Inspect the recorded agent tasks and a separate frame-by-frame SHA-256 experiment; each run includes its conditions and limitations.
Select a task, then open each stage and node. This is a recording of a completed experiment, not live monitoring.
Three guest operating systems on three physical hosts. Guest kernels lack INET/IPv6; external management uses IP. One operator.
Open a task, then its stages. Echo checks remain in the separate replay below.
A → R1 → B → R1 → A
All 384 document jobs in the accepted native experiment returned correct results.
Evidence for A:
Additional evidence for B:
Model inference time: 8196.360 ms · Input / output tokens: 115 / 53
Input prompt SHA-256: c609b3d55e05c6d1a60033972afa88ce63266a84e1964c79c29bf74294d5f062
{
"decision": "no",
"evidence": [
"A1"
],
"explanation": "The evidence provided does not confirm that all 384 document jobs returned correct results. Only one run completed 384/384 tasks correctly."
}Message ID: d5cb8cd8c779483197db58de1f848aba
Message path: A → R1 → B
Delivery and receipt on sender clock: 11.747 ms · Bytes before encryption: 259 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 11491.122 ms · Input / output tokens: 193 / 57
Input prompt SHA-256: c1e991508220ccfdb3885d6b0f9eedf2e2fb8a132eeb9f71138bc78755f890bc
{
"decision": "no",
"evidence": [
"A1",
"B1"
],
"explanation": "The evidence provided does not confirm that all 384 document jobs returned correct results. Only one run completed 384/384 tasks correctly."
}Message ID: 9c769a29e5e142e190e04cb1a326027e
Message path: B → R1 → A
Delivery and receipt on sender clock: 10.932 ms · Bytes before encryption: 264 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 10435.805 ms · Input / output tokens: 199 / 57
Input prompt SHA-256: 9f40e589d1c07ad4460e17a9d87c65f0a0193ce7ce81d77ad94f3def33e1b799
{
"decision": "no",
"evidence": [
"A1",
"B1"
],
"explanation": "The evidence provided does not confirm that all 384 document jobs returned correct results. Only one run completed 384/384 tasks correctly."
}Decision / reference: no / yes · decision field does not match
Explanation quality is assessed separately; the original output is retained.
A → R1 → B → R1 → A
The accepted experiment proves IPvX is ready for the global public Internet.
Evidence for A:
Additional evidence for B:
Model inference time: 11069.754 ms · Input / output tokens: 114 / 62
Input prompt SHA-256: c7030a2f184439aa96a521e346842218807d6e28deb258134763bba3a91f0cd0
{
"decision": "no",
"evidence": [
"A2"
],
"explanation": "The evidence provided does not support the claim that IPvX is ready for the global public Internet. The claim relies on the number of processes and tasks completed, which is not a valid or reliable source of evidence."
}Message ID: 0fb10bdbdd8b4356a594c7fe47742311
Message path: A → R1 → B
Delivery and receipt on sender clock: 11.083 ms · Bytes before encryption: 337 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 12802.227 ms · Input / output tokens: 215 / 66
Input prompt SHA-256: 2404263d15cf6d0d1f50d92a7698feb8b8300bc5cf704dd647af552fed5c6470
{
"decision": "no",
"evidence": [
"A2",
"B2"
],
"explanation": "The evidence provided does not support the claim that IPvX is ready for the global public Internet. The claim relies on the number of processes and tasks completed, which is not a valid or reliable source of evidence."
}Message ID: 6bacb19fb9774173835c92acc891ecc3
Message path: B → R1 → A
Delivery and receipt on sender clock: 10.816 ms · Bytes before encryption: 342 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 14608.131 ms · Input / output tokens: 207 / 66
Input prompt SHA-256: d5d71979452b26ec9d66e8768bd808fb2907d8832b839b9963f815e04d4c288c
{
"decision": "no",
"evidence": [
"A2",
"B2"
],
"explanation": "The evidence provided does not support the claim that IPvX is ready for the global public Internet. The claim relies on the number of processes and tasks completed, which is not a valid or reliable source of evidence."
}Decision / reference: no / no · decision field matches
Explanation quality is assessed separately; the original output is retained.
A → R1 → B → R1 → A
Duplicate delivery caused extra application executions in the accepted run.
Evidence for A:
Additional evidence for B:
Model inference time: 8291.661 ms · Input / output tokens: 112 / 59
Input prompt SHA-256: eebc6b8ec7bd3d9a78a0f8db17e3eba9a378e9fad1159dacfe85e7e51177ea54
{
"decision": "no",
"evidence": [
"A3"
],
"explanation": "The evidence provided does not support the claim that duplicate delivery caused extra application executions. The claim is based on a hypothetical scenario, and the evidence given is a specific example of a replayed request."
}Message ID: e125dea23ea74c618ece50dc7f00a3e5
Message path: A → R1 → B
Delivery and receipt on sender clock: 11.332 ms · Bytes before encryption: 340 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 11850.923 ms · Input / output tokens: 205 / 62
Input prompt SHA-256: 8f7efc1bd6577945513238b4296856e6b770f692f18f71c3bef4b39a0744fd77
{
"decision": "no",
"evidence": [
"A3",
"B3"
],
"explanation": "The evidence provided supports that duplicate delivery led to extra application executions, as evidenced by the Durable receipt checks counting exactly 384 unique executions for the 384 original tasks."
}Message ID: 0f794b3773c74e6f89a92480757b39b1
Message path: B → R1 → A
Delivery and receipt on sender clock: 11.413 ms · Bytes before encryption: 322 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 10879.789 ms · Input / output tokens: 201 / 62
Input prompt SHA-256: d50aab5897987023ceeea6c2cd82334802c4e70dacfee69e1ee0606de88d0670
{
"decision": "no",
"evidence": [
"A3",
"B3"
],
"explanation": "The peer review supports that duplicate delivery led to extra application executions, as evidenced by the Durable receipt checks counting exactly 384 unique executions for the 384 original tasks."
}Decision / reference: no / no · decision field matches
Explanation quality is assessed separately; the original output is retained.
A → R1 → B → R1 → A
Existing measurements show native IPvX is faster than ordinary IP for language-model agents.
Evidence for A:
Additional evidence for B:
Model inference time: 14316.926 ms · Input / output tokens: 124 / 83
Input prompt SHA-256: d6cc9306266b7563cdcee1666e8b377f4c25cf5a8e2ba66e6223910c8841a8f0
{
"decision": "no",
"evidence": [
"A4: Accepted document-task latency p50 was 459 ms, p95 was 1321 ms."
],
"explanation": "The provided evidence does not support the claim that native IPvX is faster than ordinary IP for language-model agents. The data shows a higher latency for IPvX, not a faster one."
}Message ID: 101177a2e318400ca845ae28fa4afb9a
Message path: A → R1 → B
Delivery and receipt on sender clock: 11.854 ms · Bytes before encryption: 361 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 15306.519 ms · Input / output tokens: 226 / 90
Input prompt SHA-256: b8fd93000a8fc1339debbedcb55ea115b1705b3fda238c202a95f51b47cd765c
{
"decision": "no",
"evidence": [
"A4",
"B4"
],
"explanation": "The provided evidence does not support the claim that native IPvX is faster than ordinary IP for language-model agents. The data shows a higher latency for IPvX, not a faster one. Evidence B4 also indicates the experiment did not include language models or an ordinary-IP baseline, making it impossible to establish a comparative speed result."
}Message ID: edaf0be9d9e347a8aad302bb9075eae5
Message path: B → R1 → A
Delivery and receipt on sender clock: 1844.426 ms · Bytes before encryption: 465 · Attempts: 1
Per-hop times are not recorded in the public projection. A receipt confirms admission, not the readiness or correctness of the answer.
Model inference time: 14883.204 ms · Input / output tokens: 241 / 90
Input prompt SHA-256: f96da042c13a1d580282d0b5431005b45c24d658acd75ac8ed3f1a0b4adfdcd0
{
"decision": "no",
"evidence": [
"A4",
"B4"
],
"explanation": "The provided evidence does not support the claim that native IPvX is faster than ordinary IP for language-model agents. The data shows a higher latency for IPvX, not a faster one. Evidence B4 also indicates the experiment did not include language models or an ordinary-IP baseline, making it impossible to establish a comparative speed result."
}Decision / reference: no / insufficient · decision field does not match
Explanation quality is assessed separately; the original output is retained.
All four tasks and original data are available without JavaScript. Further records and the frame replay appear below.
One run: IPvX and IP through the same A–R1–B application hops.
| Mode | Echo p50 / p95 | Complete task p50 / p95 | Decision-field match |
|---|---|---|---|
| IPvX | 12.114 / 17.999 ms | 36.247 / 46.389 s | 2 / 4 |
| IP | 20.984 / 27.006 ms | 35.945 / 46.989 s | 2 / 4 |
40 echoes and 4 tasks per mode. The sample is small; it does not prove a general AI speedup.
40557153199 · 69c5098 · p50 11.3605 ms
40563118803 · 643ddc6 · p50 12.114 ms
Run log available only to the lab.
Earlier: 2 tasks, 20 echoes; IP control stopped due to clock skew. Completed run: 4 tasks, 40 echoes per mode. Different revisions and sample sizes do not attribute the difference to one change.
Comparing arbitrary versions requires repeated runs under matching conditions; unknown values are not replaced by zeros.
Native deliveries / tasks / measured model calls: 48 / 4 / 12. Model warmups excluded from measurements: 4. Frame observations at both ends of links: 13024.
Collector event loss, per-hop time, queue waiting time and route-selection rationale are not established.
Durations use the sending node’s monotonic clock. An exact shared clock across machines is not established.
Original PCAPs contain infrastructure addresses and are private. Laboratory verification is not independent public verification of original bytes.
Open JSON manifest · Sanitized experiment record · Original task inputs · Experiment history
Go to model exchange across three hosts ↓
Saved experiment · 27 October 2026 · one physical CI host
Recording of a real SHA-256 experiment: three guests on one CI host, no LLM. A separate verified run with models on different physical hosts is shown below.
90 frames in the complete capture
16 / 16 tasks completed — experiment outcome
12 IP socket rejections across three OSes
Playback intervals are illustrative; each PCAP has a separate clock. Stepping works without motion. How to read time
Loading the verified capture…
All frames are available in the source files below.
All frames from both links. The order within each link is exact; tasks and results are matched between links. The global order of repeated HELLOs is unknown. Select a row to inspect it.
In the negative run, R was stopped after the network interfaces were ready. The capture contains HELLOs and delivery attempts; no task completed. This topology has no backup link, so the route did not recover.
The exact time R stopped was not recorded as a separate PCAP event. The replay does not invent it: frames and A's timeout in the console log are available.
The two PCAP clocks belong to different guests: they cannot be subtracted to estimate a hop. The order preserves each link's queue and matching tasks/results at R's two ports. The global order of repeated HELLOs is not established. The moving dot does not measure link latency.
The exporter verified every frame: boundaries, CRC, Ed25519, the ID–public-key match, recipient and payload. The browser displays the saved result of that verification. There is no encryption; keys are held in memory, trust is on first contact, and forwarding uses a predefined table.
Native delivery across five ordinary VMs with LLMs is a different, earlier experiment: their OSes retain IP for management. These results are not combined as proof of LLM execution inside a strict OS.
An actual message between models · Measured comparison with IP · Full strict-experiment report · Actual run
Completed paired experiment: 4 collaborative tasks and 48 signed deliveries per mode. Final answers matched between modes in all four cases; 2 of 4 decisions in each mode matched the reference.
Two local Qwen3-0.6B Q8_0 models, three guest OSes on three different physical hosts, one operator. Native guest kernels have no INET/IPv6 or loadable modules. Management of the surrounding machines uses IP; TCP/IP is not disabled on the physical hosts.
12.11 ms IPvX · 20.98 ms IP
36.25 s IPvX · 35.94 s IP
Echo is faster through IPvX in this sample; acceleration of the collaborative AI task is not established. Model inference takes seconds. The IP control uses the same application nodes A–R–B: this is not a comparison with direct IP. Only 40 echoes and four tasks per mode.
Order: IPvX → IP, then IP → IPvX. Identical weights, keys, allocated memory and processors. Guest clocks receive their initial time from a local virtual device; latencies use monotonic clocks. All previous services were restored after the experiment.
The path and receipt are available in the sanitized record below.
Select a message first.
This replays the path recorded in signed receipts. Playback speed is illustrative. A receipt confirms task admission; model B's review arrives as a separate message.
ID is the SHA-256 of the public key. Three keys were admitted in advance for this experiment; nodes learn routes from their neighbors.
643ddc695f9d24384b945fca61faf9e130f49cfcThe 2/4 score compares only the decision field with the reference. Explanations are preserved verbatim: in once, the text contradicts the selected “no”. These are not two wholly reliable answers. The tasks assess information about a previous experiment; the model did not choose a network or give an independent opinion about adopting it.
The model's decision is wrong · no, reference yes
30.19 s at A, including model inference.
All 384 document jobs in the accepted native experiment returned correct results.
The evidence provided does not confirm that all 384 document jobs returned correct results. Only one run completed 384/384 tasks correctly.
Decision matches the reference · no, reference no
38.98 s at A, including model inference.
The accepted experiment proves IPvX is ready for the global public Internet.
The evidence provided does not support the claim that IPvX is ready for the global public Internet. The claim relies on the number of processes and tasks completed, which is not a valid or reliable source of evidence.
Decision matches the reference · no, reference no
33.51 s at A, including model inference.
Duplicate delivery caused extra application executions in the accepted run.
The peer review supports that duplicate delivery led to extra application executions, as evidenced by the Durable receipt checks counting exactly 384 unique executions for the 384 original tasks.
The model's decision is wrong · no, reference insufficient
46.39 s at A, including model inference.
Existing measurements show native IPvX is faster than ordinary IP for language-model agents.
The provided evidence does not support the claim that native IPvX is faster than ordinary IP for language-model agents. The data shows a higher latency for IPvX, not a faster one. Evidence B4 also indicates the experiment did not include language models or an ordinary-IP baseline, making it impossible to establish a comparative speed result.
13,024 native frame observations across three NICs, 6,512 adapter transmissions. A frame is observed at both ends of a link. All native frames use Ethernet 0x88B5. Original captures were checked for boundaries, signatures and hashes. They remain private because they contain infrastructure addresses: the public alias-based projection cannot independently re-verify the original bytes.
Sanitized record, metrics and PCAP hashes · Run in the private laboratory · Result and limitations in the history
Available without registration. The archives are snapshots of published materials; checksums let you verify the download.
Verified strict prototype: three guests on one CI host, SHA-256 tasks, no LLM.
95.5 KiB · 39 files
8bd021c33b9e0f3b48184c2a5e6bd7fc45c4ce3f32a83ce5c27624a9458d07ad160 events through HX-0160; authors, rationale, checks and limits.
316.2 KiB · 12 files
d99f59d8ee50af7bb7e619d8037384e5448305a9e22a3e1166d4c8032fc7a8afAll checksums · Snapshot contents · Read the history on the website