Single-laptop, single-core, no-GPU wall-clock comparison of two zero-knowledge proving stacks — Stwo (StarkWare) and SP1 with a Groth16 wrap (Succinct Labs) — on one fixed proof statement:

There exists a reversible classical circuit C over {NOT, CNOT, Toffoli} that realises one elliptic-curve point addition over secp256k1, attested in zero knowledge against a public set of test cases.

One number ships: t_SP1_Groth16 / t_Stwo, with median, IQR, min, max, stddev over ≥10 measured runs per prover. Around it: peak RSS, proof size, verifier time, constraint count, and the trusted-setup cost of the SP1 side, reported separately.

Specification frozen. Implementation pending against milestone v0.1 (57 issues). RESULTS.md is produced on the reference rig by ./scripts/run_all.sh once implementation closes.

Apple M4 Max, 16 cores (1 used), 48 GB RAM, macOS 26.2, AC power, Wi-Fi/Bluetooth off, low-power mode off. Full toolchain pinned in versions.lock once implementation lands. A Linux x86_64 CI rig provides regression coverage but is not headline-canonical.

git clone https://github.com/AbdelStark/grover-tax.git
cd grover-tax
./scripts/run_all.sh

Expected wall time on the reference rig: ~25 min. Hard ceiling: 45 min. Output lands in RESULTS.md and results/.

Only one intentional divergence between the two sides: the circuit-commitment hash function (SP1 uses SHA-256 — upstream's native choice; Stwo uses Blake2s — a bit-oriented hash in the same structural family that Stwo Cairo has as a built-in). Both commitments are computed over identical bytes; both verifiers bind. The full justification is in RFC-0005. Other structural differences (BabyBear vs M31 field, trusted setup required on SP1 side only, macOS frequency-pinning gap) are disclosed in RESULTS.md.

• PRD.md — original technical brief, frozen as historical record.
• SPEC.md — index of the implementation specification corpus.
• docs/spec/ — eleven specification sections (overview, architecture, public API, data model, error model, observability, security, testing, performance, release, glossary).
• docs/rfcs/ — fourteen RFCs locking the load-bearing technical decisions.
• docs/roadmap/IMPLEMENTATION.md — implementation tracker keyed to GitHub issues.

The methodology is contractual. Any change that affects measured numbers requires an RFC update before merge. CONTRIBUTING.md (pending, issue #45) will document the PR workflow.

MIT. See LICENSE (pending, issue #45).