AI providers mint blocks while answering prompts.

The receipt is a proof object — a compact transcript of the model's sampling trajectory on a hash-derived prompt — that any other node can replay against the published weights. Verification can't simply re-run the forward pass and compare bit-for-bit: the same model on different GPUs, batch sizes, or attention kernels produces slightly different logits. Instead it's a calibrated statistical decision rule, run on a three-tier ladder (Quick → Smell → Full) where each tier strictly tightens what counts as honest. The verifier ships as open-source software — anyone can run one and challenge dishonest blocks.

Mining and serving collapse into one workload. The Mining API is a forked vLLM (CUDA) and llama.cpp (CPU / Apple Silicon) with proof capture wired into the sampling path itself, so generating proof material doesn't slow user requests. Synthetic backfill prompts only run when the GPU would otherwise be idle and yield to paying traffic. When a transcript clears the difficulty target it becomes a block candidate; when it doesn't, you've still answered a real prompt. No separate mining loop, no idle electricity bill.

TensorCash inherits Bitcoin Core's transaction graph, UTXO model, and signature machinery, and extends the consensus surface in a small number of well-bounded ways. Block headers commit to an inference proof blob and an adjusted-difficulty target. Every block embeds a Wesolowski VDF proof binding the block to wall-clock work. Proof validation is split into Quick / Smell / Full rungs — Quick and Smell gate relay; Full runs outside the propagation path. The presync layer scores chains by proof-of-time before proof-of-work, and deep reorgs surface a forensic advisory before the chain switches. No general-purpose VM means no general-purpose VM attack surface to inherit.

Native fungible assets sit beside the chain's coin in the same UTXO graph. Every asset-bearing output carries a single typed binary record committed to by the transaction's sighash, so asset state cannot be rebound after a signature is produced. Each asset has an Issuer Control Unit — a UTXO-shaped credential the issuer must rotate on every authorised spend, bonded by a posted native-coin stake. Regulated assets carry zero-knowledge holder-eligibility proofs (KYC) using on-chain Groth16 verification keys and rolling compliance roots. Composable with Bitcoin's existing script families, Taproot covenants, and PSBT tooling.

Issuers can publish prospectuses, term sheets, governance records, and holder disclosures as on-chain commitments. Documents are anchored by hash; the asset's compliance root binds them to consensus. QES-grade electronic signature evidence (the EU Qualified Electronic Signature standard) folds into the same anchor. The on-chain commitment is permanent and timestamped; the document itself can be public, holder-only, or shared off-chain — with the chain proving the document seen by holders is the document the issuer published.

Spot, repo, and forward primitives settle through a single financial covenant: `OP_OUTPUTMATCH`, a Tapscript check that asks whether the spending transaction contains an output with a specified script hash, amount, and (for asset outputs) asset identifier. It performs bounded matching on the current transaction — no recipient script execution, no transaction-history lookups, no mutable state. Taproot keeps contract terms confidential until the script path is taken: the public UTXO set sees the vault, not the term sheet.

A three-layer coordination stack lives behind the wallet. Nostr for offer discovery; Noise+SPAKE2 for end-to-end-encrypted bilateral negotiation sessions; HTLCs and the Fair-Sign adaptor-signature ceremony for atomic settlement. The same Qt wallet page surfaces six sub-tabs — discovery, offers, sessions, governance, discussion, cross-chain — so a trade flows from "find a counterparty" through "negotiate privately" to "settle bilaterally" without a venue. The TensorSwap Solidity contract handles the EVM leg of cross-chain swaps.

Two spending rails live from block 0. The familiar ECDSA / Schnorr stack on Witness v0/v1, and an ML-DSA stack (NIST FIPS 204) on a new Witness v2. v2 outputs are 32-byte Taproot outputs spendable only by script-path proof; the only new opcodes are `OP_CHECKMLSIG` and `OP_CHECKMLSIGVERIFY`. Because TensorCash is a genesis fork, there's no installed UTXO base to retrofit and no soft-fork activation window — users pick a rail when they fund an address, and both remain valid forever.

Order matching and price discovery have throughput, leverage, and fairness requirements that don't fit on-chain. TensorCash specifies a higher-layer DEX that uses network latency and cryptographic fragmentation — order-book state Shamir-shared 6-of-10 across independent validator subsets — to protect against front-running and manipulation. No single subset learns flow; no single subset can halt trading. Final settlement always lands back on the base chain via the same atomic-swap primitives a bilateral trade uses.