docs/VERIFIED_RESULTS.md

Verified local operation

Current package: RC.3.5 Windows installation

The download pair is the RC.3.1 standard Safetensors checkpoint with the RC.3.5 ComfyUI integration. The guided installer prepares its own Python 3.11 worker, pinned libraries, FFmpeg/FFprobe and automatic-reference software. Models are prepared afterward through the workflow's Prepare / download models panel. Installation guide · Exact files and hashes.

RC.3.5 changes software installation and recovery. Learned weights, the numerical generation graph and the previously qualified inference runtime are unchanged. The video, audio, speed and resource measurements below retain the identities of the builds that produced them; they are not relabeled as new RC.3.5 generation results.

The installer regression suite passed 37 CPU tests, covering pinned downloads and hashes, interruption/resume, incomplete environment repair, node updates, saved UUID model profiles, reader configuration, rollback and refusal to modify a detected running target ComfyUI. These are software-installation checks, not model loading or GPU-generation tests. They do not certify every Windows installation or third-party node combination.

The exact RC.3.5 ZIP was installed through INSTALL_WINDOWS.bat on a separate, clean official ComfyUI Portable v0.40.0 installation. It downloaded and built its own managed Python 3.11 software environment, without borrowing an existing Highway worker. A repeat launch preserved the configuration. A deliberately incomplete worker with missing pip package records and an incomplete reader environment were preserved and rebuilt successfully from verified cached wheels. An attempt to update while that ComfyUI was running was refused without stopping the app.

The installed ComfyUI opened the supplied workflow and its model-preparation panel; the local scan correctly reported the missing model files, with the managed worker and reader software already detected. This acceptance used a CPU host and did not download model weights or generate a new video.

RC.3.3 installed desktop qualification — 14 October 2026

This historical desktop qualification used the RC.3.1 standard Safetensors checkpoint with the RC.3.3 ComfyUI integration. The current download pair is listed separately in Downloads. RC.3.3 changes only host terminal-socket cleanup relative to RC.3.2; model tensors, numerical graph and worker execution remain unchanged.

Four additional jobs were submitted through the actual ComfyUI Run button on the installed RC.3.3 pair: A text-only, intentional one-second timeout, cancellation during denoising, then B with four image references, Combat 0.3 and ReleaseSession. B succeeded after both intentional failures in the same host. Both completed MP4s, including audio, match the reference videos I had accepted exactly. They contain 124 frames at 24 fps at 1344×768 and stereo 32 kHz audio. All 20 evaluations of all 50 transformer blocks and 932 owned weight stores are recorded; the original transformer checkpoint was not read according to the scoped runtime guard.

The forced timeout was explicit, produced no completed media and returned in 10.87 seconds, including teardown. The first sampled observation without the cancelled worker was 12.894 seconds after the host received the UI interrupt. This is a bound for the measured case, not a universal cancellation guarantee. The final B workflow explicitly released its worker; a text-only workflow without ReleaseSession can retain a cache as configured.

Native terminal-cleanup tests passed 13/13, including the unpatched failing control, preserved unrelated errors and native IOCP traffic. Exact ZIP installation and actual host HTTP controls passed 34/34; final UI media and recovery verification passed 39/39. The qualified host is Windows CPython 3.13.12 with the checked callback digest. Other Python callbacks are not claimed repaired. See Recovery.

The complete request times for these additional sentinels were 377.3 s for A and 530.6 s for B. These are qualification observations on a shared desktop, not a controlled speedup comparison. Earlier resource measurements below retain their actual build identities.

The historical reset's exact remote connection was not recorded. The repaired callback failure is established by the discriminating controls; no claim of historical harmlessness or recovered peer identity is made. Applicable licenses are listed in License and attribution.

RC.3.2 desktop lifecycle qualification — 14 October 2026

This earlier qualification used the RC.3.1 standard Safetensors checkpoint and RC.3.2 ComfyUI integration. The current download pair is identified separately in Downloads. The numerical graph and learned tensors are unchanged. This revision changes worker request serialization, timeout, cancellation and shutdown behavior.

Four jobs were submitted with the actual ComfyUI Run button on Windows / RTX 4090: an intentional one-second worker timeout; a cancellation during denoising; a valid text-only job; and a valid job using four references with Combat at 0.3. The timeout and cancelled jobs reported failure rather than a completed video. The following text-only job succeeded in the same host. The last reference job used the exact final integration ZIP and its ReleaseSession node closed the worker after saving the result.

Both completed outputs are 1344×768, 124 frames at 24 fps, with 32 kHz stereo audio. Both whole MP4 files match the reference videos I had previously accepted byte for byte, including audio. Each valid job ran the fixed 20-evaluation program through all 50 transformer blocks. Scoped runtime guards recorded no reads of the original H3 transformer checkpoint; this is not an operating-system-wide file-access audit.

The first three jobs used the same executable files as the final ZIP. A later correction changed only attribution documents, package metadata and checksums; that delta is recorded rather than represented as a fresh rerun of all three jobs. The final ZIP was installed and used for the fourth job. Installer verification passed 26 assertions; final media, pair and cleanup verification passed 29. A separate ten-case CPU campaign used real worker subprocesses and IPC to exercise queued timeout/cancellation, close races, interrupted callbacks, recovery, Unicode paths and owned-process cleanup. These CPU cases do not substitute for the actual GPU jobs.

The worker subtree was absent at the first observation 4.753 seconds after the host received the UI interrupt. This is a sampled upper bound for this cancellation, not a universal latency guarantee. Successful generation may retain a cached session unless ReleaseSession is used; the text-only case retained one as configured, whereas the final reference workflow explicitly released it.

RC.3.2 qualification case	Full ComfyUI request	Transformer phase	Tree RSS peak	Whole-device VRAM peak	GPU mean / peak power	GPU mean / peak temperature	Sampled GPU energy
A — text-only after timeout/cancellation	379.2 s	239.6 s	30.01 GiB	6.85 GiB	265.8 / 392.2 W	62.7 / 79 °C	27.72 Wh
B — four references, Combat 0.3, explicit release	548.4 s	313.1 s	30.83 GiB	12.47 GiB	252.2 / 404.9 W	63.5 / 81 °C	38.28 Wh

Case	Worker CUDA allocated / reserved peak	Observed power window
A	3.31 / 3.78 GiB	375.40 s of 379.21 s
B	3.81 / 4.03 GiB	546.43 s of 548.39 s

These are single qualification runs on a shared desktop, not a controlled speedup benchmark. Full request time includes verification, preparation and export. Summed process RSS can count shared pages more than once; GPU figures include other applications. Energy covers the valid sampling window only. Earlier showcase and speech results below retain their original build identities and have not been relabeled as new measurements.

The historical Windows connection-reset failure has a narrowly qualified RC.3.3 cleanup fix described in Recovery. Successful media alone was not treated as proof that the original warning was harmless. Distribution conditions are described in Rights.

Model card

RC.3.1 metadata, installer and recovery qualification — 14 October 2026

The RC.3.1 pair qualified in this section uses a 20,968,864,872-byte standard Safetensors checkpoint. All 932 tensor payloads, totaling 20,967,541,480 bytes, match the prior accepted artifact. The 1,136 logical handles and graph execution order are preserved. All 23 original DaSiWa metadata fields now retain their exact string values. The matching reader permits passive string-valued provenance metadata while still checking required Highway fields, bounds, unique JSON keys, descriptors and integrity records.

Source metadata describes the inherited source. Its original title, architecture and native-resolution labels do not replace Highway's active model definition or identify a separate benchmark. This packaging change does not modify learned weights or the numerical generation policy.

Two repairs were qualified: preserving source metadata, and naming incomplete installation stages with the .disabled suffix so ComfyUI does not import them as enabled custom nodes. Tests used actual interrupted staging and restoration, missing/corrupt package and dependency inputs, and real Windows file-handle contention. The accepted default installation was preserved.

Four actual ComfyUI Run submissions exercised two versioned boundaries:

1. On RC.3, cancel active generation, confirm interruption and worker exit, then complete text-only A in the same host.
2. Rehearse exact-file rollback to the accepted APP158b pair, start its real host and discover the Highway node without neural generation, stop it, then install RC.3.1 from the final ZIP.
3. On RC.3.1, submit a profile with a deliberately missing checkpoint, observe a diagnostic and no video, then generate B with four image references and Combat 0.3 in the same host. ReleaseSession closes the worker afterward.

Both completed MP4 files match the previously accepted controls byte for byte, including audio. Each has 1344×768 pixels, 124 frames and 32 kHz stereo audio; both fully decode. B proves that tested RC.3.1 pair loads and executes 20 evaluations through all 50 blocks. Runtime counters recorded all 932 owned tensor stores, verified payload integrity and zero source-transformer checkpoint reads. These are scoped runtime guards, not an OS-wide I/O trace. The original prompt remained preserved; all four visual references were recorded.

Case and tested build	Request	Transformer	Tree RSS	Device VRAM	CUDA allocated / reserved	Power mean / peak	Temperature mean / peak	Sampled energy
A — RC.3 recovery	327.42 s	211.39 s	29.93 GiB	6.79 GiB	3.31 / 3.78 GiB	299.8 / 424.3 W	63.9 / 80 °C	27.07 Wh
B — RC.3.1 final pair	493.69 s	319.71 s	31.02 GiB	12.04 GiB	3.81 / 4.03 GiB	286.6 / 426.0 W	65.0 / 79 °C	39.08 Wh

These are recovery/qualification measurements on a shared RTX 4090 workstation, not an isolated performance comparison. Summed RSS may count shared pages repeatedly; whole-device VRAM, power, temperature and energy include other applications. CUDA reserved includes allocated. Energy integrates observed sample intervals only. Cancellation evidence is attributed to RC.3; it is not silently relabelled as another full RC.3.1 generation. The rollback rehearsal did not generate on the old pair.

The final host also logged a Windows asyncio connection-reset warning (10054) after the completed job. History success, durable output, complete-file equality, decode and owned-worker exit were independently confirmed. The precise disconnected socket was not identified, and the warning is not claimed repaired. It is retained as a separate transport observation.

Power sample coverage for A: 325.10 s of 327.42 s.

Power sample coverage for B: 490.94 s of 493.69 s.

This qualification does not assert exhaustive concurrency, timeout or multi-machine behavior. Public distribution and commercial rights remain separately controlled. Recovery instructions · Rights.

RC.3 Safetensors and installed-package qualification — 14 October 2026

The RC.3 candidate was installed from its versioned integration ZIP into a separate official ComfyUI instance. Its worker was built in a new Python 3.11 environment from pinned, hashed software wheels; it was not a clone of the development environment. The optional reference reader used its own pinned library overlay.

The checkpoint is a standard Safetensors file containing 932 individually addressable physical tensors and the saved Highway model definition, including 1,136 logical parameter handles. The official Safetensors reader opened it. Every tensor's dtype, shape and payload bytes were compared with the accepted source artifact. Only serialization, storage offsets and content-verification records changed; learned weights and numerical generation policies were preserved.

Two jobs were submitted with the actual ComfyUI Run button: text-only generation, then four image references with Combat LoRA at 0.3. Both produced 1344×768, 124-frame videos with 32 kHz stereo audio. Both complete MP4 files match the reference videos I had previously accepted byte for byte. This includes video and audio; it is not a comparison of selected frames.

Each job executed 20 evaluations through all 50 blocks. Runtime guards recorded zero reads of the source H3 transformer checkpoint and zero calls to the old model's forward path. These counters are scoped runtime evidence, not an operating-system-wide file-access trace. The reference reader exited after use. The final workflow's ReleaseSession node closed the worker, and the host remained idle with no worker subtree.

RC.3 qualification case	Full ComfyUI request	Transformer	Tree RSS peak	Whole-device VRAM peak	GPU mean / peak power	GPU mean / peak temperature	Sampled GPU energy
Text-only, first request	329.0 s	210.2 s	29.82 GiB	7.13 GiB	293.5 / 425.1 W	62.8 / 76 °C	26.67 Wh
Four references, Combat 0.3	445.0 s	290.1 s	30.80 GiB	12.11 GiB	305.5 / 434.1 W	63.7 / 79 °C	37.75 Wh

These are qualification runs on a shared Windows / RTX 4090 workstation, not a new speedup comparison. The full request includes initial file-content verification, preparation and export. Transformer time covers the numerical denoising phase. Tree RSS includes the isolated ComfyUI host and child processes and may count shared pages more than once. Device metrics include other applications. Energy is integrated only between valid observed samples. The earlier showcase and APP158b measurements below keep their original build identities.

Case	Worker CUDA allocated / reserved peak	Observed power window
A	3.31 / 3.78 GiB	325.45 s of 328.98 s
B	3.81 / 4.06 GiB	443.18 s of 445.01 s

The two-job RC.3 check does not relabel the earlier long-speech, editing or cancellation evidence as a fresh run of every feature. Full content hashes identify compatible external files after download or relocation; timestamps from the original machine are not required. Missing or altered contents remain errors. Public distribution rights and implementation disclosure remain separately controlled in Rights.

Actual ComfyUI UI — APP158b

Windows, RTX 4090, ComfyUI 0.28.0 / frontend 1.45.21. The host used Python 3.13 and the isolated worker Python 3.11.13 with PyTorch 2.10.0+cu130.

Case	Output	Full request	Transformer	Tree RSS peak	Whole-device VRAM peak	GPU mean / peak power	GPU mean / peak temperature	Sampled energy
Text-only, after cancellation	1344×768, 124 frames	290.9 s	220.7 s	30.32 GiB	7.84 GiB	333.5 / 422.7 W	66.4 / 77 °C	27.03 Wh
Four references, Combat 0.3	1344×768, 124 frames	445.3 s	290.6 s	30.92 GiB	13.02 GiB	312.5 / 433.9 W	66.7 / 80 °C	38.48 Wh
Same references, Combat off	1344×768, 124 frames	359.2 s	282.1 s	33.75 GiB	12.98 GiB	358.5 / 431.4 W	70.5 / 80 °C	35.74 Wh

All outputs include stereo audio at 24 video frames per second. The text-only file exactly matched an earlier accepted output. I reviewed and accepted both reference videos. The reference tests confirmed that all four images reached the request; the automatic name-binding stage resolved two characters, while the environment and prop still entered as visual references.

Combat was applied in the enabled case and absent in the following case. They used the same worker without another backend open. Cancellation at evaluation 5/20 closed the worker; the next request succeeded. ReleaseSession closed the worker after the final request.

These are single runs on a shared desktop. Whole-device metrics include other applications. Summed process RSS may count shared memory more than once. Warm/cold differences prevent interpreting the difference between the two reference cases as a LoRA speedup.

English speech sample — APP158b

One uninterrupted shot, one adult speaker, no visual references or LoRA. The prompt requested speech during seconds 1–13 of a 15-second clip.

I know the road ahead is uncertain, but we have come too far to turn back now. Stay close, keep your eyes open, and when the sun rises, we will find our way home together.

The actual output contains 362 frames at 1344×768, approximately 15.08 seconds, with 32 kHz stereo audio. I listened to the output and confirmed that the English was clear and the requested passage was complete. Exact speech onset and offset were not separately annotated. No replacement TTS, translation or dubbing was applied.

The complete request took 1233.3 seconds; transformer time was 1026.7 seconds for 20 evaluations. Tree RSS peaked at 33.28 GiB; whole-device VRAM at 12.77 GiB. Mean/peak board power was 397.1/455.9 W, mean/peak temperature 73.2/82 °C, and sampled energy 135.97 Wh. Worker release was verified.

This documents the accepted sample, not a general guarantee of exact wording, language or lip synchronization. These APP158b results are not relabeled as measurements of a later checkpoint format.

RC.3.4 workflow model setup — 14 October 2026

RC.3.4 adds first-use model preparation to the Asset Profile node. It keeps the RC.3.1 checkpoint and the previously qualified numerical runtime, worker and transport code unchanged.

Check	Result and scope
Install before model files are present	The exact worker environment, FFmpeg and optional reader software overlay are checked; model verification is explicitly deferred until setup.
Actual ComfyUI interface	Opened the supplied workflow, used its setup button, scanned the registered model folders, stopped verification, then prepared the files successfully.
Existing model reuse	All 21 selected files passed full SHA-256 checks, including the optional reader. Large model files were reused locally; the six-field asset profile and reader locations were saved automatically.
Saved workflow	Existing input widget positions were preserved. Credentials and generated asset-profile paths were absent from the saved workflow.
Transfer and failure handling	CPU fixtures exercised actual HTTP transfer, range resume, a server ignoring Range, interruption/resume, truncated data, wrong hashes, conflicting files, low disk space and denied access. A small pinned source file also passed download and checksum verification over HTTPS.
Configuration integrity	Failed preparation preserves the active model profile; no partially verified profile becomes active.
Local endpoint protection	Missing setup nonce, cross-origin requests and invalid actions were rejected.
Worker handoff	The existing worker accepted the automatically saved profile through its CPU doctor path, without importing Torch, loading a model or querying the GPU. Blank optional reader fields resolved to the saved reader configuration.

The two CPU regression suites passed 36 tests, followed by 18 passing aggregate checks including those suites and the installed-interface evidence. These counts describe setup and integration checks, not generation-quality tests.

No new video was generated for RC.3.4, and no multi-gigabyte model was downloaded again. Large-file transfer behavior was exercised with smaller fixtures; the full existing model contents were verified from disk. The earlier video and performance results retain their original build identities. Repository access restrictions and component licenses remain in force. First-time setup.