A proposed post-quantum stealth-address scheme for Ethereum, targeting a new ERC-5564 scheme ID and working against the deployed ERC-5564 announcer with no protocol changes. Meta-address distribution (naming services, off-chain sharing) and the ERC-6538 registry are outside the proposal's scope (D-014 scope update, 2026-09-28).
Repository: https://github.com/Skanislav/pq-sap · Docs site: https://skanislav.github.io/pq-sap
- Detection — ML-KEM-768 (FIPS 203). The announcement's ephemeral key is a KEM ciphertext; the shared secret drives a one-byte view tag and the address derivation. This is the part "harvest now, decrypt later" threatens, so it is the priority.
- Address derivation — the commitment format (
0x02) pairs a 32-byte spending-key commitment with the ML-KEM viewing key: 1,217 bytes total. The sender derives a fresh commitment and account address using the shared secret and an agreed deployment configuration, without learning the spending secret. Construction A (0x01, 5,633 bytes) remains available for accounts using a per-address blinded ML-DSA key. Whether0x02becomes the only normative format remains a decision for the ERC draft (D-024). - Spending — delegated to the account model. The browser demo proves knowledge of a spending secret and commitment opener (D-025). Its current UltraHonk backend is not post-quantum sound. Direct SPHINCS-C13, blinded ML-DSA, and a direct ML-DSA-44 committed-key profile (D-027, local contract path only) provide separate implementation evidence, with their own privacy and security boundaries; the direct routes reveal the key at spend time.
The proposal's focus is ML-KEM-secured discovery and sender-computable
addresses under existing ERC-5564 announcements. Spending integrations have
additional account and network requirements. See the
EIP-8288 summary for a prospective
aggregation route and the security analysis for
the current key-exchange review boundaries. Start with the
ERC evidence map; spending specialists can follow the
separate research guide, whose measurements are
assembled in the spend stretch-goal report.
The cohort's end-of-work revision of the proposal is
docs/PROJECT_RETROSPECTIVE.md (D-035).
This repository is the artifact trail of our EPF cohort work: the scheme we specified, the vectors it generates, the checks we ran, and the decision log recording why. Related external proposals are not mirrored here as code or text (D-032). ERC-8441 is not prior art: its first public draft was committed 2026-08-30 and its ERC merged 2026-10-29, both after this project's implementation, vectors, and proofs existed (first commits 2026-08-06/07), so it is a concurrent design — compared in review, not something we built on or lagged behind.
| Path | What |
|---|---|
docs/erc-draft.md |
Reserved for human-written ERC text (target scheme ID 2) |
docs/TECHNICAL_SPEC.md |
Working technical specification |
docs/DECISIONS.md |
Dated ADR log of design decisions |
python/ |
Executable Python spec, reference library, test vectors, benchmarks |
js-client/ |
TypeScript scanning client that reproduces the vectors byte for byte |
ui/ |
Demo web UI (Vite + React): receive, scan, and spend (EOA + ERC-4337 routes) |
docs/ERC_EVIDENCE.md |
Required claim evidence, relevant proofs, and open commitment-profile composition |
docs/research/cost-report.md |
Measured L1/L2 announcement and spend gas, registration, and scanning costs |
docs/SPENDING_RESEARCH.md |
Optional account, signature, and ZK research |
lean/ |
Generic KEM results, Construction A research, and maintainer tooling |
wiki/ |
Mirrored documentation site source |
The reference libraries, deterministic vectors for both address formats, generic KEM results, Construction A security analysis, and browser ownership-proof demo exist. Python defines the byte-level behavior; TypeScript reproduces the vectors; Lean supports specifically scoped mathematical statements.
Next: review the security boundaries of each route, prepare a self-contained vector checker and ERC assets, and bring the proposal to community review. The ERC text is human-written only under D-024 and is not frozen. See the submission gap analysis.
The conformance vectors and checker are intended to enter the
ethereum/ERCs asset tree under CC0 when
the PR is opened.
Parts of this repository were drafted with AI assistance. The disclosure is precise rather than gestural, because a reader deciding how much to trust a lemma or a vector deserves to know what produced it:
- Humans are the authors of every project-facing specification, the ERC text (D-024), and every security decision in the decision log. No AI tool is listed as a co-author, and no AI tool signs off on a claim.
- The models used were Anthropic's Claude. Git trailers record the exact identifiers, so the list below is mechanical, not recalled: Claude Fable 5 (35 commits), Claude Fable 5.1 (18), Claude Opus 4.8 (15), and Claude Opus 5 with 1M context (2).
- Scale. 70 commits carry a
Co-Authored-By: Claudetrailer, spanning 2026-08-06 to 2026-09-10; three commits are authored asClaude <noreply@anthropic.com>. No commit after 2026-09-10 carries an AI trailer. - Where it was used: documentation drafting, test scaffolding, exploratory code, wiki/sync plumbing, and audit passes over the tree. The conformance vectors are generated from the Python reference — determinism, not trust in a model, is what makes them checkable — and every implementation claim is backed by a run recorded in this repository.
Read the code and the vectors, not the prose: the vectors are the contract, the
Lean proofs are audited for axioms and sorry on every build, and the tests
replay byte-for-byte across Python, TypeScript, and Solidity. Where that
machinery does not cover a claim, the claim is marked as an assumption or an
open obligation rather than asserted — see
docs/ERC_EVIDENCE.md.
To audit the assist rate yourself:
git log --format='%b' | grep -c 'Co-Authored-By: Claude' # 70 co-authored commits
git log --format='%b' | grep -oE 'Co-Authored-By: Claude [^<]+' | sort | uniq -c
# 35 Co-Authored-By: Claude Fable 5
# 18 Co-Authored-By: Claude Fable 5.1
# 15 Co-Authored-By: Claude Opus 4.8
# 2 Co-Authored-By: Claude Opus 5 (1M context)Licensed under the MIT License.
The reference implementation builds on widely used, permissively licensed cryptography libraries. Runtime and benchmark dependencies include:
kyber-py(ML-KEM): MIT / Apache-2.0dilithium-py(ML-DSA): MITpycryptodome(Keccak/SHA-256): BSD-2-Clause / public domainliboqs-python: MIT (optional audit backend)coincurve(secp256k1): MIT / Apache-2.0 (optional, classical hybrid benchmark)@noble/post-quantum: MITviem: MIT@openzeppelin/contracts: MIT
There is no GPL or other copyleft dependency in the hot path. The ERC
submission still intends to place the conformance vectors and checker in the
ethereum/ERCs asset tree; the ERC repo
uses its own CC0 process for the submitted spec and assets.