This document states what the seventeen-oracle family is — and is not — from a cryptography and production-hardening perspective. Read it before treating oracle outputs as the sole trust anchor for high-value money flows.
Related: Chronos SECURITY.md · oracle-core SIGNING.md · Factory known-issues (KI-6)
| Claim in marketing | Honest status (2026-07) |
|---|---|
| Seventeen verifiable-math oracles | True — live capabilities, unit tests, AIMarket v2 integration |
| Ed25519-signed manifests & receipts | True — canonical forms match the Hub |
| Hybrid post-quantum (Ed25519 + ML-DSA-65) | Partial — implemented in oracle-core, off by default; Hub verifies Ed25519 only; no external crypto review |
| Chronos Wesolowski VDF | Research-grade primitive — standard construction, parameters in source, not independently audited, no formal verification |
| Hardened production cryptographic service | False today — see below |
The family was built in roughly two months of focused engineering. That is enough for a strong research / prototype stack (correct math, tests, demos, lottery integration) — not enough for a fully hardened cryptographic service with published setup ceremonies, external audits, and proof-of-correctness artifacts reviewers expect before mainnet-scale TVL.
What exists in-repo
- Fixed RSA-2048 challenge modulus (not per-request generation) — see
chronos/vdf.py. - Documented parameters: generator derivation,
Tclamp[1, 1_000_000], 128-bit Fiat–Shamir challenge prime, verification equation. - Unit tests including adversarial cases; Foundry test-vector alignment for on-chain consumers.
- SECURITY.md with setup transparency and threat model.
What a hardened production VDF service would still publish — and we do not yet
| Gap | Why it matters |
|---|---|
| Published parameter-selection guide | Operators need T ↔ wall-clock mapping on reference hardware, security margin vs ASIC/GPU speedups, and recommended values per use case (lottery vs ordering). Today T is caller-chosen within a cap. |
| Standalone security analysis | A reviewer-facing document (assumptions, adaptive-root usage, RSA-2048 anchor, grinding bounds) separate from README prose — suitable for citation in an audit scope letter. |
| External cryptographic audit | No Trail of Bits / NCC / similar report; no public audit PDF. |
| Formal verification | No Coq/Lean/TLA+ (or third-party) proofs tying implementation to Wesolowski semantics. |
| Operational transparency | No public attestation log of modulus fingerprint, deployed T policies, or hardware profiles used in production. |
Until closed: use Chronos for demos, testnet lottery, and grinding-resistance experiments. Do not treat it as the only trust anchor for large real-money flows without an independent review.
What exists
- Shared
oracle_core/signing.py: Ed25519 for manifests and receipts; optional additive ML-DSA-65 (FIPS 204 viadilithium-py) whenORACLE_PQC=1. - Hybrid verify requires both signatures when PQ fields are present.
- Tests in
core/tests/test_core.pyand Platon signing tests (skipped ifdilithium-pyabsent).
What is still thin
| Gap | Detail |
|---|---|
| Normative hybrid spec | AIMarket Protocol v2 documents Ed25519 receipts; the PQ extension (pq_algorithm, pq_public_key, pq_value) is implementation-defined in oracle-core, not a frozen RFC with test vectors in aimarket-protocol. |
| Hub parity | Live Hub verification path checks Ed25519 only; PQ fields are ignored unless the consumer implements verify_signature_object. |
| Default posture | PQC is off in production configs; the README badge describes capability, not deployment state. |
| Proof-of-correctness | No independent review of canonical string formats, key lifecycle, or hybrid composition (e.g. binding PQ key to Ed25519 identity). |
| Key management | File-based keys (data/*_signing_key, chmod 600) — no HSM/KMS integration guide for oracle operators. |
Specification: core/docs/SIGNING.md — describes what the code does today; not a substitute for audit or protocol freeze.
Each oracle ships domain math + invoke handler + tests + portal scene. Shared infrastructure (protocol, signing, metrics, rate limits) is real and reused.
What cannot be done credibly in a short sprint for all seventeen:
- Per-oracle cryptographic threat models reviewed by outsiders
- Side-channel analysis of hot paths
- Constant-time guarantees on non-crypto oracles (many are numerical, not constant-time by design)
- Unified security incident response and key-rotation runbooks
Honest tier: research / prototype with production-style integration (Hub, lottery, on-chain vectors for Chronos/Sortes). Not a bank-grade or L1-grade crypto service.
We will treat the oracle crypto layer as production-hardened only when all of the following are true (tracked as KI-6):
- External cryptographic audit report published (scope:
oracle-coresigning, Chronos VDF, Sortes ECVRF at minimum). - Chronos: published parameter guide + operator attestation checklist. — done 2026-10-12:
chronos/docs/PARAMETERS.md. - Hybrid PQC: normative extension in
aimarket-protocolwith negative test vectors; Hub verifies both layers when present. — done 2026-10-12: protocol §7.3.7–§7.3.8 + vectors. - Key-management runbook (rotation, compromise, multisig or HSM option) for oracle operators. — done 2026-10-12:
key-management.en.md.
Item 1 is still open. Self-verification on 2026-10-12 (standard RFC 8032 / RFC 9381 vectors and
negative cases) found and fixed three real issues — Sortes VRF without validate_key, Ed25519
verification accepting small-order keys, Chronos y vs N − y — see KI-6 in the Factory
known-issues. That is not an audit.
Until then: testnet, demos, and bounded pilots only — aligned with the Factory pre-mainnet checklist.
Cryptographic issues: open a GitHub issue on alexar76/oracles
with [crypto] in the title. Do not post working exploits publicly before a fix.
Other languages: crypto-maturity.ru.md