Evidence & benchmarks · v0.2.0
What ran, what it measured, and how to check it.
Everything below ran on Solana Devnet on 18 October 2026 through Octoarka’s API. Each landed transaction is linked to Solana Explorer, so you can verify it independently.
Proofs
Each outcome, labelled honestly.
| Proof | What happened | Kind | On-chain |
|---|---|---|---|
| Order MATCHED | Buy 2 tSOL at an average of at most 106; filled at 103.88. | Real | 2iC7Ch…EvLQq ↗ |
| Order PARTIAL | Buy 15 against a 10-tSOL book; 10 filled within the limit, the rest cancelled. | Real | 5LxrNX…CSMtq ↗ |
| Order FAILED, proven not executed | Limit far below the book: refused, then proven not executed after its lifetime expired. Nothing moved, no fee. | Real | not landed by design |
| Lost reply → UNKNOWN → resolved | The reply was discarded on purpose; after a restart the order was resolved MATCHED from chain evidence, with one send. | Fault injected | 2eV5tb…EqYs8 ↗ |
| DvP settlement through the API | Atomic asset-for-payment settlement submitted by a trading application; MATCHED. | Real | 2o8kZG…yqVCs ↗ |
| Kill switch | With admission paused, a new order was refused before anything was sent; pause, refusal and resume audited. | Real | not landed by design |
| Duplicate refused on-chain | A rebuilt copy of an executed order was rejected by the Execution Guard (receipt exists). Bypasses Octoarka on purpose. | Labelled test | not landed by design |
| First DvP settlement (v0.1.0) | Lost acknowledgement injected; UNKNOWN; restart; MATCHED with exactly one signed attempt. | Fault injected | 128Jwa…n5tkM ↗ |
Octoarka Execution Guard program: 44LuVy…NhG5 ↗. Its on-chain bytes match our build.
Benchmarks
Where the time goes, stage by stage.
Durations in milliseconds, computed from the runtime’s own journal timestamps. The engine’s own work takes milliseconds; the rest is reading Solana at the strongest confirmation level and waiting for the network.
| Stage | Order MATCHED | Order PARTIAL | DvP |
|---|---|---|---|
| Accepted → state resolved (finalized reads) | 2,553 | 2,516 | 3,685 |
| Resolved → plan authorized (policy) | 14 | 15 | 16 |
| Authorized → exact message bound | 2,459 | 2,341 | 1,562 |
| Signer request → signature verified | 20 | 9 | 15 |
| Verified → attempt recorded durably | 4 | 2 | 1 |
| Recorded → RPC acknowledged | 364 | 242 | 362 |
| Acknowledged → reconciled (finalized) | 3,582 | 3,650 | 3,615 |
| Reconciled → final report | 2 | 2 | 3 |
| Total: accepted → final report | 9,004 | 8,779 | 9,272 |
24.5k–33.8k compute units
per guarded order, 26.1k for the DvP settlement; 5,000 lamports fee each. Refused transactions paid no fee.
49 µs p50 commit
Local measurement over 1,000 acceptances (p95 66 µs, p99 147 µs). Power-loss durability of the device was not tested.
727 ms p50 book age
Age of the live Mainnet reference book as the trading app displays it, over 120 s (p95 1.85 s).
Scope: single runs, one machine, one RPC provider, Solana Devnet only. These are not production latency claims, SLAs or throughput figures. Demo-market liquidity came from our own maker.
Engineering verification
How the release was checked.
372 automated tests
Lifecycle, policy, journal, recovery, reconciliation, API and both apps; 0 failed.
Formal model
The recovery logic is model-checked in TLA+, with a control that proves the check can fail.
Reproducible packages
Two independent builds produce bit-identical packages; tested on clean Ubuntu systems.
Security checks
Secret scanning, dependency audit, sandboxed services; a TLS advisory fixed before release.
The complete evidence pack (logs, manifests, reproduction steps) is available to evaluators. Ask for it.