Every number, re-run on a fresh machine

22 September 2026. A new c8g.8xlarge in Virginia, 32 Graviton4 cores, nothing on it. I clone the public repository at commit 320d21d and re-run the receipts in docs/evidence: the zero-allocation gate, the layout, algebra and handshake tests, the ingest benchmark, the core scaling gate. Then three node processes on the same box converge, one is killed with kill -9, and it recovers from its write-ahead log. Five minutes. The numbers are the ones on the screen, including the ingest run that came in below my receipt.

Recorded

22 September 2026, 16:30 UTC

Machine

c8g.8xlarge · 32 Graviton4 cores · us-east-1

Commit

320d21d

Length

5 min 09 s · captions in English

Keys ←→ 5 sJL 10 sK play / pauseF full screenC captions0–9 jump

On the screen, against the receipts

Three runs of the same commands, all public: the receipt in the repository, the recording above, and scripts/reproduce.sh run on the recording's machine after the camera stopped. Benchmarks move from run to run; the spread below is that movement.

Measurement	Receipt in docs/evidence	On the screen	Same machine, after the recording
Write path, per insert	0 B, 0 allocs	0 B, 0 allocs	0 B, 0 allocs
Ingest, 32 cores, one signature per 100 deltas	5,681,818 deltas/s	4,608,294	4,444,444
Ingest, 32 cores, one signature per 256 deltas	6,024,096 deltas/s	6,369,426	6,410,256
Core gate, 16 cores	85,096,809 ops/s	83,482,668	79,106,994
Core gate, 32 cores, floor 50,000,000	68,278,197 ops/s	69,554,250	69,102,044

The three-node run at the end stays on one box: three processes, three loopback addresses standing in for three regions, 1,000 keys written into node 0, all three Merkle roots equal by the time node 0 went quiet at 3.01 s. Then kill -9 on node 0, and it answers again after 113 ms with the same root. That checks the mechanism, not the three-region result; the 100-process run across three regions has its own receipts, convergence-100-node.txt and crash-recovery.txt, in the same folder.

What is edited

The typing of one long command is speeded up. Where a line of the voice needs longer than the run took, the screen is held. The voice was recorded afterwards, one line at a time. The terminal output is not touched. For the web, the sound is level-normalised and the picture is encoded once more at constant quality, so that the whole recording is one 17 MB file the browser plays on its own; the re-encode scores 0.997 against the export on the SSIM scale, where 1.0 is identical.

Transcript

1. The machine
2. I am Harsh Rawat. I am seventeen. I am in Kota, in India.
3. This is a fresh cloud machine. 32 cores.
4. Neoverse V2. That is Graviton4.
5. The long command asks the cloud what the machine is.
6. c8g.8xlarge, in Virginia.
7. A clean box. Nothing of the project on it yet.
8. I will re-run the numbers from my application. Then a three-node mesh.
9. From nothing
10. This is the public repository.
11. Commit 320d21d. Everything here runs from this tree.
12. The memory rule
13. First, the memory rule.
14. The write path allocates nothing. This test fails the build if that ever changes.
15. Zero bytes, zero allocations per insert.
16. The records live outside Go's heap. The garbage collector never touches them.
17. Layout, algebra, handshake
18. Three quick checks. Layout first. Struct sizes and field offsets are pinned by tests.
19. Algebra. Merge in any order, in any grouping, merge twice, same result. One hundred random cases each.
20. This is what lets nodes merge without a leader.
21. Handshake. The default TLS config negotiates the post-quantum hybrid key exchange.
22. Signing stays classical by default. The hybrid signature is opt-in.
23. The ingest path
24. Now the real receive path, not the core.
25. One signature check per frame of 100 or 256 deltas. Then a merge for each delta, in process. No socket, no disk.
26. My receipt says 5.7 to 6 million deltas a second.
27. N = 100: 4.6 million. Lower than the receipt.
28. N = 256: 6.4 million. Higher than the receipt.
29. Benchmarks move run to run. Both runs are public: the receipt, and today's transcript.
30. The ceiling rows give every core its own engine, so nothing contends. That best case is 11 and 13 million.
31. The gap between the rows is contention on one shared engine. That is the part I am working on.
32. The core gate
33. The in-process core gate is 50 million a second.
34. Today: 83 million at 16 cores.
35. 69 million at 32 cores.
36. It passed.
37. Three nodes, kill -9
38. This is a three-node mesh across three simulated regions.
39. 1,000 keys injected into node 0, and all three roots converged.
40. Then kill -9 on node 0.
41. It recovered from checkpoint and log in 113 milliseconds, with its root matching again.
42. Close
43. Every number I said has a receipt in this folder, with the exact command and machine.
44. About 57 percent of version one is done.
45. The straggler tail and the durable path are the open problems.
46. This is the engine today. Thank you.

Files

• The commit the recording runs from: 320d21d
• The receipts: docs/evidence
• The re-run on the same machine, after the recording: reproduce-c8g-8xlarge-2026-09-22.txt
• The two scripts the recording runs: scripts/reproduce.sh and scripts/local-mesh.sh