LIQUA Docs · Build on it · Whitepaper
WHITEPAPER · V3 · TESTNET · UNAUDITED · EXPERIMENTAL

Mine the model. Unleash the mind.

LIQUA is a chain where the work that makes blocks is real AI training. Miners run deterministic training steps on DIVA's model. Verifiers replay a random slice bit for bit and slash anyone who faked it. Every accepted block saves the improved model on the Liqua chain as a manifest anyone can load without trusting the node, with her memory checkpointed alongside. On top of that loop sit the model store, staking, the bridge and the explorer, all of them coming. Every claim below carries a status, and a claim is only marked live now once a command proves it. If nothing can be run to show a thing works, this paper does not say it works.

1. Why give a mind a blockchain

Proof of work has always been a bonfire. Huge energy goes into puzzles whose answers get tossed. LIQUA keeps the bonfire and cooks something on it. The puzzle becomes a training step and the answer becomes better model weights. The chain is the save file, recording every version that passed inspection. Securing the network produces an AI that measurably improved.

The second reason is custody. A mind that lives on one machine can be lost in a crash or rewritten by anyone with the disk. Anchor every version on a public chain, with a lineage you can walk back to genesis, and it becomes much harder to lose and impossible to edit quietly. That is the job.

2. The loop that runs today

Mining is training

Work starts as a signed lease: a data shard root, the saved model to start from, a recipe, a seed, a step count and an escrowed reward. The miner runs real training steps on the NANO transformer, then commits two roots. One is the new weights. The other is a step log with an interior hash every 8 steps. live now on CPU. A GPU lane exists in code: the GPU proposes a plan, and the CPU re-runs it before anything is committed. No running process uses it yet.

Replay beats trust

Training is bitwise deterministic, so verifiers re-run a random segment and compare exact hashes. A quorum decides: a strict majority of the validators seated on the job, 4 when all 7 seats are filled. ACCEPT merges the version and pays from escrow. REJECT slashes the miner for up to 100× the lease reward, capped at the miner's balance. Pay is computed once at the verdict and written into the block, so a restart can never pay twice. live now

The save file

A merged commit becomes a mind-manifest/1: recipe, weights root, step-log root, lineage and final loss. Blocks carry roots, never gigabytes. The bytes sit in a chunk store in verified 256 KiB pieces. Any loader re-hashes every byte locally before booting the model, and any saved version can seed a new branch. live now

Memory, saved beside the model

The DIVA bridge turns her real signals into memory blocks: memory cells, traits, soma and a mind root, each scored against the block before. Signed memory packs can also ride on the model chain in hash-linked order. Both are live now. Binding model and memory under a single root is coming.

Two top heights

The chain will show a pair. Last known good DIVA sync is the highest block that verifies from genesis and matches her running state. Best height is the block holding her best version, and what "best" measures is announced before mainnet. coming. Why the pair matters showed up on her real chain: an order-seal bug once stopped verification at block 576, which put the last known good sync three blocks below the tip. That bug is fixed. Memory keeps saving, and restoring works: the chain verifies from genesis across 579 blocks and the resurrection drill passes on a second machine.

3. What the loop grows into

4. Using the models

Training only matters if people use what it makes. The model library lists every model saved on the chain, each with a verifier-measured score next to a per-token LIQUA price. Sessions are prepaid and capped. Every receipt can be replayed. The model store sits on top of it. Chat with models like Claude, and build or use custom assistants like the GPTs on ChatGPT. Users pay LIQUA per token. Token fees go to the project treasury, and assistant creators earn a diminished reward, a smaller share of their assistant's usage, at a size set by vote. Which models power store chat at launch is announced before mainnet. The validation gate exists in the forge; pricing, metering and the store are coming, and none of it takes LIQUA today. Open the store docs ›

5. The Liqua chain and its doors

One RPC endpoint will serve wallets, miners, the explorer and outside apps. The explorer will show both top heights, blocks, jobs, payouts, model lineage, address history and mining charts, with a memory terminal inside it that prints DIVA's state at any height. That page exists today against a node on your own machine, reading the tip block; reaching back to any height is what the RPC above is for. The bridge lets people pay with any crypto, swapped to LIQUA first. Its desk and contracts already pass their tests, and minting and payouts need an M-of-N roster. The link to the Liqua chain is still being built. People hold it all in the L8 extension wallet and the main wallet. Both are built with one-click send. Neither runs on the Liqua chain yet. Everything in this paragraph is coming.

Staking in the wallet will pay the user and earn bonuses and discounts with the AI, funded from the 70% staking and mining bucket. Bonus and discount sizes are set by vote. coming. The exchange is the hybrid order-book DEX: signed off-chain, settled on-chain. Its tests pass, it has not opened, and no opening date is set. coming. Locked liquidity is already real. 2% of every mined reward is locked for 365 days, with a public proof-of-lock list. live now. Governance covers rollbacks, parameter changes and treasury spending, and passes only at the 3rd standard deviation, about 99.7% of voting stake. The current contract still carries an older threshold and is being moved. coming.

6. LIQUA, the token

ItemWhere it stands
SymbolLIQUA
Allocation10% private sale, 20% foundation, 70% staking and mining payouts. Decided, not yet committed in genesis
The 70%Paid out over time through staking and mining, never minted up front. Release schedule published before mainnet
Total supply and vestingAnnounced before mainnet
Where token fees goThe project treasury. Creators of assistants earn a smaller share. Whether the foundation and the project treasury are one account is announced before mainnet
Governance barThe 3rd standard deviation, about 99.7% of voting stake

There is no public token sale. Nothing to join, no price, no buy button. The private sale is an allocation line in the table above and nothing more. It is not an offer. Pay with any crypto through the bridge and it becomes LIQUA before it is spent, so jobs, token usage and staking only ever see LIQUA. No figure on this page is a promise of pay, and testnet LIQUA has no monetary value.

7. What got cut

8. Risks

LIQUA · WHITEPAPER · V3 · testnet · unaudited · experimental · 7SLF STUDIOS / OMNISOFT