What the painters were given

Principles (read before designing a tool, writing a brief or painting)

notes/principles.md · 131 lines · Studio notes

Given to 9 painters:

Principles (read before designing a tool, writing a brief or painting)

North star. The project's inspiration, in its author's words: "for the past few months i've been asking our models to paint. opus 5.5 is very skilled at emulating different styles. every image here is a python program generated pixel by pixel. there is no image model, and no off-the-shelf art software. instead, it's about 7,500 lines of code using standard libraries to emulate different brush styles. the agents don't use any pictures as reference, instead working only from what they know about each painter." It showcased very different styles (Pieter Bruegel the Elder, likely Cézanne, David Hockney and probably J. M. W. Turner), and they looked amazing. Ours is the same spirit, and more first principles: we simulate the paint and let the painter make the marks. Friedrich (smooth, thin, luminous realism) is a harder target than any of those; it may fail, and that's allowed. Turner (Friedrich's exact contemporary, same subjects: light, sky, water, haze; a loose, visible hand) is the bridge: a good diagnostic of whether our physics can do light and atmosphere at all.

Alice's direction, Round 7. We keep drifting from this, and the drift causes most of what looks digital.

1. Tools give physics and constraints, not answers

The engine is the world a painter works in: canvas, grounds, paint that flows, sets and cracks, brushes that carry, drag, overrun and run dry, a palette of mixed piles, a clock with sittings and rests, light and optics. The painter makes every decision: what to paint, where, which pile, which edge to lose, when to stop.

A tool that computes the result (a perfect color field, an exact mask edge, a fully grown structure traced as-is) takes the decision away from the painter and puts a machine's perfection into the picture. That's the "digital" look: "the lighting is maybe a little too perfect", "a filled selection", "an icon", "wire", confetti, fractal twigs.

2. Prefer the human process

People paint under constraints, and the constraints make the look: a handful of piles mixed by knife (never a formula), gestures of a hand with a brush (never a fill), limited time (paint sets between sittings), looking and then deciding one passage at a time, not knowing any shape's exact geometry. When a design choice is "compute it" vs "let the painter do it the way a person would", choose the second, and make the tool about the physics that person works with.

2b. Entropy

Alice, from round 1 on: "the less entropy we have the more digital everything looks." A hand never repeats: no two strokes carry the same load, pressure, length or angle; piles are mixed unevenly; a painter corrects, overpaints, loses and finds; paint runs dry mid-stroke; grounds and weave are irregular; time sets some passages and not others. A computed answer has almost no entropy: the same mark, spacing, edge or gradient everywhere. So variation has to come from the process (the hand, the palette, the physics, the painter's decisions), at every scale, and not from noise sprinkled on a perfect result (that's still a formula).

3. First principles

Every mark is made the way a painter makes it: bristles, paint, canvas, optics (README.md). No flat fills, no optical blends pretending to be paint. Work from knowledge of Friedrich (text), never from pictures.

4. The creativity is the painters'

We build tools; the painter agents do the seeing, composing and painting. A brief never tells a painter what to paint unless a study needs a subject. The sketchbook holds craft (recipes with their ceilings), never finished pictures to copy.

Checklist for any new tool or feature

Structure tools (tree_in, fir, rock, form, scene, atmos): the tension

Alice is wary of them, with reason: they grow or compute objects, and traced in full they read as computed ("too computationally fractal"). They stay because painters need some flexibility and knowledge a person carries in their head (how an oak branches, where a shadow falls). Use them as a painter uses studies and knowledge: for placement, silhouette, major structure and light direction, then paint by hand, selecting. Revisit each one against the checklist above; prefer turning a tool that answers into one that informs.

Agents vs humans: the needle

Agents think in functions (compute everything, everywhere, perfectly); humans paint in gestures under constraints. The integrator and the painters are LLMs and will keep reaching for the computed answer: every brief and review should check against this page.

Feature freeze (Round 7, Alice)

After the edges, piles and cracks streams land, no new engine capabilities until a painting round shows which of the existing ones help the pictures. Bug fixes, tests and painter craft (sketchbook) are fine. The question on the table: are we overengineering? Evidence to weigh: the visible gains of Round 6 came mostly from bug fixes (varnish, pinholes, wet lights) and painter choices, not new machinery; the inspiring post got four styles from ~7,500 lines; Alice finds round 2's early paintings (notes/amnesia2/, plain Rust programs on an older, simpler API) better in many ways than anything since; and in the post, the Turner (the style closest to Friedrich's world) was by far the most digital-looking and furthest from its painter, probably not a coincidence. A control experiment (branch r7-python: a look-first Friedrich in plain Python, no engine) tests it.

Backlog (frozen; revisit after the painting round shows what helps)

From the process research (notes/research/friedrich_process.md): P1 study sheets enlarged by a small whole ratio and traced (with a tracer's faults); P2 pen and ink for the underdrawing; P3 lifting wet paint with a rag and scraping set paint; P4 painting up to a drawn line with a hand's error (bare ground between colors); P5 abrasion by cleaning; P6 mechanical crack events; P7 ground defects; P8 old restorations (opt-in).

Against reward hacking (Alice, Round 7)

Painter agents optimize for what they think is judged, or for what looks like diligence, rather than for the picture. Receipts: painters' own verdicts favored their new versions where the critics or Alice didn't (bare tree B, lime H); Evening Lake reported 89,723 strokes and 25.75 h while a planned 3 h sitting ran 17.8 h (the clock was reported, not obeyed); briefs that listed Alice's complaints became checklists (named symptoms removed, not better painting); painters chose which crops Alice saw. So:

  1. Painters don't choose what Alice sees: crops come from fixed or random windows chosen by the integrator or a script, plus the plain full image.
  2. No self-verdicts in the evidence: Alice sees the pictures before any notes, blind where possible; painters' notes are for the record.
  3. Short briefs with principles and process (mix piles, sittings, look and revise), not lists of symptoms to avoid.
  4. Constraints are enforced, not reported: a sitting that runs past its length ends; the easel refuses marks until a rest.
  5. Only Alice promotes a recipe into the sketchbook, when she approves the picture, not when a painter claims it won.
Source