Changelog

[0.3.0] — 2026-09-16

Coordinated release: Trisha 0.3.0, with Trident 0.3.0/compiler API 3 and Joy 0.5.0. Validation receipts accompany the release archives.

• Own Triton lowering/emission, cost and neural target integration, Neptune runtime modules/network states and all 43 independent manual baselines.
• Upgrade the native and recursive proof backend to pinned Triton/tasm-lib 7.0.0, native claim version 5 and stark-triton-v7. Regenerate older proofs. Shared CPU/wgpu validation rejects noncanonical inputs, oversized domains, trailing proof items and malformed binary encodings.
• Verify complete expected native claims in the production recursive SDK. Fixed Neptune transaction/native-currency policies bind canonical programs and the caller's full commitments. File witnesses use private, new files.
• Add the pinned trisha-neptune adapter: construct canonical compiled-lock outputs, prepare complete SingleProof transaction intents, and submit through an explicitly authenticated gateway. Program inspection remains offline; generic deployment without transaction intent remains unsupported.
• Repair pinned Neptune mining fork/hash layouts, portable CPU hashing, CPU/GPU ownership, bounded nonce reservation and template refresh. Wallet import is interactive; selected network and wallet command outcomes are checked. Add the explicitly local local-testnet1 state for real-proof tests.
• Correct compiler/stdlib behavior against independent execution and negative vectors. Fresh full proofs, device tests, formal analysis and node admission are recorded separately in validation records.
• Marshal typed source entry parameters from public input and constrain Bool/U32 values. Preserve declaration-order struct fields and terminal return values; emit canonical native Field literals. Reject unresolved TIR diagnostics.
• Record each freshly verified classic/hand baseline proof with its exact program and public claim; installed gates require the complete proof pairs.
• Add deterministic source archives, locked fixture generation and isolated installed compiler/warrior smoke checks on coordinated source candidates.

The default prover is the CPU backend. Shader inventory and historical mining numbers do not establish a complete GPU prover; current boundaries are in GPU architecture.

[0.1.0]

Historical development notes. Performance claims below were not reproduced by the current release review; see archived GPU claims.

added

• trisha run — execute Trident programs on Triton VM
• trisha prove — generate STARK proofs of execution
• trisha verify — verify STARK proofs
• trisha deploy — stub (prints digest, no neptune-core)
• trisha guess — GPU nonce mining, 24M H/s on M1 Max
• batch mode for all commands (trisha <op> batch)
• GPU backend: 7 WGSL shaders (goldilocks, ntt, tip5, fri, poseidon2, gemv, mine)
• kernel fusion: NTT + Merkle in single command encoder
• VRAM budget management with CPU fallback
• persistent buffers: twiddle cache, two_inverse, Tip5 constants
• staging pool: 8 pre-allocated readback buffers
• proof file format: .proof.toml (TOML envelope + bincode + base64)
• vendor patching: patches/apply.nu injects GPU hooks into triton-vm
• 4.3× proving speedup on small programs (M1 Max)