A protocol to make AI verified and verifiable.

The receipt is a compact summary of how the model arrived at its answer. Any other operator can replay it against the public copy of the model and decide whether the answer was honest. A naive 're-run it and compare every digit' check doesn't work in AI: the same model on different graphics cards produces slightly different numbers under the hood, even when the answer is genuine (why?)The same AI model on different graphics cards produces tiny numerical differences (attention kernels, batch sizes, logits) — harmless, but it rules out simple bit-for-bit checks.. So instead we use a statistical testA pass/fail test calibrated so honest providers pass with high probability and cheaters get caught — used because exact replay isn't possible in AI., tuned so honest providers pass and cheaters get caught. The test runs in three escalating tiers — a fast check, a medium-depth check, and a full audit (Quick / Smell / Full)Three escalating layers of verification: Quick (under a second), Smell (medium-depth sampling), and Full (a complete audit). Each is stricter than the last. — and each tier is stricter than the last. The checking software is open-source: anyone can run it, and anyone who spots a dishonest block can challenge it.

Serving and mining are one job, not two. We run a customised version of the standard open-source AI serving enginesForked versions of vLLM (the standard GPU serving engine) and llama.cpp (a CPU / Apple-Silicon engine) — modified to record the proof transcript while answering. — capture of the proof happens inside the same step that produces the answer, so users see no slowdown. Idle GPU time is filled with internal practice prompts that step aside the moment a paying user shows up. If an answer happens to clear the network's difficulty barThe rarity threshold a proof must beat to be accepted as a new block — automatically adjusted to keep block time steady., it becomes a candidate block; if it doesn't, you've still served a paying customer. Either way the electricity isn't wasted, and there is no separate mining hardware to buy.

TensorCash is forked from Bitcoin Core. We keep how Bitcoin tracks coinsBitcoin's accounting model — instead of bank-style balances, coins exist as discrete 'unspent outputs' from prior transactions., how it signs transactionsBitcoin's existing way of signing and verifying transactions with cryptographic keys., and how blocks are passed between nodes — then add a small number of new rules, all narrow on purpose. Each block carries a compact AI-work proofProof-of-inference — a compact, replayable receipt attached to every AI answer, so others can verify that a real model genuinely produced it. and a difficulty targetThe rarity threshold a proof must beat to be accepted as a new block — automatically adjusted to keep block time steady., plus a cryptographic clockVerifiable Delay Function — a cryptographic clock based on the Wesolowski construction, proving that real time has elapsed even with unlimited parallel hardware. that proves real time elapsed during mining. Proofs are checked in three escalating layers (Quick / Smell / Full)Three escalating layers of verification: Quick (under a second), Smell (medium-depth sampling), and Full (a complete audit). Each is stricter than the last.: two fast ones that gate how blocks travel between nodes, and one full audit that runs in the background. Before accepting a competing chain, nodes also weigh how much real time was honestly spent on it (proof-of-time)A measure of the wall-clock effort behind a chain (via VDFs), used alongside proof-of-work to score competing chains., not just raw compute — and any deep rewrite of recent history triggers a forensic warning before the chain switches over. Crucially, we did not bolt on a general-purpose programming layer (no smart-contract VM)Unlike Ethereum and most modern chains, TensorCash has no general-purpose smart-contract programming language — the attack surface that has caused most DeFi hacks does not exist here. — the kind that has caused most of the high-profile DeFi hacks of the last decade. The chain only does what it needs to do, and nothing more.

Issued assets live in the chain the same way TSC does — not as IOUs running inside a smart contract. The issuer fills in an asset recordThe on-chain record where an asset's parameters live — ticker, supply cap, transfer rules, legal terms, governance rules, and bond. that fixes the rules: ticker symbol, decimal places, maximum supply, who can transfer it, who needs to pass identity checks, the legal terms, the voting rules, and a refundable security depositA refundable native-coin deposit (Issuance Control Unit bond) that an issuer locks when creating an asset — released once the asset has accrued enough activity to prove it isn't spam.. There is no listing fee, but the deposit stays locked until the asset has paid enough miner fees to prove it's a genuine asset, not spam. Once published, the rules are enforced by every node on the network: the chain itself refuses to issue more than the cap, refuses to reuse or rename a ticker, refuses transfers to ineligible recipients, and refuses silent edits to the legal terms.

Issuers can publish the legal documents that belong to an asset — prospectus, term sheet, governance records, holder-only disclosures — and pin them directly to the asset on-chain. Public documents are readable by anyone. Holder-only documents are encrypted; only the wallets of actual holders receive the key to unlock them. Either way, the chain records a permanent fingerprintA short, fixed-length cryptographic fingerprint of a document — any change to the document produces a completely different fingerprint. of the document, with a timestamp. If anyone tries to swap in a different version later, holders notice immediately. EU-regulated digital signaturesQualified Electronic Signature — the highest tier of EU-regulated digital signature under the eIDAS framework, legally equivalent to a handwritten signature., notarised records, or PGP signatures can all point to that same fingerprint — the chain doesn't need to understand the legal artefact itself, only to vouch for what was signed and when.

Three of the most common trade structures in finance — buy now (spot), lend collateral against cash (repo), and agree today to deliver later (forward) — all share one rule the chain enforces: the spending transaction must contain exactly this output, of this amount, in this assetOP_OUTPUTMATCH — a new Tapscript opcode introduced by TensorCash. It enforces a single rule: the spending transaction must contain an output of exactly this amount, in exactly this asset, locked to exactly this script.. That single rule is enough to express every variant of these contracts, without the chain needing a general-purpose programming language — and without the security holes that come with one. The check looks only at the current transaction: no rummaging through history, no mutable contract state, nothing to exploit. Trade terms stay private until the moment a leg of the trade is executedA 2021 Bitcoin upgrade that keeps the detail of a contract hidden until the moment it's executed — onlookers only see a generic locked output. — the public chain sees a generic locked output, not the term sheet behind it.

The wallet handles the full pre- and post-trade workflow in one place. Three things happen behind the scenes: first, a decentralised messaging networkA simple, censorship-resistant messaging protocol used here for advertising open trade offers without going through a central venue. lets you browse open offers without going through any single venue; then, once you find a counterparty, your wallets open a private, encrypted channelTwo well-studied cryptographic protocols (Noise and SPAKE2) that together open a private, password-authenticated channel between two wallets. only the two of you can read; finally, a cryptographic ceremonyA trade that either completes for both parties or for neither — never half-way. Built here using Hash Time-Locked Contracts and adaptor signatures. guarantees the trade is atomic — either both sides receive what they expected, or neither side moves any funds. One wallet page covers all six steps of the trade lifecycle — discovery, offers, negotiation, governance, discussion, and cross-chain swaps. For trades that span TensorCash and Ethereum, a companion smart contractTensorSwap — a small Solidity contract on the Ethereum Virtual Machine that handles the Ethereum side of a TensorCash↔Ethereum cross-chain trade. handles the Ethereum side.

From the very first block, TensorCash supports two parallel signature schemes side by side: the same proven ones Bitcoin uses todayThe two signature schemes Bitcoin already uses — ECDSA (legacy) and Schnorr (modern). Both supported in TensorCash from day one., and a new quantum-resistant oneML-DSA (NIST FIPS 204) — a signature scheme designed to remain secure even against a future quantum computer. Standardised by NIST in 2024. standardised by the U.S. National Institute of Standards and Technology in 2024. Users pick which scheme they want at the moment they create an address — and both schemes stay valid forever. Because TensorCash launched with both from the start (a genesis fork)A blockchain that starts as its own genesis (block zero) rather than splitting off from an existing chain — so new rules can ship from the first block, with no upgrade vote needed., there is no awkward upgrade later, no waiting for the network to vote in a new rule, no need to migrate old funds.

Some parts of trading — fast order matching, leverage, deep liquidity — can't be done on a blockchain at the speed serious traders need. TensorCash specifies a separate, faster exchange layer that sits above the chain. What makes it different is that no single server ever sees the whole order book: it's split cryptographically across ten independent validatorsA cryptographic technique that splits a secret across N parties so that any K of them can reconstruct it, but fewer than K learn nothing. Here, 6-of-10 across independent validators., and it takes six of them, working together, to act on any piece of it. No single party can peek at the order flow before a trade lands, and no single party can halt trading. Once a trade is matched, it settles back on the base chain using the same atomic-swap rules that any two-party deal uses.

Mining difficulty rises and falls with how much AI compute the network is running, so it works as a live index of AI-compute demand. Two parties post margin into separate private vaultsA 2021 Bitcoin upgrade that keeps the detail of a contract hidden until the moment it's executed — onlookers only see a generic locked output. and the chain settles them against that difficulty at a fixed future blockThe contract reads difficulty at one committed, already-buried block — not the block that settles it — so neither side can time settlement around a difficulty change. Once that block is final, the payoff is fixed. — manipulation-resistant, because the payoff is locked once that block is buried. The simplest form is a capped contract-for-differenceA contract-for-difference whose loss is capped: each side posts collateral up front and can lose at most that amount, with no margin calls and no liquidations.: go long or short the compute cycle, and the most you can ever lose is your own margin. A small day-one premium turns it into an optionPaying a one-off premium turns the contract into a covered call or put on AI compute — the buyer's loss is limited to the premium, and the writer earns it for posting collateral. — a covered call or put on compute. Miners and AI providers use it to hedge revenue when difficulty climbs; outside hedgers get exposure to AI-compute trends without running a single GPU. Like everything else here it is enforced natively by consensus (no smart-contract VM)Unlike Ethereum and most modern chains, TensorCash has no general-purpose smart-contract programming language — the attack surface that has caused most DeFi hacks does not exist here., and a cooperative settlement looks just like an ordinary payment.