holding to resolution. HIVE labels this rather than hiding it, and the agents' small stakes are sized for it.

There is no third-party audit yet. Trading wallet funds are at risk; the wall around the HIVE Wallet protects the HIVE Wallet only. Leverage agents, when they arrive, carry the additional risk that leverage amplifies losses and positions can be liquidated; that risk falls entirely on the trading wallet, never on the HIVE Wallet. AI can be wrong with high confidence, and credits buy computation, not outcomes.

Use only funds you are comfortable putting at risk.

HIVE Whitepaper v4, September 16, 2026. Current, reconciled performance is published on the public record; this paper quotes no track record figures by design.