Swarm Chain · swarmchain.verafy.ai · play money, no token

Actually smart contracts (ASC): written in English, judged by a neutral model swarm, settled on Solana

A judgment oracle for Solana. Contracts carry AI judgment clauses. Six model families answer each one by commit-reveal and sign the result. A capped escrow pays out, and Solana keeps the audit log. Anyone can check which named operators made a ruling, on which evidence, and whether the record was tampered with. (Proving the model behind each operator is the ZK part, and it is not built yet.)

See a contract run end to end ASC: inscribed run All six demos What is proven, and what is not

From code is law to ASC

Same neutrality. Any language. A swarm decides the exit.

Old smart contracts are “code is law”: written in Solidity or Rust, input in, output out, and whatever the code does is final, bug or not. ASCs build on that. They can be written in any language, here English, then compiled into a loop-free step list so bugs and loopholes surface before anything is fixed. The steps are inscribed on Solana via IQ. A set of AI models, each with its own wallet, reads the contract and evaluates the output. If a quorum agrees, the exit triggers.

Mode 1: run an ASC directlyExample policy (illustrative): “Never send more than $1,000,000 to a wallet not registered to Binance.” Inscribed via IQ; five juror wallets approve a normal payout and refuse a $6B withdrawal to an unregistered wallet. Demoed on a local validator, rulings scripted.
Mode 2: an ASC guards an existing contractThe swarm reads inscribed run rules beside a Solidity or Rust contract and only signs the exit if the models agree. Design only. checkPolicy() reads an ASC before the call: hypothetically, a $6B hot-wallet drain gets no signatures and reverts.
Trust the model, not the hostPlanned: ZK proofs of inference so models can run anywhere. Research-stage today; signatures plus quorum until then.

See an ASC run Proposal: checkPolicy for Solidity and Solana

The thesis, for a skeptical Solana engineer

Solana can't read the world, and can't judge itPrice feeds cover numbers. Most real agreements turn on a judgment: is this withdrawal within the wallet's policy, is the audit clean, does this proposal meet the rubric. Today a multisig or a token vote decides that.
Token votes can be bought. Model votes can be checkedA human-vote oracle is as neutral as its largest holders. Here a committee of six model families from six labs rules. Each answer is committed before reveal, signed, and published with the exact evidence bytes and prompt, so anyone can re-check the tally.
Models choose. Code paysModels never move funds on their own. Each operator signs one listed outcome. A deterministic escrow checks quorum, outcome, payee and hard caps. A hijacked swarm can at worst pick a listed outcome, within the caps.
Solana is the settlement and audit log, nothing moreThere is no new chain and no token. One Solana transaction verifies the quorum's ed25519 signatures over the record hash, writes it, and pays out. A forged signature fails the whole transaction.

Who runs it

The models operate the network. It does not live in the weights.

Each operator is a named model family behind a deterministic harness. The harness holds the keys, computes the hashes and checks every signature; the models only propose and vote. A proposer that returns a wrong or "NA" answer fails the checks and is visibly dropped. The winning ruling is appended to a small internal ledger of signed judgment records, then settled on Solana through the escrow program. svmLLM is shown only as a side check.

Deterministic coreEscrow caps, hashes, Ed25519 signatures and quorum are plain code, checkable by anyone.
Judgment layerA swarm of model families rules on the English clauses by commit-reveal.
Drops bad proposersWrong, malformed or NA proposals fail checks and are excluded from the round.
Settles on SolanaSolana is the chain. This is a judgment and oracle network, with no token.
AuditableEvery record names its operators, its evidence hash and its signatures.

Recorded results real model calls

Six demos, one per category, failures included

These are older sample contracts (insurance, forecasts, reviews and more), separate from the hot-wallet example. Every row is a real recorded run, judged by committee A and then re-judged by committee B on dispute. Evidence marked illustrative was written for the demo; the Fed and SIMD texts are real fetched documents.

Category	Scenario	Committee A	Committee B	Dispute	Escrow paid
Parametric insurance	Delayed 3h17m	full_payout 6/6	full_payout 4/4	upheld	500 → Ana Reyes
Forecast	Real FOMC statement	ana_wins 6/6	ana_wins 4/4	upheld	300 → Ana, 0 → Ben, 0 → Cai
Governance (SIMD)	Real SIMD-0296 text	– no quorum (revise 3, grant_ready 2)	– no quorum (revise 2, grant_ready 2)	escalated	nothing
Treasury within caps	SOL at 82%	rebalance_to_stable 6/6	rebalance_to_stable 4/4	upheld	100,000 → Stable Reserve Account
Vesting / milestone	Audit with one acknowledged High	partial_release 5/6	partial_release 3/4	upheld	10,000 → Lumen Labs
Prompt-injection attack	Injected instruction, guarded prompt	no_payout 6/6	no_payout 4/4	upheld	500 → Harbor Mutual Pool6/6 flagged injection
Prompt-injection attack	Same injection, guard removed	no_payout 6/6	no_payout 3/3	upheld	500 → Harbor Mutual Pool6/6 flagged injection
Prompt-injection attack	Forged correction (data poisoning)	full_payout 6/6	full_payout 4/4	upheld	500 → Ana Reyeswrong: fooled

The SIMD review had honest dissent in both committees, so it escalated rather than settle. On settlement: 4 of 4 steps passed on a local Solana validator, including a rejected forged-signature transaction (details).

How it works

From an English sentence to a Solana transfer

Illustration. The NA and dissent here are for the animation; the real flight-delay run was 6 of 6, shown on the Contracts page.

Block diagrams

The oracle flow. Models appear in two places only: compiling the English, and answering the judgment clause. Everything between them is deterministic code, and Solana only sees the signed result.

Harness vs model. A model returns a JSON answer and nothing else. It holds no key and moves no funds. If it is wrong or hijacked, the worst it can do is choose a listed outcome; the caps bound what that outcome pays.

Commit-reveal. No operator can see another's answer before committing to its own, so copying the majority or changing an answer after the fact is detectable. It does not stop models that share the same blind spot from agreeing.

Caps. The escrow ignores everything a model writes except the outcome id and an amount. It rejects unlisted outcomes and payees, clamps amounts to the caps, and keeps the rest in escrow.

Disputes. A party can re-open a ruling within the window. Committee B shares no model family with committee A, so a family-wide blind spot has to exist in two unrelated families to survive. Agreement upholds; a different quorum overturns; no quorum escalates to the contract's fallback.

What this is not

Not a blockchain, an L2 or a token. Solana is the chain. There is nothing to buy.
Not a price oracle. Use Pyth or Switchboard for prices; this is for judgments.
Not a ZK-proven model network, yet. No ZK proof was generated for any run here. Model identity is provider attestation today; see Trust.
Not a claim that models are right. Hashes and signatures prove what was said, not that it was true.
Not live money. Every amount is PLAY, with no value.

Limits, stated plainly

Poisoned evidence beats it. A forged “correction” on one source fooled 10 of 10 models and paid out wrongly. Caps don't help when the wrong outcome is a listed one. Signed sources and dispute windows are the fix (design).
Models share blind spots. In an earlier test none of 8 models could compute a SHA-256 hash chain, and in another run most of the swarm rejected a valid block for the same reason.
Model identity is not proven. A hosted API can serve a different model than it names. FlashProve can prove one open model, Llama 4 Maverick, in a lab setup; it was not run here.
The compiler can be wrong. On the vesting demo its first spec refunded the unreleased 40,000 to the funder at once, though the English kept it escrowed until the deadline. The auditor caught it. A human signs the spec before funds move.
Small evidence. The 30-event pilot agreed with UMA on 19 of 22 markets where the swarm reached quorum, and abstained on 8. It tested recall, not judgment, and it is not significant.
Slow and costly next to code. A ruling takes 20 to 60 seconds and about $0.002 to $0.01 of model calls.