Front page: the tutorial. The full model, with every plate and page, is MODEL.md in the study folder; this is the part a pixel painter uses.
Kazuo Oga, 男鹿和雄画集Ⅱ (Ghibli THE ART series, 2005). 148 scanned pages in Japanese:
The scan's image number is the printed folio plus two. Write that at the top of your notes.

nb157. The front matter as a contact sheet. Make these first, twenty pages to a sheet, to see the book's shape before reading.
It never states its model as a list. Each principle comes with one painting, often the colour board beside the final background, and Oga or Tanaka says what changed and why. The plate is the teaching. I read every page as half-page crops at about 1.15×, so I could read the Japanese and look at the painting in the same glance:
# work/crop.py IDX part [scale] part: t,b (halves), tl,tr,bl,br, or x0,y0,x1,y1 fractions
P={'t':(0,0,1,.52),'b':(0,.48,1,1),'tl':(0,0,.52,.52),'tr':(.48,0,1,.52),'bl':(0,.48,.52,1),'br':(.48,.48,1,1)}
f=P[part] if part in P else tuple(map(float,part.split(',')))
c=im.crop((int(f[0]*W),int(f[1]*H),int(f[2]*W),int(f[3]*H)))

nb030. p.44, "unmuddied mixing" (濁らない混色), C-0081: the greens that bridge light and shadow.
Each principle has the book's term, the page, and the rule I took from it. The rules in bold are the ones that carried into pixels and have a how-to.
| Principle | Book | Rule |
|---|---|---|
| Ground stage (地塗り) | pp.46–49, cover steps 5–8 | Settle light direction, overall colour and atmosphere at the lowest level of detail, "as close to finished as possible". Detail is a thin last layer that must not break it. |
| Grasp the crowd (群れを捉える) | p.48, C-1646 | A mass of like things is one organised impression with a few extracted members, never member by member. "Drawing everything crushes the space." |
| Coloured greys | pp.48, 57, 103 (C39-5) | Neutrals are always tinted: brown-, green-, purple-, red-, yellow-grey. In low light the greys carry the whole range of hue. |
| Unmuddied mixing | p.44 (C-0081), p.45 (C-0804) | A big value range is held by saturated mid-tones; accents laid in unmixed; many half-mixed colours side by side. |
| Living brushwork | pp.50–51 | "Know beforehand what shape each brush motion makes." Moss is not drawn leaf by leaf. In light, a few lit marks stand for the unpainted mass (C-0822). |
| Space and planes (抜け, 面) | pp.52–55 | "Intentionally severing connected space gives depth" (C-0064). The ground is one plane with undulations; near grass seen from above, far from the side (C-0095). |
| The light source | pp.56–57, 94–97 | "Don't spread the light around; concentrate it." Shadow about 2/3 (C-0792). Reflected light takes the sky's colour, loosened with the object's own. The light's colour at its edge (際, C-0772). Snow: two shadow families (p.96). |
| Density at the points | pp.58–59 | "Concentrate density at the points and the whole feels dense." A cut's length decides how much it gets. |
| Omission | pp.60–61 | Keep form, colour and light; drop members. Shadow contrast softens with distance (C-0007 board vs background). |
| Air | pp.98–99 | Far things approach the sky colour; dry air gives crisp distance, humid air narrows the range. |
| Seasonal note | pp.104–108 | One seasonal flower works like a kigo (season word): a minimum of objects conveys the season. |

nb053. p.56: C-0772 (the sun's colour) above C-0792 (the gorge, shadow about two-thirds). These two plates became how-to 4.

nb045. p.52, "awareness of space and planes": C-0064 at the foot of the page. A ridge line and a band of tree shadow sever the near meadow from the far village.
This is the most practical part of the book, because it's a procedure:

nb014. Step 7, "whole colour": the splayed flat brush on the shrubs.

nb020. The finished cover on the desk.
I measured his plates in OKLab and found his darkest 2% at L 0.35, against 0.20 for photos. Oga lifts his black point, I thought. The printed text on the same pages measured L 0.29–0.35: it was the printing and the scan. Correct each plate to its own page's ink and paper first (work/plates.py):
k=np.median(a[o[:max(50,len(o)//1000)]],0); w=np.median(a[o[-len(o)//100:]],0) # page black, page white
out=np.clip((im-k)/(w-k),0,1)
After correction his black point is like a photo's. And whole-picture statistics (shadow share, grey share, tint count) barely separate his plates from photos of the same subjects at all. What makes a picture his is where things go. Measure passages, not pages.

fig. What does survive a global look: his paintings, area-reduced to 160 px, still read almost completely. The ground-stage claim, seen from the other side.
Next: copying a passage, which is where the model starts to mean anything.