Current implementation: the shared-RAM contract and engineering record
FD&F acceptance record.
| Suite / checkpoint | Recorded result | Coverage |
|---|---|---|
| r19-a14 fdf-tun | 20/20 | Three scopes with the same served prefix and assigned leaf address; bidirectional aliases, concurrent traffic, atomic configuration refusal, fd/protect path and teardown. |
| r19-a14 reach | 68/68 | Reach rules, identity-only resolution, ambiguity refusal, aliases, session loss and the Kotlin bridge stage. |
| r19-a14 region-loop | 20/20 | Real daemon and region/hub in namespaces, including Communicator endpoint resolution. |
| r19-a14 graph / hub | 17/17 · 25/25 | Graph and hub behavior in their individual suites. |
| r19-a14 placement / leaf-tunnel | 18/18 · 10/10 | On-mesh placement and routed leaf behavior. |
| r19-a14 member selftest | 24/24 | Planner, keys, ledger, quarantine, concurrency and bridge loss. |
| r19-a15 regression | Region-loop PASS | A runMerge preserves leaf names; failed against r19-a14. |
Counts are reproduced from the checkpoint records. They have not been rerun as network tests while editing this website, and they must not be summed into a fictitious single suite.
Recorded hardware demonstration, October 12.
The Windows machine begins on its ordinary Internet connection and cannot reach the private FrogNet address. Starting the bridge produces a forward-via-hub path. Ping and SSH work; the PC appears in the node’s hosts file and monitor. The monitor and live topology map are accessible.
An Android client joins and participates in a live Communicator call with Windows. The displayed statistics report 1280×720 at 20 frames per second. Stopping the bridge removes the Windows machine’s private-network access.
This is a recorded operational demonstration, not a bandwidth benchmark. The latest take reports roughly 1.2–1.5 Mb/s at the receive-side stats. The Android device’s presence in /etc/hosts is recorded; an earlier take explicitly notes that its monitor row was not yet visible. Read the demonstration sequence.
| On-camera observation | Result | Context |
|---|---|---|
| runMerge rebuilding the network view | 4.9 s, then 0.79 s | Timed command after removing the hosts file; the second invocation is a repeat run, not a separate failover measurement. |
| Windows–Android call | 1280×720 · 20 fps · roughly 1.2–1.5 Mb/s | Displayed receive-side statistics in this demonstration, not a bidirectional total or a portable bandwidth requirement. |
| System code size | Under 35,000 lines | Reported in the demonstration for the system’s action code. |
| Android application code size | Under 6,000 lines | Reported in the demonstration for the Android app. |
The line counts are reported in the October 12 demonstration, not a fresh source audit. A reproducible LOC comparison requires the checkout, counting tool and exclusions for tests, third-party/vendor code, generated files and build output.
Discovery and liveness measurements.
| Experiment | Recorded observations | Conditions |
|---|---|---|
| r15 address-change pipeline | 66 ms facts + 7 ms announce + about 5 ms local apply; about 80 ms end to end | LAN namespace experiment; WAN adds region RTT for publication and world read. |
| r17→r19 per-event processing | r17: 22–29 ms; r18: 16/18 ms; r19: 33–45 ms | Namespace runs, median event totals. r19’s steady path is unchanged from r18; the record identifies environment variance. |
| Five-second liveness configuration | Silent member loss about 3.9 s; silent region loss about 4.0 s | C++ and Kotlin compared on the same namespace setup. |
| Loss tradeoff | More false deaths at 10–20% random packet loss | One 15-member-minute comparison per condition; not a production reliability guarantee. |
| r16 databasehost switchover simulation | Unanimous re-selection; about 19 ms for five simulated nodes after known death | The store is not queried. Actual detection time precedes this local selection. |
FrogTorch and Redis measurements remain independent.
Psychedelic FrogTorch’s four DDP campaign ratios are 1.95×, 1.64×, 1.03× and 1.75× Gloo elapsed time divided by FrogNet elapsed time. The compared semantics and correctness evidence belong with each campaign. A separate loopback result was unfavorable to the Python backend. Read the complete compute record.
The recorded Redis behavioral checkpoint has 1,502 OK / 207 N/A over 29 files on one core. The ARM64 parallel checkpoint reports 1,177 OK / 94 errors over eight files. Those errors remain errors. The original evidence also records a small-cache i7 control slower than stock Redis on several operations.
Redis scaling and topology-change tests ask another question. RT7 recorded stock moving 24,936 keys in 23.3 s while the Ribbit front moved none and served in 0.067 s. These timings do not mean every Redis command runs hundreds of times faster. Read the conversion record.
Earlier transport and call experiments.
The earlier Communicator operating point is bidirectional 160×120 video at 24 fps with uninterrupted prioritized audio at about 121 Kbps. Its route included the earlier broker and a 900 MHz segment. It is not HD video at 121 Kbps, and it is not a measurement of the new October 12 call.
Published semantic workload ratios range from 2.3× through 8,741× for different payload shapes; the large result involves mostly static megabyte structure with six changing fields. A 21-byte SAME frame is a protocol-level figure, not the complete IP/TCP transaction cost. The earlier echo comparison averaged 201 ms direct Apache and 118 ms hot SAME/DIFF on that recorded setup.
Boundaries that remain.
No general promise of a globally coherent heap, arbitrary cross-region transactions, automatic history reconciliation or application continuity across re-homing is established. Region permissions and recovery policy belong to the operator or vendor.
The old roughly 56-node tunnel allocator limit belongs to the broker-era implementation. Region routing uses a different hub/mapping path. The October 11 records do not characterize a maximum region size, so neither that old limit nor a designed 59,500-address replacement is a current region capacity result.
The source and records are author-produced. External reproduction is not established here. The Windows and Android operational demonstration adds hardware evidence; it does not replace the individual failure and scaling gates.