THOR RECORD / J-011 / 28 AUGUST 2026
Thor's arrival changed the machine—and the architecture.
J-010 ended before the target computer arrived. Since then, the Jetson AGX Thor has been brought up, the audio and camera foundations have been reproduced on the new platform, a bounded Qwen3-Omni reference has been completed, the governed Core identity gate has passed, and the current experimental speech direction has changed.
The important result is not that every part worked. It is that the evidence exposed where a monolithic omni model stopped serving the product, early enough to replace it with a modular realtime design that is still being integrated and qualified.
These photographs establish arrival and physical custody only. They do not prove boot, runtime support, performance, full duplex, or continuous perception.
What changed since J-010
The pre-arrival plan became measured Thor work.
Milestone disposition and capability evidence are kept separate. A passed milestone proves only its declared scope; a component experiment does not become an integrated product capability.
The controlled programme, traceability, evidence vocabulary, risks, and source authority are established.
Thor arrived, but untouched factory software state is permanently unknown because mutation preceded the controlled capture.
The supported NVIDIA platform is operational; independent recovery evidence remains open.
XVF3800/Pebble hardware AEC and C920 capture passed their declared standalone Thor fixtures.
A fixed-corpus Qwen3-Omni reference is retained as historical evidence, not the active conversational architecture.
Core/provider text governance passed within its declared scope, and one creation-with-lineage event was completed; no archived state was imported.
Realtime speech integration is current work. Continuous scene-state perception follows only after that loop is stable.
Target platform
What is now established on Thor.
These are local fixture facts, not vendor projections and not a production qualification.
- Compute
- Jetson AGX Thor Developer Kit · 128 GB-class unified memory · 1 TB-class NVMe
- Platform
- L4T 39.2.1 · JetPack 7.2.1-b49 · CUDA 13.2
- Configured mode
- 120 W platform baseline; this is a power mode, not a measured consumption claim
- Audio fixture
- XVF3800 USB capture · analog AUX reference/output · Creative Pebble V3
- AEC result
- About 485× peak separation between the speaker-only and double-talk LEFT-channel fixtures
- Camera fixture
- Logitech C920 · 1280×720 MJPEG · approximately 30 FPS · standalone capture
The useful failure
Qwen3-Omni produced a conditional reference—and lost the architecture.
The fixed NVFP4 corpus exercised text, audio, image, video, synchronized audiovisual input, native speech presence, and 50 English/Spanish controls. That bounded result is real and remains useful.
It did not answer the realtime product question. The official BF16 path exceeded the practical memory envelope in the tested runtime, while the native speech path later showed reproducible digital amplitude collapse before playback. Runtime-specific cleanup and cancellation complexity kept accumulating around a provider that still did not deliver the required conversational path.
What did not change
The Core—not the model—still owns Evopien.
MS-06 completed the provider boundary and created the Thor-generation identity exactly once with declared engineering lineage. Archived conversations, operational memory, relationships, and identity state were not imported.
Identity, permissions, relationship separation, governed memory proposals, audit, and cancellation remain Core responsibilities. The model sees filtered context and cannot grant itself database, network, filesystem, credential, or actuator authority.
Foundation v1.2 remains controlling. Foundation v2.0.1 and Alpha v0.3 remain drafts; the current voice-and-vision-first engineering sequence does not silently amend the approved Alpha scope.The modular pivot
One realtime path, several replaceable specialists.
Each row below has a bounded role. Several have separate Thor experiments; the chain as a whole has not yet passed an end-to-end duplex gate.
- 01Audio boundaryXVF3800 + Pebble hardware AECThor fixture
MS-03 standalone fixture passed. LEFT supports near-end reflex evidence; RIGHT carries the recognition feed.
- 02Streaming recognitionNemotron 3.5 Streaming ASR 0.6BExperimental
Two experimental operating points are retained: a faster realtime candidate and a slower accuracy candidate. Real-room bilingual integration remains open.
- 03Turn intelligenceSilero + DualTurn 240 ms + contextual UltraVADExperimental
Fast speech evidence, conversational endpointing, and semantic context advise the deterministic Turn Manager; no classifier owns the floor alone. UltraVAD remains contextual-only and its production licence status is unresolved.
- 04AuthorityEvopien Turn Manager + CoreCore-owned
One generation identifier must own request state, speculation, cancellation, memory proposals, synthesis, PCM, and playback.
- 05CognitionCommunity-produced Huihui Qwen3.8-27B abliterated NVFP4 + SGLangExperimental
One small English control and one Spain-Spanish-labelled no-thinking control passed on Thor. Reasoning, tools, image, video, and broader language quality remain unqualified.
- 06Speech generationMagpieTTS v2607 + NanoCodecExperimental
English and Spanish generation ran; Spanish physical playback and digital cancellation/recovery slices also ran. Castilian naturalness remains open, and the path depends on an unmerged vendor PR plus a local experimental cancellation patch.
- 07Output boundaryGeneration-owned PipeWire playbackOpen gate
The digital player can be cancelled quickly, but microphone-observed last-sample silence has not been measured.
Non-negotiable language gate
English and Spain Spanish pass together—or the baseline does not pass.
Spanish execution exists in the component experiments. That is not the same as accepted Castilian speech quality, matched endpointing, matched interruption behaviour, or end-to-end parity.
- English
- Component references exist; integrated overlap and recovery corpus pending
- Spain Spanish
- Component references exist; Castilian naturalness and paired parity gates pending
Evidence boundary
What this record still does not claim.
The component chain is not integrated and the paired bilingual acceptance corpus has not passed.
Digital cancellation measurements do not establish speaker-observed acoustic silence.
C920 capture is a standalone fixture. Current/stale scene state and the Qwen3.8 VLM path remain MS-09 work.
The Core-governed typed gateway remains the design; the replacement stack has no qualified integrated research or tool path.
The modular architecture checkpoint is experimental and is not yet a merged, green production baseline.
Real-user privacy, recovery, reliability, actuation safety, independent isolation, and deployment evidence remain future gates.