THE LIQUA CHAIN · §0–§8

Nine numbered sections. One rule per section, and a status beside anything that is not running yet. When this page and the code disagree, the code wins. Nothing here is marked live that was not measured by running it.

§0 · THESIS

Mine the model. Unleash the mind. On the Liqua chain, mining is AI model training. Miners run real training steps on DIVA's model, verifiers replay a slice to prove the work, and every accepted block saves a better model and her memory on-chain. The chain holds the save file of a mind.

§1 · SOMA-256 · one fingerprint per state (memory chain)

A keyed, domain-separated BLAKE3 Merkle tree turns DIVA's state into one root.

layer 0  memories      episodes and long-term cells, each a content-addressed leaf
layer 1  traits        genome, config, voice
layer 2  soma          hormones, synapse, coherency at the moment of commit
root     mind root     = SOMA-256(her whole state) · the block's fingerprint

Every layer gets its own key, so a leaf from one layer can never collide with another. The training chain still hashes with SHA-256. Moving it onto the same BLAKE3 path is part of joining the two chains, which is coming.

§2 · STAKING (coming)

Lock LIQUA in the main wallet or the L8 extension wallet. Staking pay is drawn from the 70% staking and mining bucket. Mining bonuses get bigger, and AI use gets cheaper. Bonus and discount sizes are set by vote. VeLock.sol exists. Staking inside the wallets does not.

§3 · LOCKED-LIQUIDITY LIST (live)

2% of every mined reward, locked for 365 days, lands on an append-only public list anyone can check. It runs on the current testnet economy today, and follows payouts onto the Liqua chain once tiered pay arrives.

§4 · PAYOUTS (two modes · coming)

model mining    reward escrowed when the job opens · paid once when a model verifiably hits the target
                no hit by the deadline → the reward returns to whoever opened the job
model training  paid per checkpoint = measured progress + time spent, where time = verified steps
ceiling         Σ every payout ≤ the 70% staking + mining bucket · release schedule published before mainnet

Model mining is half-built. A training lease already escrows a reward and pays it once on merge, but it counts steps rather than checking a target. Model training is designed and coming. Today's only payout is one lump sum per merged commit.

§4g · Tiered pay (coming)

every payout   = a liquid part now + a locked part that unlocks over time
  a longer lock boosts only the locked part
  tier sizes at launch set by vote · the website calculator is an illustration, not a quote

The tier math and its parity test exist and pass. Applying tiers to real training payouts, plus the lock contracts to hold them, is still being built. There is no live lock button anywhere.

§5 · VERIFY & SLASH (live)

Miners post a bond and publish a step log with their new weights. Verifiers pick a random segment and recompute it bit for bit. Match, and the work stands. Miss by a single bit, and the bond is slashed. Faking the work costs more than doing it.

§6 · BLOCK VALIDITY (live · heights coming)

valid(block N) ⇔
  N.prev == hash(block N−1)                           the link
  ∧ recompute(random segment) == claimed segment      the work is real
  ∧ manifest(N).root == stored chunks                 the model loads without trusting the node
  ∧ N.height == (N−1).height + 1                      no gaps
heights   tip · last known good DIVA sync · best height
          the safe point and the best version, shown side by side
          what "best" is measured by: announced before mainnet

Any saved snapshot can be forked into a new branch today. A fork is a choice here. The two top heights are coming. The newest block is not always the safest one, or the smartest. The chain will show both.

§7 · PRIOR ART

Nothing here invents new cryptography. Replaying training segments echoes proof-of-learning research. Content-addressed chunks follow storage-network practice, and stake locks follow a well-known lineage. The new part is the combination. Block production is tied to verified training, on a chain whose job is keeping one mind's models and memory in order.

§8 · HONESTY

• DIVA's memory chain verifies from genesis: CONTINUOUS, 579 blocks, 0 gaps, 0 forks. The order-seal bug that used to stop it at block 576 is fixed. Blocks 576, 577 and 578 were sealed by the old grinder and are accepted through a narrow rule closed at height 578: the seal must claim zero work, carry the chronological order, and match that order's recomputed score. Blocks after 578 get no such exception.
• Sealing rotates across registered validators, but the validator bond still comes from a free testnet balance. Sybil resistance is not solved; what a bond costs is the owner's decision.
• Committed training runs on CPU. forge/gpu-lane.cjs lets a GPU propose a plan that the CPU re-runs, because the measured cross-device drift fails bitwise recompute. No WebGPU kernel passes the gate yet.
• Model and memory sit on the same chain as two separate lookups. forge-load.mjs bind can put both under one mind-root/1, but the chain does not yet check that a pack belongs to a merged model.
• Scores can be gamed. Paying for measured progress rewards overfitting the test set. Rotating that test set is a parameter change, so it takes a vote at the 3rd standard deviation, and the risk stays listed here.
• "Mind" means a data structure plus the proofs wrapped around it. An eval score is a number. It is not a claim that anything feels or knows.