warriors/trisha/audit/legacy-roadmap-claims.md

Historical roadmap claims — retained 2026-09-12

The text below was copied from the earlier roadmap before correcting its status assertions. It is historical material, not current release evidence. The default CLI proves on CPU; real cycle/height metadata is implemented. Claims of complete GPU proving, speedups and mining readiness below lack current pinned end-to-end receipts. Current evidence belongs in this audit directory and the coordinated Trident release ledger.



tags: trisha, roadmap crystal-type: spec crystal-domain: cyber

trisha roadmap

trisha closes the loop: source → compile → run → prove → verify → deploy.

status

layer status what works gap
runner done trisha run executes TASM via triton-vm cycle count always 0
prover done trisha prove generates real STARK proofs cycle_count, padded_height metadata
verifier done trisha verify validates proofs
batch done trisha <op> batch for all operations
proof file done TOML + bincode, roundtrip tested
deploy stub prints digest, says "not yet available" no neptune-core, no RPC
miner (CPU) done trisha mine --neptune — rayon, 0.33 MH/s on M4 Max
miner (wgpu) done Mine.wgsl + Tip5.wgsl wired into wgpu accelerator not connected to neptune_mine.rs
miner (honeycrisp) done MSL Tip5 kernel, 1.1 MH/s (3.3× CPU), --features gpu NEON/unimem zero-copy (Phase 2/3)
GPU shaders done 7 WGSL shaders (goldilocks, ntt, tip5, fri, poseidon2, gemv, mine)
GPU proving done all 7 hot paths: hash, iNTT, NTT, Merkle, FRI fold, GEMV
GPU verify done FRI fold on GPU, parallel batch via scoped threads double-buffering
kernel fusion done NTT + Merkle: single command encoder, zero per-layer sync
persistent bufs done twiddle cache, two_inverse, Tip5 constants
staging pool done acquire/release across all 7 readback sites
VRAM streaming done budget awareness, chunked hash/GEMV, CPU fallback guards streaming NTT (four-step FFT)

what's real

full prover pipeline on GPU. 7 compute shaders dispatched to Metal/Vulkan/DX12 via wgpu. kernel fusion eliminates per-layer CPU↔GPU sync. VRAM budget auto-detected from adapter limits; oversized buffers gracefully fall back to CPU. dependency patching injects GPU hooks into triton-vm without maintaining a fork. 4.3× proving speedup on small programs.

neptune HardforkBeta CPU mining: 0.33 MH/s on M4 Max (rayon + triton-vm Tip5). honeycrisp GPU mining: 1.1 MH/s (39 Tip5 permutations per nonce; Metal MSL kernel; compute-bound at ~65% GPU utilization — practical ceiling for this algorithm on M4 Max). GPU path activated via --features gpu; trisha mine --neptune --bench-gpu 10 benchmarks. CPU+GPU combined: ~1.43 MH/s. wgpu Tip5 + Mine.wgsl shaders exist and compile but not wired into neptune_mine.rs.

what's scaffold

deploy (prints metadata, no blockchain interaction). honeycrisp Phase 2 (NEON Tip5 via acpu) and Phase 3 (unimem zero-copy IOSurface).

proposals

proposal status goal
honeycrisp-npt-mining phase-1-done MSL Tip5 kernel live (1.1 MH/s); Phase 2: NEON+AMX; Phase 3: unimem zero-copy
gpu-proving-at-scale open streaming NTT for 2^25+ row traces
mining-tree open M=29 neptune mining tree (obsolete for mainnet — HFB removed memory-hard PoW)
proof-merging open trisha merge — proof-that-verifies-proof
real-workload-benchmarks open actual speedup numbers across trace sizes
double-buffered-batch open pipelined GPU batch verification
deploy open neptune-core integration, on-chain programs
missing-metadata open cycle_count, padded_height in proof file
helical-parallelism open multi-strand coprocessor proving: GPU hash + AMX field + NEON bookkeeping in parallel; O(log N) sequential depth

confidence milestone

# cpu pipeline (passing)
trisha run /tmp/hello.tri
trisha prove /tmp/hello.tri --output /tmp/p.toml
trisha verify /tmp/p.toml

# batch (passing)
trisha run batch /tmp/a.tri /tmp/b.tri
trisha prove batch /tmp/a.tri /tmp/b.tri --output proofs/
trisha verify batch proofs/a.proof.toml proofs/b.proof.toml

# GPU (passing)
trisha prove /tmp/hello.tri  # dispatches NTT/Merkle/FRI/GEMV/Hash on GPU

# mining (passing)
trisha mine /tmp/hello.tri --difficulty 1000000000000000000

# not yet passing
trisha deploy /tmp/hello.tri --state testnet
trisha merge proof1.toml proof2.toml --output merged.proof.toml

Graph