Ferros doesn't optimise the EVM. It bypasses it entirely for known operations — and isolates everything else into lanes that never compete for resources.
Every parallel execution engine eventually crashes into the same two walls. We tested every published strategy. None solved both.
DashMap, ConcurrentHashMap, AccountsDB — every chain uses some variant of shared mutable state during execution. Under parallel writes, threads compete for shard locks. This caps scaling regardless of how clever the scheduling is.
A native ETH transfer through revm costs ~2us. The same operation as pure Rust — two HashMap lookups and two writes — costs ~200ns. For known operations, we pay 10x overhead for interpreter dispatch, gas metering, and memory management.
| Approach | Used By | Peak Speedup | Why It Breaks |
|---|---|---|---|
| Frame Barriers | Batch-parallel schemes | 1.8x | N-1 synchronization barriers cap parallelism |
| Multi-Version State | Database-style MVCC | 0.16x | Lock contention in version chains — slower than sequential |
| Block-STM | Aptos | 0.4x | 80-90% wasted execution at any real contention |
Pre-computed dependency graph + work-stealing execution. No multi-version state store. No speculative re-execution. Every transaction executes exactly once. Both bottlenecks eliminated.
None of them question two axioms: all transactions must execute through a VM, and gas must be priced in a volatile token.
Formally verified Rust functions that produce identical state transitions to their EVM equivalents. No interpreter. No gas metering. No bytecode dispatch.
Each CNP is mathematically proven to produce identical state transitions, events, and reverts as its EVM reference implementation.
Your ERC-20 transfer is $0.002 — worst case $0.008 during rare congestion. No gas auctions, no volatile token pricing.
New templates are added by governance vote. Not a precompile, not a JIT compiler. Business logic, not cryptographic ops.
No transaction modifies shared state during execution. Every lane reads from an immutable Arc<HashMap> snapshot — O(1) creation, zero locks, zero contention.
Writes are captured as write-intent logs and applied in a deterministic commit phase using optimistic concurrency control. Conflicts are detected by read-set validation and resolved by re-execution in global tx order.
This is fundamentally different from every other approach. Aptos speculatively writes to shared state and rolls back. Solana locks accounts. Monad processes sequentially. Ferros lanes execute against an immutable snapshot — no locks, no rollbacks, no contention.
Every block, 1% of Lane A transactions are randomly re-executed through the EVM and compared against the native result. If they ever diverge, the native executor is automatically deactivated, the registrant is slashed, and affected transactions are re-executed through the EVM. Probabilistic audit adds <1% overhead while catching bugs within a few blocks.
Known operations run on certified native Rust — purpose-built, audited, and 10-50x cheaper. Unknown operations run on a full EVM. Both are available to everyone. Economics decide what lane you drive in.
ERC-20 transfers, approvals, payments. Native Rust. No EVM interpreter. No gas metering.
Native currency transfers. Zero calldata. The simplest and cheapest operation on the network.
DEX swaps, lending, staking. Certified native execution for common DeFi patterns.
Multi-step compositions, complex DeFi. Native execution for advanced certified patterns.
Full EVM compatibility. Deploy anything — Solidity, Hardhat, Foundry. The permissionless innovation lane.
State Gravity: a single canonical state root over all state, proof-carrying cold reads, and fair pricing. No state expiry. No user-provided Merkle proofs. No data loss.
When a transaction touches dormant state, the proposer includes a Merkle witness (key + value + proof) in the block body. Validators verify the witness against the parent state root — no live archive dependency during execution. One-time thaw fee of $0.005 per key, then the state is hot again.
Creating a new storage slot costs $0.005 ($0.002 permanent + $0.003 refundable bond). Delete the slot later and get $0.003 back. This creates a natural incentive for state cleanup — without forcing it. Active users pay nothing extra. Only returning dormant users pay the thaw fee.
Agents need guardrails that survive code bugs and key compromises. Ferros enforces them at the protocol layer — before execution, not during it.
Spending limits, velocity caps, whitelists, business hours, and selector gates — enforced at the protocol level before execution begins. Zero gas consumed on rejection.
Policies cannot be removed by the wallet owner alone. Removal requires multi-sig authorization + timelock. A compromised key cannot silently disable spending limits.
Not ERC-4337 bolted on as a smart contract. Genesis-level architectural decision. Every account has a validation scheme dispatched through a layered pipeline.
Smart Accounts run validation in a restricted sandbox: own storage only, no external calls, no environment opcodes, no state writes. Prevents validation DoS attacks that plague ERC-4337.
Standard nonces force strict ordering. An agent sending 100 independent transactions must serialize them. If tx #50 fails, txs #51-99 are stuck.
Nonce domains solve this: each account has multiple independent nonce sequences. An agent sends transfers on domain 0, governance votes on domain 1, oracle updates on domain 2 — all in parallel, none blocking each other.
Lane A native executors are not trusted blindly. Each template is pinned to a contract codehash, verified against storage invariants, and must pass millions of differential tests (native vs. EVM producing identical state transitions).
Registrants stake collateral. If probabilistic audit detects divergence, the template is deactivated and the registrant is slashed — automatically.
FerrosBFT: a 2-phase pipelined BFT protocol built on HotStuff-2. O(n) happy path. Immediate-propose pipeline. No hand-waving on safety.
Leaders propose immediately on QC/TC receipt — no waiting for timeout intervals. The pipeline is self-sustaining: each committed block triggers the next proposal at network speed.
Finality requires 2 consecutive Quorum Certificates from ≥67% of validators. No single QC is treated as final. Provably safe under partial synchrony with f < n/3 Byzantine validators.
When a validator receives timeout messages for a view it already passed, it contributes its own timeout to help form a Timeout Certificate. Prevents deadlocks when validators join late or miss proposals.
Equivocation (double-voting) = full slash, provable on-chain. Missed slots = reputation penalty only. Proposer rotation + mandatory mempool utilization prevents sustained censorship.
3 validators with rotating leaders producing identical block hashes at blocks 100, 1,000, and 5,000. Stablecoin gas collection, fee distribution, and the full consensus pipeline are running end-to-end. 312 tests passing, zero warnings.
A platform where useful technology is the point — and the economics reward everyone who makes it work.