research: poseidon-constraint-attribution

[alexandre-mrt]

Summary

Tonight's nightly research-loop run. Queue item #1 (on-chain gas per entry point) was re-attempted first and re-confirmed BLOCKED — this sandbox's egress proxy returns 403 for both github.com (needed for a sui CLI release) and fullnode.testnet.sui.io (a fallback JSON-RPC read), an org-level allowlist denial rather than a per-tool prompt this time. Per the sandbox's own proxy docs, that's not retried or routed around, so the run moved to the next un-settled queue item.

That turned into this experiment: the depth-20 Merkle membership proof, not the "four named Poseidon instances" README credits, is the dominant non-linear-constraint cost in transfer.circom and compliance.circom.

• Built 10 isolated single-gadget probe circuits (scripts/bench/circuit-probes/*.circom) and a reusable script (scripts/bench/constraint-attribution.mjs) that compiles each probe plus fresh copies of all three production circuits, then reconciles a hand-written per-circuit "recipe" against the real compiled total.
• Result: Merkle path = 76.0% of transfer.circom's and 81.2% of compliance.circom's non-linear constraints, vs. only 18.0%/14.1% for the named domain-tagged Poseidon calls. withdraw.circom (no Merkle path) is the one place the README's original claim holds (78.0%).
• Attribution reconciles to the constraint: 0/3/0-constraint deltas across the three circuits (the compliance delta is fully explained — two boolean-enforcement constraints + one AND gate not present in the isolated comparator probes).
• Per-Merkle-level cost is exactly 246 non-linear constraints (243 for Poseidon(2) + 2 for MultiMux1(2)'s selection + 1 boolean check), architecture-fixed — a real number for the queue's Merkle-depth-vs-anonymity-set item to compute against instead of guessing.
• Toolchain note: used circom2 (npm-bundled WASM build of the real Rust circom compiler) since github.com/cargo's crate host are blocked here — it reproduces the 2026-07-22 baseline's native-circom constraint counts exactly, a useful fallback for future nights hitting the same block.

Docs updated: README.md and docs/research/BASELINE.md corrected (the "four Poseidon instances dominate" claim was materially misleading for two of the three circuits), docs/research/EXPERIMENTS.md re-ranked (Merkle-depth work promoted ahead of a Poseidon2 swap; Poseidon2 re-targeted at the Merkle-path arity, not the named hashes), docs/research/LEDGER.md appended, full write-up in docs/research/2026-09-27-poseidon-constraint-attribution.md.

No circuit, Move module, or frontend proving code was modified — this is a measurement/attribution night only.

Test plan

• node --experimental-vm-modules test/{transfer,compliance,withdraw}.test.mjs — 108/108 pass (fallback/simulated mode — real-proof mode needs the pot15 ptau file, and storage.googleapis.com is also blocked in this sandbox; a new, previously-unseen blocked host, recorded honestly in BASELINE.md)
• cd scripts && bun run src/test-converter.ts — 109/109 pass
• cd scripts && bun run src/test-compliance-utils.ts — 67/67 pass
• cd frontend && bun run test — 19/19 pass
• cd contracts && sui move test — NOT RUN, sui CLI unavailable (same blocker as queue item feat: Veil privacy payment protocol — full stack MVP #1); no contract code changed
• cd scripts/bench && npm install && node constraint-attribution.mjs — reproduces the numbers in this PR's report exactly

🤖 Generated with Claude Code

https://claude.ai/code/session_016Kg1ZFhKEHbxyx3JwiqEh5

---

Generated by Claude Code

[alexandre-mrt]

CI is red on this PR (3 of 4 jobs), but both failures are pre-existing and unrelated to this PR's diff — this PR touches only scripts/bench/, docs/research/, README.md, and .gitignore, none of which either failing job's steps depend on.

1. Frontend (vitest + biome + tsc) and Proof converter + compliance utils — fail at "Set up job", before any repo code runs:

##[error]Unable to resolve action `oven-sh/setup-bun@735343b667d3e6f658f44e0a462d63b48e3324cb`, unable to find version `735343b667d3e6f658f44e0a462d63b48e3324cb`

Confirmed this is not caused by this PR: the exact same failure, on the exact same two jobs, reproduces on main's current HEAD (f942fca, this PR's own base) — run 30348614126. The pinned SHA for oven-sh/setup-bun no longer resolves (upstream likely force-pushed or deleted that ref). I don't have GitHub access outside this repo (alexandre-mrt/veil) to look up a current valid SHA for oven-sh/setup-bun, and re-pinning .github/workflows/ci.yml blindly is outside this PR's scope (a circuits-research PR touching CI config for an unrelated action would be guessing at a fix I can't verify). Proposed fix for whoever picks this up: re-pin oven-sh/setup-bun in .github/workflows/ci.yml to a current release commit SHA (check https://github.com/oven-sh/setup-bun/releases) — both jobs should go green immediately once that's done, no code change needed.

2. Circom circuits (real Groth16 proofs) — fails downloading the Powers of Tau file:

curl -fsSL -o build/pot15_final.ptau https://storage.googleapis.com/zkevm/ptau/powersOfTau28_hez_final_15.ptau
curl: (22) The requested URL returned error: 403

This same URL also returns 403 from my own sandbox tonight (see this PR's report, docs/research/2026-09-27-poseidon-constraint-attribution.md) — for me that's my sandbox's egress allowlist, but here it's GitHub's own hosted runner getting 403 directly from storage.googleapis.com, a different network entirely. That points at the object/bucket itself (possibly rate-limited or access-restricted), not a one-off network blip. I don't have permission to re-run the job to confirm (rerun_failed_jobs returned 403 Resource not accessible by integration), and this PR doesn't touch circuits/scripts/compile*.sh or the CI workflow, so there's nothing in this PR's diff to fix here either. Flagging for visibility — the same host also blocks a step of tonight's own research (see the report's Open Questions).

No production code, circuit, or CI config was touched by this PR, so I'm not pushing a fix for either — both are infrastructure/CI-config issues that predate and are independent of this change. Will keep watching this PR.

---

Generated by Claude Code