Logbook: the first four-loop rounds on the desk card
Everything the four loops asserted was written before a card was switched on. Two EditScore calls have now run. They confirmed one number, moved none, and killed one assumption, and this entry is where those measurements live now that fourloops.md is gone.
Apparatus
The desk card is 24 GB. Services are reached over HTTP and models through pixi, and a service that is not up produces a connection error rather than a low score, which is the intended behaviour: an unmet precondition is a failure, never a quiet zero.
| what | where | needed by |
|---|---|---|
| CycleGAN style transfer | localhost:8000 |
loop 3 |
| Pixal3D | localhost:8002 |
loop 4 |
| VoxHammer | localhost:8003 |
loop 4’s latent arm, which is stubbed |
pixi environment omnigen2 |
the corpus repository | loops 2, 3, 4 |
pixi environment editscore |
the corpus repository | all four |
pixi environment anny |
the corpus repository | loop 1 |
The notebooks run on the Livebook server in 7-service/service-livebook, started with
MIX_ENV=prod mix run --no-halt -e 'IO.puts ServiceLivebook.start()'which prints a URL carrying a boot token. The two heavy environments cannot be merged and the notebooks do not try: OmniGen2 pins torch 2.6.0+cu124 and EditScore pins cu128. The notebooks pin only Pillow, NumPy, Matplotlib and Requests in their own embedded interpreter and reach the models by subprocess.
A Windows service is closed. A Mix release installs one through erlsrv, whose install command requires a node name, and Livebook starts distribution itself with its own EPMD module, so a release that already started a node aborts with {:already_started, _}. That was measured rather than predicted: the first container did exactly this until the release env set RELEASE_DISTRIBUTION=none.
Measured on the desk, 2026-08-24
| what ran | weights | peak | seconds |
|---|---|---|---|
| OmniGen2 bf16, 1024x1024, 30 steps | 14.75 GiB | 17.14 GiB | 131 |
| OmniGen2 NF4, same input and seed | 4.33 GiB | 6.72 GiB | 133 |
| EditScore NF4, 512x512 | 6.7506 GiB | 28 to 36 |
Four bits bought memory and not speed: 133 s against 131 s, because dequantisation costs about what the narrower reads save on this card. The EditScore peak is the one number the run confirmed rather than corrected – the scoring script already cited 6.75 GiB, and 6.7506 GiB is that number measured.
logbook-rfd1016-model-repos.md carries the same NF4 pair against the ASUS UGen300’s 8 GB and is not repeated here.
The scale is 0 to 10, and the router assumed 0 to 1
EditScore returns each component out of ten and overall as their geometric mean, so a score lands in 0..10.
| instruction | prompt following | consistency | perceptual quality | overall |
|---|---|---|---|---|
| a matching instruction | 10.0 | 9.2 | 2.0 | 4.29 |
| nonsense, the same pair | 0.0 | 10.0 | 2.0 | 0.00 |
The components are recorded rather than only their mean, because the mean is what the two runs agree on and the components are where they differ. Consistency rising to 10.0 on the nonsense run is the tell: an edit that was not asked for was not made, so the pair is maximally consistent and overall is correctly zero. The nonsense instruction is the negative control, and a scorer that returned anything above zero for it would be certifying its own defect.
Loop four’s threshold has now been wrong twice for one reason. First as a variance against 0.15, which almost never fired; then as a standard deviation on a range ten times larger, where it fires on almost anything. A constant compared against an assumed range is the defect, not the constant, so the router becomes scale-free – spread over mean – and the threshold is set from measured view scores rather than from a third guess.
The quantisation asymmetry is a rule rather than an oversight. A quantised generator does not write corpus data; a quantised verifier may, because condition 5 is about what produces the corpus and a scorer produces a number.
Where the rest of that document went
fourloops.md also carried the loop shapes, the two-counts hazard and the holdout rules. None of that was a measurement, and all three are now stated once in a layer a gate reads: fourloops-plan.usda for the task graph and fourloops-etnf.usda for the schema, the hazards and the columns deliberately absent. scripts/check_fourloops_etnf.py reads the second against the code, which is why the numbers above are cited from here.