| # Phase 7 Design: Byzantine Validation and Proof-Carrying Consensus |
|
|
| Phase 7 makes peer collaboration adversarially aware. A node is not trusted merely because it participates: proposals must be signed, fresh, non-replayed, and accepted by a quorum that tolerates a declared number of faulty nodes. |
|
|
| ## BFT decision contract |
|
|
| `bft.py` implements a deterministic weighted decision layer. For fault tolerance `f`, the classical minimum participant count is `3f + 1`; the default decision also requires a two-thirds weighted quorum. Conflicting votes from one node in the same proposal round are classified as equivocation and withhold acceptance. Incentives are transparent accounting deltas, not a currency or an economic security guarantee: aligned votes receive a small reward after an accepted decision, equivocation receives a penalty, and honest dissent remains visible without punishment. |
|
|
| The protocol is intentionally separate from Lean. Consensus establishes distributed agreement about an artifact; it does not establish mathematical truth. A node’s incentive must never depend on suppressing dissent or upgrading a claim beyond its kernel-backed status. |
|
|
| ## Signed envelopes and replay protection |
|
|
| `security.py` now signs a canonical envelope containing `node_id`, payload, timestamp, and nonce. Verification checks the Ed25519 signature and content digest. `ReplayGuard` rejects stale timestamps and previously accepted `(node_id, nonce)` pairs. The Cloudflare Durable Object applies the same checks at the edge, stores nonce keys with a bounded expiration, rejects malformed or stale messages, and broadcasts only verified envelopes. |
|
|
| Production deployment should add peer admission, key rotation and revocation, room authorization, durable audit events, rate limiting, clock policy, and protection against storage exhaustion. A public Worker should not use a convenience key-generation endpoint as its identity authority. |
|
|
| ## Groth16 verification |
|
|
| The circuit in `circuits/private_premise.circom` proves a bounded private lower-bound relation. The local workflow now compiles the circuit with `circom2`, generates an initial Groth16 proving key from an 8-power ceremony artifact, generates a witness and proof for `secret_value = 7` and `public_floor = 5`, exports a verification key, and verifies the proof with snarkjs. The Flask endpoint `/api/zk/verify` invokes the same verifier against the local verification key or an explicitly supplied key. |
|
|
| The demo uses an initial proving key for reproducibility; it is not a substitute for a production multi-party ceremony. The ceremony artifact provenance and hash must be pinned, and the final system should use a reviewed contribution chain, independently verified keys, circuit audits, public-input binding tests, malformed-witness tests, and proof-replay policy. The API reports verification only when snarkjs returns success. |
|
|
| | Layer | Evidence | Status | |
| |---|---|---| |
| | Circuit source | Circom constraints | Implemented | |
| | R1CS/WASM compilation | circom2 output | Implemented locally | |
| | Groth16 setup | Initial zkey from pinned demo ptau | Implemented locally; ceremony trust limited | |
| | Proof generation | Witness and proof JSON | Implemented locally | |
| | Proof verification | snarkjs and Flask endpoint | Implemented locally | |
| | Production privacy guarantee | Audited ceremony and circuit | Not claimed | |
|
|
