pass record

pass-02

status: complete
evidence: verified by manifest
frames: 43,200 image, 43,200 text
config hash: 7859a7f006e2
runtime: 30:00 at 24 fps
gallery: image channel · text channel · both channels · 2,160 sampled stills
prompt readout: every prompt, frame by frame · 30 minutes at datacenter-clean-v3

frame 1 · 0:0021,600 · 15:0043,200 · 30:00

pass-021–43,200

664ba85279001–2,651

542b74c6887f2,652–10,636

826a790f994e10,637–43,200

verified by manifestobserved in the logno recordrecovered from staging

what it looks like · 17 frames across the pass

Open any of them here. These are sampled stills, one every 24 frames — the wall has all 1,800.

Channels

channel	files	bytes	covers	contiguous	how it was counted
image	43,200	58.3 GB	1–43,200	yes	rebuilt by hashing the published channel on the share; the rendering runs' checkpoint rows were lost with their processes
text	43,200	1.5 GB	1–43,200	yes	rebuilt by hashing the published channel on the share; the rendering runs' checkpoint rows were lost with their processes

Two independent stacks: separate sequences, directories, hashes, manifests and video files. Never composited, never re-timed. Rendered by one driver iteration, which is what frame-locks them.

Runs

3 rendering runs. On the axis above, the square tick is where each run stopped and the hatched segment beside it is the frames recovered from the renderer's local staging afterwards. Those segments are tens of frames out of 43,200, so they are drawn at a minimum width rather than to scale; the exact ranges are below.

664ba8527900

1–2,651

outcome: failed
frames: 1–2,651 of 1–43,200 requested
started: 2026-09-09 03:04-07:00
finished: 2026-09-09 07:14-07:00
cause: driver exited 1073807364 (DBG_TERMINATE_PROCESS) at frame 2,651 - Windows Update restarted the machine at 07:14
reached the share by checkpoint: 2,640
recovered from staging: 2,641–2,651 (11 frames)
driver, reconstructed: 30bee52fad0f
driver, declared by recipe: 5047c0c726f0feb3e3dd2393763e9d368319785c

f001321 a frame this run rendered

Not observed: this run predates per-run driver hashing. The driver at the time was commit 5047c0c; the reconstructed value is that commit's content with CRLF line endings, which is how it sat on the share. It is a reconstruction, not a measurement, and is recorded separately from driver_sha256 for that reason.

542b74c6887f

2,652–10,636

outcome: failed
frames: 2,652–10,636 of 2,652–43,200 requested
started: 2026-09-09 08:59-07:00
finished: 2026-09-09 21:35-07:00
cause: WinError 10055 (WSAENOBUFS) at frame 10,636 - the driver opened a socket per ComfyUI API call, ~23 per frame, polling /history every 0.25s
reached the share by checkpoint: 10,571
recovered from staging: 10,572–10,636 (65 frames)
driver, reconstructed: 30bee52fad0f
driver, declared by recipe: 5047c0c726f0feb3e3dd2393763e9d368319785c

f006649 a frame this run rendered

Same pre-fix driver as the previous run. Not observed; reconstructed as above.

826a790f994e

10,637–43,200

outcome: completed
frames: 10,637–43,200 of 10,637–43,200 requested
started: 2026-09-10 15:48-07:00
finished: 2026-09-12 21:30-07:00
checkpoints: 271
seconds per frame: 5.936
driver, hashed at launch: 4f3406d1abea
driver, declared by recipe: 5047c0c726f0feb3e3dd2393763e9d368319785c
interpreter: %USERPROFILE%/anaconda3/python.exe

f026929 a frame this run rendered

MEASURED on 2026-09-13 from the driver still on the share, which is byte-identical to comfy-batcher commit 8853843 (the websocket + keep-alive fix) as checked out with CRLF. The recipe declares 5047c0c; this run did not use it. 32,564 of the pass's 43,200 frames were rendered by 8853843.

Ran after the driver moved to a ComfyUI websocket plus one keep-alive HTTP connection, which removed the socket churn that ended the previous run.

Driver

the driver changed mid-pass

THE DRIVER CHANGED MID-PASS: 2 distinct driver files rendered this pass — 1-2,651 by 30bee52fad0f (reconstructed), 2,652-10,636 by 30bee52fad0f (reconstructed), 10,637-43,200 by 4f3406d1abea. The recipe hash is unaffected by design (the driver is deliberately outside CONFIG_HASH_INPUTS so a harness fix does not split a pass's recipe identity), which is exactly why this has to be stated here instead.

frames	task	sha256	provenance
1–2,651	664ba8527900	`30bee52fad0f`	reconstructed
2,652–10,636	542b74c6887f	`30bee52fad0f`	reconstructed
10,637–43,200	826a790f994e	`4f3406d1abea`	measured

observed_sha256 is the driver file each run actually launched, hashed at launch. declared_commit is what the pinned recipe SAYS the driver is — a declaration, in a different identifier space, which nothing verifies. When they disagree, or when runs disagree with each other, the observed hashes are the fact.

Render log

frames logged: 43,158 covering 2–43,159
denoise: 0.35 – 0.72
luminance: 28.1 – 167.6 of 255
chain breaks: 0
gpu logged: 70 h 22 m
mean per frame: 5.870 s

withheld

Every frame's sampler seed is in this log and is NOT exported. It is a complete reproduction recipe; available on request.

Frames per era

era	frames
front	8,640
aerial	8,639
side	5,760
aerial-to-front	4,320
front-to-side	4,320
side-to-3q	4,320
3q-to-aerial	4,284
3q	2,875

How this pass was assembled

Assembled from three rendering runs and two staging rescues. Frames 1-2,640 and 2,652-10,571 reached the share by checkpoint; 2,641-2,651 and 10,572-10,636 were recovered from the renderer's local staging after each crash, decoded first to prove they were not torn, and published through the same verify-on-read path. A fourth task (29f7fbddc0ff) never rendered at all: it was declined before the lease because 'python' resolved to Gaffer's own venv, which deliberately has no PIL. These manifests and this pass.json were REBUILT on 2026-09-13 by hashing the published channels on the share, because the original run wrote a manifest covering only its final 44 frames (43,157-43,200) and the per-checkpoint rows were lost with the process that held them. The frames themselves were never at risk; only the record was.

← pass-01 all passes pass-03-rehearsal →