Each row is one run: what it measured, the figure, the objects per send and the connections, the box and the date. The rail moves 33 bytes per send; a company's records stay in its own book.
| what | figure | per send · connections | where, when |
|---|---|---|---|
| The live run through the whole rail | 100,000 sends a second for the run's first 33 minutes; the cell folded 59.7, 60.1 and 60.1 million objects into one root per ten-minute period (100,167 a second), nothing dropped at the door or at the cell's intake, every root sealed 5–9 s after the close see it → | 1 × 33 B · 8 persistent TCP lanes, a paced synthetic source | Telemach, the door on hash001 → the rail's cell, 29 Sep 2026 19:30–20:03 UTC |
| The same pace with the edges' proofs | 60,099,588 objects in one root; the cell's proof index ready 18–21 s after the close; every real edge's period FINAL on time see it → | 1 × 33 B · 8 lanes, four edges alongside | Telemach, 29 Sep 2026 23:00 UTC |
| Objects folded into roots since 26 September | the running total, read live from the cell's own period counts see it → | — | Telemach, live |
| Four company edges, 40 hours | 631,750,544 records booked at 1,000 a second per edge; every period with records FINAL; 11,080 FINAL periods re-verified offline, strict (Ed25519 and ML-DSA-65), 0 failures see it → | 1 × 128 B per record · one edge per company | Telemach, 29 Sep 23:33 – 1 Oct 15:33 UTC; verdict 8 Oct 2026 |
| An edge's memory | about 250 MB per edge container (122 twin edges: 257 MiB on average, 213–285; the four pilot edges: 225 MiB); the collector beside it 8–14 MiB | — | Telemach, kubectl top, 21 Oct 2026 |
| what | figure | per send · connections | where, when |
|---|---|---|---|
| The door to a test cell | 29,999,992 of 29,999,992 records in 300 s = 100,000 sends a second; 0 errors, 0 dropped, 0 duplicates; the door 3.94 cores, 1.96 MB of memory | 1 × 33 B · 16 persistent TCP | hash001, 25 Sep 2026 21:27–21:32 UTC |
| The door into the rail's cell, 15 minutes | 90,000,065 received = 90,000,065 forwarded at 100,000 sends a second; 0 dropped, 0 errors; queue peak 314 of 4,096; the door 3.6 MB of memory | 1 × 33 B · 16 lanes | Telemach, 27 Sep 2026 20:30–20:45 UTC |
| Ten thousand companies through one door | 120,000 sends in 120.6 s, one fresh connection per send; 0 refused, 0 dropped; 0.41 ms mean, 7.7 ms slowest | 1 × 33 B · one connection per send, up to 512 in flight | neo-7 (2 cores), 27 Sep 2026 |
| The binary accounting cell | 30,000,000 records in 300 s = 100,000 a second; 0 errors, 0 dropped; the cell killed with its hour open and resumed with all 30,000,000; the root FINAL | 1 × 33 B · 16 persistent TCP | hash008, 25 Sep 2026 |
| The DLT alone, four hours | 1,439,500,000 objects in 14,395 s = 100,000 a second; 0 errors, every object present once and in lane order; each ten-minute period sealed 195–204 s after its close | 1 × 33 B · 16 persistent TCP | 27 Sep 2026 12:00–16:00 UTC |
| The light first server | 299,984 objects a second acknowledged into the first server's memory on 0.67 core, no cryptography on that step; 54,001,283 objects, 0 lost, 0 duplicated | 1 × 32 B · 16 signing sources | hash008, 21 Sep 2026 |
| One edge's intake | 100,000 records booked in 20.1 s = 5,000 records a second into one 60-second period; the period FINAL; the strict verify of 110,000 records 0.21 s | 1 × 128 B · 1 TCP | neo-7, 27 Sep 2026 |
| what | figure | per send · connections | where, when |
|---|---|---|---|
| Their audit-logging signal, 10,008 companies | 58,623,047 audit records answered in two hours, 0 lost, 0 refused, 8,140 requests a second; 1,200,960 periods sealed, all admitted see it → | 1 record per request · 10,008 companies, one book each | rig-1 (16 vCPU), 20 Sep 2026 10:46–12:46 UTC |
| Behind ApeiroRA's collector, strict | unsigned records refused; signed records admitted; signed then altered refused; a real VM lifecycle 4 of 4 verified see it → | 1 record per request | rig-1, 25 Sep 2026 |
| A month on OpenNebula, run 3 | 14 of 14 CLU checks; OpenNebula's own accounting agrees 12 of 12; 14 forged packs refused; a provider's host lost and restarted within the run see it → | — | rig-1, 25 Sep 2026 |