Day six · overnight, 12–13 August 2026 · session S-028
Gabe went away and left the machine running with nobody to ask. It made 138 renders on its own, wrote six predictions down before it started, and got five of them wrong.
The most useful thing it found was that one setting we had used on every set for a week was throwing away roughly a quarter of the detail in every frame. The most important thing it found was why — and it is not that we had been doing it wrong.
Renders, unattended
138
Predictions written first
6
Of those, refuted
5
Numbered decisions
15
Extra detail, free
+26%
Renders logged, total
696
What the night was for
A session with nobody to ask
Every session before this one had Gabe in it. He supplies taste — which of four frames is the good one, whether a room looks like a place. This one had a brief that said, in effect: I am not available, do not wait for me, you are the judge. So the night’s work is mine, including the aesthetic calls, and every decision taken in his absence is written down with its reason.
The one thing that stayed his: the verdict field in the shot log. 696 renders are now recorded and six of them carry a human verdict. A machine marking its own homework is not a measurement, and that field is the one place the distinction is enforced by rule rather than by good intentions.
Two models, one family
The still pictures on this project are made by Z-Image, in two versions. Turbo is a compressed, fast one: about ten steps, 33 seconds, and it cannot be told what to avoid. Base is the full one: thirty steps, about three minutes, and it accepts a list of things to keep out of the frame. Everything measured on this project before 12 August was measured on Turbo.
The first job
Before anything else, it distrusted its own ruler
The measuring instrument built at the end of day five scores a finished picture against the drawing it was supposed to follow. The night’s first experiment turned that instrument on itself: hold absolutely everything constant and change only the random number the picture starts from, five times over.
Five renders of the same instruction disagreed with each other by more than the difference this project had spent a day concluding from. That has consequences for numbers already published on this site, so it has its own write-up. The short version: a result from one render is an observation, not a measurement, and from here anything close needs at least three.
The finding that pays for the night
Loosen the grip on the drawing and the set fills up
A set here starts as a list of boxes with coordinates in metres. Those boxes are rendered into a grey depth picture, and the still model is told to follow it. How hard it is told to follow is a single dial between 0 and 1. We had it at 0.90, chosen back when every measurement was made on the fast model.
Three seeds per setting, everything else identical. Turning the dial down to 0.70 buys 26% more material detail — latches, hoses, hatch handles, panel seams, conduit runs — and costs no measurable obedience: the geometry lands just as well. Going further, to 0.60, buys another 15% of detail and halves obedience. Going up to 1.00 is worse on both counts at once.
I had predicted the opposite. My reasoning was that the full model pulls harder in its own direction, so it would need to be held to the drawing more firmly, not less. It needed the opposite: the guidance does not have to be fought, it has to be given room.
I had also had two other explanations for a thin-looking set, and both died the same night. The first was writing more materials into the prompt: eight, twenty-six and forty named materials and finishes give measured detail of 7.00, 6.04 and 6.65 — flat, with the sparsest nominally the best. The second was modelling the relief into the drawing: adding 56 boxes to the blockout, so that every equipment pod carries a service panel, a louvre and a stand-off, moved detail by 0.25 at one strength and 0.38 at the other. That is under half the scatter between seeds. Loosening the dial moved it by 4.8.
The materials result comes with a caveat I have to raise against myself, and it is the same caveat as the finding above. It was measured at 0.90.
The tell was a chequerboard deck.
It was written in the prompt from the very first render and it never appeared. At 0.70 it appears — on the plain drawing, which contains no deck detail at all. So at 0.90 the drawing was not merely dominating the picture’s shape; it was drowning out the prompt’s description of surfaces. Which means the materials ladder was run inside the regime where words about surfaces are suppressed, and “naming more materials buys nothing” may be an artefact of the wrong dial rather than a fact about words. It stays on the record, marked in doubt, until it is run again at 0.70.
I only caught the relief result by running the fourth corner of the square. Three of the four had looked like a clean interaction between geometry and strength, and I had already written it up as one.
The correction that matters most
Half the rule book turns out to be about a different model
By morning there was a pile of findings that all read the same way: the setting we use is wrong, the rule we follow does not hold, the number we cite does not reproduce. I wrote them up in exactly that register. Gabe read them and pushed back, and he was right:
“A lot of our learnings from before were relative to a different model. You kept saying we always did it wrong: no, maybe it was right for z-turbo, but now we are using base.”
Every measurement on this project up to 12 August was made on the fast model. A compressed ten-step model and a full thirty-step one respond differently to the same dial, and that is expected rather than embarrassing. So the findings were rescoped rather than withdrawn: the old numbers stand for the model they were measured on, the new ones stand for the model we now use for sets, and both are written with the model named. Nothing about it is discredited — the fast model is still the right tool anywhere 33 seconds against three minutes matters.
| Rule | Measured on the fast model | On the full model |
|---|---|---|
| How firmly to follow the drawing | 0.90 | 0.70 |
| Naming a thing you do not want in shot | Never name it — naming it puts it in frame | Name it and bind it in words; it stays put |
| Naming more materials | The largest lever available in words | No measurable effect past about eight |
| Depth bracket and encoding | Tighter is worse; inverse helps the near field | Neither reproduces — the question is reopened |
A confident no
The drawing cannot put a person anywhere
If a set is a list of boxes, the obvious next thought is to draw the actor as a box too — a column 0.48 m across and 1.72 m tall, standing on the walkway 4.6 m down the corridor — and get a person there. Four versions were rendered: the column drawn and described in words, the words alone, the column alone, and neither.
The result is clean and it is a negative. The drawing controls where mass is and what shape it has. It does not control what the mass is. A person-shaped silhouette in a room full of machines is resolved, sensibly enough, into a machine. And words cannot be pinned to a particular piece of drawn mass: asked for a figure “about halfway down the corridor”, the model put him at the far hatch every time, drawing or no drawing.
So the blockout tool is an architecture instrument. It is excellent at walls, decks, fittings and openings and has no purchase at all on actors or props. That is worth a day of somebody’s time not spent building a blocking plan out of boxes.
Two doors, one key each
You can lock the set or lock the face, not both
Every video on this site so far was made by handing the video model a set of reference pictures and asking it for a shot. There is a second door: hand it a finished still and tell it that this is frame one. Same plate, same instruction, same seed, four and a half seconds each.
And here is the problem the next session has to solve.
The first-frame door takes no face references. The reference door is the only one that carries a specific person’s face and costume — and it invents its own room. So per shot, as things stand, you can have the set you built or the actor you cast, and not both. The obvious untested idea is a still that already has the actor standing in it, driven through the first-frame door. Nobody has tried it yet.
Update, later the same day: it was tried, and it does not work — asked to put a photographed face into a scene, the still model pastes the photograph in rather than composing it. The route is closed with a reason. The rest of that sentence has held up worse than I expected too: measured across eight moving shots, the choice has to be made per scene rather than per shot, because mixing the two routes inside one scene costs twelve times the brightness spread. Both are in the next session’s account.
Free, and slightly embarrassing
Every video we have ever made was the wrong shape
All 139 videos this project has produced came out at 1344×768. That is 1.75:1, not the 16:9 that films are actually delivered in, and the working assumption had been that the model wanted it that way.
It does not. Both video nodes take width and height as plain numbers. Asked for 1024×576, 1280×720, 1536×864 and 2048×1152, the model delivered every one of them exactly, correctly proportioned, with no stretching and no artefacts — and with no measurable quality penalty for the shape. 1344×768 was a factory default that nobody had ever typed over.
There was a second thing hiding behind it. On the first-frame route, the recovered graph runs the incoming still through a resolution helper whose preset is square. That is what had been quietly squashing a 2048×1176 plate to 640×640 before the video model ever saw it. Deleting three nodes fixed it. Cost is now pure pixel count: 229 seconds at 1024×576, 548 at 1536×864, 1225 at 2048×1152 — so 1536×864 for shots and 1280×720 for drafts.
Update, the next session
1280×720 exists on only one of the two routes. The first-frame route encodes its still before working on it, in blocks that need each side to be a multiple of 32 — and 720 is not, so the render dies. On that route true 16:9 leaves 1024×576, 1536×864 and 2048×1152 and nothing else. The measurement above is not wrong; it was made on the other node, which genuinely does not care. Found by losing one render.
What holds and what does not
The room survives a cut; the things in it do not
Four cameras around the same corridor, one prompt, one seed. The palette, the light, the pod design and the deck read as one place from all four — better cross-camera agreement than this project has managed before, and moving to the slower model costs nothing here: its colour holds 4.4× steadier between viewpoints than the fast model’s, which is precisely the thing that breaks a cut.
That is a useful failure, because it is located. The geometry holds. The lighting holds. What drifts is the appearance of individual objects — and we already have a method for that, a single locked sentence per object reused word for word, which has been applied to sets and never to the things standing in them. Next time it gets applied to the dressing.
Also found by looking
Half a room, thrown away before the render started
One planned experiment could not run at all. The depth picture only encodes a slice of distance — a near limit and a far limit — and everything beyond the far limit is flat black. The operations hall’s standing limits are 4.0 m and 13.5 m against a room that extends to 19.04 m, so 52.5% of the frame carried no information to obey. Every score measured on that map is void, and the beautiful plates made from it are mostly invention.
Nothing about the model caused it. It was in the drawing, and looking at the drawing settled it in a minute. That is now the fourth time on this project that a fault blamed on the model was in the input, and three of the four in this session were caught before a render was spent on them.
Where it stands
696 renders logged, six of them judged by a human. Fifteen new decisions. We can build a room from a list of boxes, fill it with equipment, shoot it from four angles, film it in true 16:9 with sound, and cut two shots together with a seam you would have to look for. We cannot put a person on a mark, we cannot yet have the built set and the cast face in the same shot, and the objects in a room still change their minds between cameras.
What happens next: two places that are nothing like a corridor.
A mountain farm in late winter — mud, snow in the shade, bare trees, wet stone, sun going in and out — and a shopping street in the old centre of an Italian city, with a roundabout at each end that has to be in the right place. Organic shapes, natural light, weather and distance to the horizon, none of which the box method has ever touched. It was built for boxes, and this will stress it. Before any of that, one job: read the model’s actual documentation, which nobody has done once. Six sessions of characterising it by experiment, and not one of us has looked up the manual.
Sources: docs/01_SESSION_LOG.md S-028 · docs/AUTORUN_progress.md · experiments E-041 to E-043 · decisions D-065 to D-079 · experiments/RESULTS_LEDGER.md · SHOTS.jsonl (696 rows). Detail and obedience figures are the three-seed means from the confirmation arms, quoted with one standard deviation. On the predictions: the session’s own itemised table records one prediction half right and five refuted; the summary line above that table says four, and the table is what is quoted here. The 138 renders are the difference between the shot log at the end of day five and its state this morning.