Before: the paint simulator
This build follows a run on an open-source oil-paint simulator. There the critic was built, language-model painters drew by writing Lua strokes on an easel, and the hybrid method (default map, exploration, pause, drawing) was first used on a seated figure.
The same simulator is the painting studio of Study Wall, a sibling build: briefs read before any picture exists, walls of studies, and paintings that are complete only when they survive a critic's test.

The brief
Brief (hybrid-01)
A charcoal or pencil drawing of a seated female figure on the ground in an informal, vulnerable pose. Her hands are clasped and defined, but her face is left indistinct, dissolving into gesture or shadow. Let the marks search, falter, or declare. Allow them to outline, shade, and carve out the air between her and the space she occupies. Absence should be intentional. Line density should ebb and surge across the form. Drawn with tension between form and omission. The marks should assert where the artist knows--and stutter where they don't. Hands resolve; the face dissolves. Surface weight shifts across the figure. Let the drawing feel unfinished, if finishing would betray it.
The default map (Midjourney)
Measured spread of the default map
Default map: measured spread per batch
Median [25th-75th percentile] per key. Same readers and registry meanings as critic/vtl/registry.py and critic/color/registry_color.py; tonal/marks thresholds are provisional.
| group | key | NotSteered | Steered |
|---|---|---|---|
| placement | dx | -0.00355 [-0.0236-0.0174] | -0.00276 [-0.014-0.0301] |
| placement | dy | 0.0906 [0.0627-0.0989] | 0.0674 [0.0442-0.0743] |
| placement | dr | 0.0595 [0.0499-0.0707] | 0.0434 [0.0302-0.0701] |
| placement | dx_ton | 0.00242 [-0.00459-0.00537] | 0.00262 [-0.000857-0.00583] |
| placement | dy_ton | -0.0138 [-0.0255--0.0106] | -0.00696 [-0.00891--0.00468] |
| structure | rv | 0.924 [0.913-0.934] | 0.924 [0.916-0.931] |
| structure | mu | 0.821 [0.81-0.874] | 0.899 [0.879-0.919] |
| structure | sdi | 0.205 [0.198-0.211] | 0.174 [0.169-0.177] |
| structure | vcli_g | 3.11 [3.03-3.29] | 3.17 [3.13-3.26] |
| structure | sci | 2.85 [2.81-2.91] | 2.86 [2.81-2.94] |
| structure | drc | -0.0312 [-0.0485--0.0248] | -0.0122 [-0.0178--0.00655] |
| structure | n_islands | 6 [5.75-7.25] | 6.5 [5.75-8] |
| structure | drain_margin | 0.0345 [0.0117-0.0494] | 0.0281 [0.0165-0.0676] |
| tone | mean_zone | 7.07 [6.55-7.26] | 7.43 [7.26-7.79] |
| tone | shadow_mass | 0.0766 [0.0582-0.104] | 0.0445 [0.0355-0.0751] |
| tone | highlight_mass | 0.754 [0.706-0.784] | 0.834 [0.802-0.849] |
| tone | midtone_share | 0.155 [0.141-0.165] | 0.106 [0.0973-0.112] |
| tone | range_p2_p98 | 77.8 [74.4-81.1] | 72.7 [71.2-79.8] |
| tone | gini_mass | 0.537 [0.511-0.575] | 0.628 [0.605-0.639] |
| tone | white_area | 0.374 [0.329-0.632] | 0.726 [0.692-0.774] |
| tone | black_area | 0.0132 [0.00902-0.0304] | 0.00933 [0.00583-0.0138] |
| tone | withheld | 0.58 [0.547-0.612] | 0.663 [0.642-0.689] |
| marks | stratification | 0.472 [0.441-0.5] | 0.458 [0.417-0.512] |
| marks | unanimity | 0.0887 [0.067-0.18] | 0.153 [0.126-0.181] |
| marks | pressure_cv_mean | 0.887 [0.752-1.09] | 1.18 [0.964-1.37] |
| marks | detail_democracy_cv | 0.971 [0.882-1.18] | 1.25 [1.18-1.32] |
| marks | withheld_share | 3.41e-05 [0-0.0252] | 0 [0-0.00182] |
| colour | col_okc_C_mean | 0.0107 [0.00783-0.0158] | 0.0147 [0.0127-0.0164] |
| colour | col_light_sdi | 0.356 [0.309-0.453] | 0.334 [0.303-0.454] |
| colour | col_dark_sdi | 0.291 [0.279-0.317] | 0.292 [0.278-0.296] |
| colour | col_focal_count | 16.5 [15-18.2] | 15.5 [13.5-16.2] |
| colour | col_primary_focal_dominance | 0.154 [0.128-0.176] | 0.21 [0.191-0.275] |
| colour | col_border_salience | 0.0454 [0.0276-0.0611] | 0.0185 [0.0118-0.0242] |
vcli_label
- NotSteered: earned tension 12/12
- Steered: earned tension 12/12
regime
- NotSteered: gradient-field / controlled instability 12/12
- Steered: gradient-field / controlled instability 12/12
col_regime_tags
- NotSteered: quiet field with sharp accents 12/12, high-light ground 12/12, low-chroma field 12/12, diagonal force 11/12, distributed salience 10/12, narrow hue field 8/12, dispersed light 5/12, compact dark 5/12, color rides structure 2/12, color-structure decoupled 2/12, edge-activated field 1/12
- Steered: quiet field with sharp accents 12/12, high-light ground 12/12, low-chroma field 12/12, diagonal force 11/12, distributed salience 7/12, dispersed light 5/12, compact dark 5/12, narrow hue field 4/12, color rides structure 4/12, edge-activated field 2/12, discrete chroma objects 1/12
The critic’s map of the default
Pass 5: The default map (handoff to S2, exploration)
Reading: the brief, PROMPTS.md and Russell's note; both contact sheets; Steered #03 and #10 heads at full resolution; the rows files for every per-image count below; Passes 1-4. Counts carrying a [key value] come from measured/summary.md or measured/*.rows.json; counts without one are judgement (mine or an earlier pass's, credited). No number is attributed to A27, A4, A5, A30, CTG or guards. dx, dy and sdi describe the edge centroid and its spread, not where the figure is (registry: dx cannot_say where the subject is); where I say "figure" the claim is judgement.
Where the passes disagree, and which reading I followed
- Pyramid lean (Pass 1 "same right-leaning triangle 12/12" vs Pass 2 "no single direction"): followed Pass 2; the rows agree, Steered dx is negative in 8 and positive in 4.
- What white_area shows (Pass 1 "half the frame empties" vs Pass 2 "partly paper tone"): followed Pass 2; white_area is the largest region above a fixed white threshold, not ground.
- Which Steered face is lost (Pass 1 #03, Pass 3 #10, Pass 4 none): followed Pass 4; at full size #03's nose and mouth are findable inside the shadow and #10's eye, nose and mouth are legible, so findable 12/12. #03 is still the one image where shadow covers it (Pass 1's point, kept).
- NotSteered resolved faces (Pass 1 11/12 vs Pass 3 12/12): followed Pass 3, who opened #07 at full size.
- Face-over-hands lists (Pass 1 and Pass 3 both 6/12, different images): kept the count, not a list.
- Route (Pass 2 "geometry first" vs Passes 3-4 "order of finish first"): followed Passes 3-4 for priority (prompt 1 below is the finish allocation), and Pass 4 for form: one refusal per prompt, so Pass 2's geometry moves appear as separate prompts, never stacked on prompt 1.
- Pass 3's four "ask for instead" items vs Pass 4's warning that they become adjectives: followed Pass 4; no prompt or sketch brief below uses search, falter, stutter, dissolve, omission, unfinished or vulnerable.
The default map
- Edge mass centred side to side and below centre: |dx| under 0.05 in 18/24 [Steered median dx -0.00276; NotSteered -0.00355]; dy positive (down) in 24/24 [Steered median dy 0.0674]; brightness centroid above centre in 23/24 [Steered median dy_ton -0.00696; exception Steered #11, dy_ton 0.00102].
- Height-filling, then cut loose from the frame: head at or near the top edge 24/24 and body in the central third 24/24 (Pass 2, judgement); bottom-edge contact 9/12 NotSteered to 0/12 Steered (Pass 2); border activity low in 12/12 Steered [col_border_salience max 0.0314, median 0.0185] against 5/12 NotSteered above 0.05.
- One pose and one viewpoint: knees up with arms over them 20/24, front or three-quarter body 22/24, eye level 24/24, back view 0/24 (Pass 2, judgement).
- The face is always there: findable in 24/24 (Passes 3-4 and my full-size check); direct gaze 6/12 NotSteered; head bowed 12/12 Steered. The steer turned the gaze down and kept the face.
- Hands below the face in finish and in the frame: hands at feet or shins 9/12 Steered (Pass 2); face more resolved than hands 6/12 Steered (Passes 1 and 3); hands out-know the face 0/24 (Pass 3, judgement).
- Finish by object class: skin stumped smooth 11/12 Steered (Pass 1), cloth as angular scribble (at least 10/12 NotSteered, Pass 1), a ruled horizontal ground stroke 12/12 Steered (Passes 1-2).
- Paper reserved around the figure, never quiet inside the drawing: largest white region over 0.6 in 11/12 Steered [white_area median 0.726; exception #04, 0.285], while the marks reader finds a quiet zone under 0.03 of the area in 12/12 Steered [withheld_share max 0.026, median 0].
- No committed black and no dark off the figure: largest black region under 0.035 in 22/24 [Steered max black_area 0.0287; exceptions NotSteered #02 0.0625, #07 0.0418]; darks owned by the figure 21/24 (Pass 2); midtones under 0.15 in 12/12 Steered [midtone_share max 0.144].
- The measured steer is unevenness, placed by category: density varies between zones above 1.0 in 12/12 Steered vs 5/12 NotSteered [detail_democracy_cv Steered median 1.25, NotSteered 0.971]; the key says the unevenness exists, not where it falls, and on the sheets the dense zones are cloth and seat and the thin zones skin and paper (Passes 3-4, judgement).
- The steer narrowed the set: tighter IQRs on mu [0.81-0.874 to 0.879-0.919] and white_area [0.329-0.632 to 0.692-0.774], and three habits reach 12/12 (bowed head, ground stroke, clear bottom edge). Same kernel label 24/24 [vcli_label earned tension 12/12 each batch], which tracks none of the lines above.
Keep from the brief: charcoal or pencil, a woman seated on the ground, informal pose, clasped hands. Refuse what the generator added: the pretty finished face, the bowed head as the sign, the knees-up pyramid at eye level, smooth skin, the ruled ground stroke, paper as the only air.
S2 brief
(a) Midjourney prompts (Russell runs each with --style raw at --ar 4:3)
Each is one refusal, run as its own batch; they are alternatives, not a stack (Pass 4: one cost at a time).
1. Refuses line 5 and 6 (finish by class, face above hands). Russell's note says the polish cannot be asked away; this removes the two places it lives. `` graphite pencil drawing on white paper, a woman seated on the floor holding her clasped hands up at chest height, the two hands drawn in full detail with knuckles, thumb, fingernails and the shadow of one hand on the other, the hands the darkest and most finished area of the page, her head an outline of the skull only with no eyes, nose or mouth, her arms and legs plain thin contour lines with no shading, no smudging, no blending --ar 4:3 --style raw ``
2. Refuses line 2 (height-filling, centred block). `` pencil drawing of a small woman seated on the ground in the lower left corner of a wide sheet, her head well below the middle of the page, a large plain field of paper above and to the right of her, the figure taking up less than a quarter of the sheet, her hands clasped in her lap --ar 4:3 --style raw ``
3. Refuses lines 3 and 4 (eye-level front view, bowed head as the sign). `` charcoal drawing seen from above and behind, a woman seated on a bare floor with her back to the viewer, the top of her head and her shoulders toward us, her hands clasped behind her neck with every finger drawn, the floor seen as a tilted plane of long parallel strokes around her, no face visible --ar 4:3 --style raw ``
4. Refuses line 8 (no committed black, every dark owned by the figure). `` charcoal drawing, a woman seated on the ground drawn in light grey line, a large flat solid black shape on the floor and wall behind her that is not her shadow, its edge running close along her back and knees, deep black against bare white paper, her clasped hands pale against the black --ar 4:3 --style raw ``
5. Refuses lines 6 and 7 (ruled ground stroke, paper as the only air). `` charcoal drawing, a woman seated on the ground, the space all around her filled with dense crosshatching that stops just short of her outline, so she is a pale shape held by the dark around her, the floor and the air drawn with the same strokes, no horizon, no floor line, her clasped hands at the centre of the sheet --no horizon line --ar 4:3 --style raw ``
What should move if these land, per prompt: 1, thinner midtones and contrast concentrated in one area [Steered median midtone_share 0.106, gini_mass 0.628, col_primary_focal_dominance 0.21 should each move with the hands]; 2, edge centroid off centre [|dx| over 0.05 in only 6/24 now]; 4, a committed dark [black_area under 0.035 in 22/24 now]; 5, a quiet zone inside the drawing and less reserved white [withheld_share 0 and white_area 0.726 Steered medians]. 3 has no measured habit to move; it is judged by eye.
(b) Easel sketch briefs (pencil or black chalk, 4:3 sheet saved at 600 px, a few chunks)
Tools are only those in the easel guide's "Drawing: pencil, chalk and eraser": pencil grades 9H-9B, chalk(), sketch / line / rule / hatch, erase, fix. Coordinates in easel units (1000 wide, origin top left). Each makes a decision the map shows the generator never takes.
Sketch A: the hands out-know the head (placement of finish).
- Chunk 1: pencil("4H"), sketch the whole seated figure at pressure 0.2, a few passes, sitting low and left (head near x 300, y 330; seat near y 640), so the block-in is all there is of head, arms and legs.
- Chunk 2: pencil("2B"), line the two clasped hands at chest height (around x 330, y 430) with corners kept (smooth=false): which hand is in front, the thumb plane, the knuckle row. pencil("4B") hatch the planes turned from the light inside the hands only, two directions. Sharpen between.
- Chunk 3: erase one stroke down through the head at strength 0.6, leaving its ghost. fix() over the hands. Nothing else is touched again.
Sketch B: a real earlier answer, then a frame that cuts (correction and placement).
- Chunk 1: pencil("HB"), line the head and near arm at one angle: head upright, around x 760, y 250.
- Chunk 2: erase that head and arm at strength 0.7 (the ghost stays), then pencil("2B") line them again lower and turned (head near x 820, y 360), so the first answer is visibly rejected and the second is different in size and angle. The figure runs off the right edge at the shoulder and knee.
- Chunk 3: pencil("3B") line the clasped hands fully; leave the far leg as the HB block-in because attention left it. No rule anywhere: there is no ground stroke.
Sketch C: the air is drawn, and a dark she does not own (mark and pressure).
- Chunk 1: pencil("2H"), line the figure as contour only, centred low (seat near y 700), no hatch on her body.
- Chunk 2: chalk(), hatch the floor and wall around her at pressure 0.5, spacing tight, stopping about 12 units short of her contour, so she is reserved paper held by marks rather than a drawing on blank paper.
- Chunk 3: chalk(), hatch one block at upper left (about x 0-380, y 0-420) in two crossed angles at pressure 0.9 until it is the darkest thing on the sheet, its lower edge 20-30 units above her head. pencil("6B") line the clasped hands as the only dark on her. fix().
Exploration
The pause
S3/S4: the choice
Russell's decision (2026-10-06): let the critic's recommendation stand (s2_explore/critic/passes/5_pause.md). The eye was offered and not used here: this choice is the system's, unattended.
Chosen direction: p1 #4's structure (finish placed in the hands, the face held back, the frame cutting the head), with RP_B #1's drawn-and-taken-back head in place of the blank face, and no inset hatched panel. Midjourney is used sideways for hand reference only (transfer of p1 #4's hands as contours and value zones: s5_painting/ref/ref_hands.lua).
Sources: p1 #4 = Midjourney (prompt p1 + easel sketch C as image prompt); RP_B #1 = Midjourney (Russell's prompt + easel sketch B); the head's draw-and-erase move = easel sketch B (Claude painter).
The sheets





Sheet 3’s material test (why the easel’s black floors near L* 37)
Material test (run by the orchestrator, 2026-10-06, on a scratch sheet; not this drawing)
Question: does chalk or pencil take again where it has been erased? Sheets 1-3 have treated "erased chalk will not take chalk again" as a fact, and the critic deferred the head take-back because of it.
Measured (median L*, lower = darker), same canvas and ground as this sheet:
- pencil 4B, drawn, erased (strength 0.95), redrawn at the same pressure: 72.3; drawn twice, never erased: 68.3
- chalk, drawn, erased, redrawn: 50.2; drawn twice, never erased: 47.7
- chalk, drawn light, fix(), erased, redrawn hard: 49.0; the same without fix(): 50.5
So both take again over erased areas, fixed or not, close to their normal darkness. The earlier observation most likely met chalk's floor instead: chalk cannot go darker than about L* 37 however much it is layered, so pressing over an existing grey seems to "not take". A take-back of the head (erase, then redraw) is feasible on this sheet. Nothing here changes the critic's moves; it corrects the premise one of them rested on.
The unattended studio

What carried over
- The critic: the VTL canon frozen verbatim, the registry of what each number may and may not say, tonal and mark readers, Russell's colour analyser, Protocol v8 routing, the guard, and the engine passes.
- The hybrid method: a default map from a generator, exploration, a pause, the drawing, and a sideways look.
- From Codex's parallel build (read only): Hands as the motor, every stroke paying for itself, generator images as targets only, and intention kernels with edge-integrity guards.
What did not
- Language models writing coordinates (the drawing was mechanical).
- Body meshes and CLIP-based sketching (non-commercial licences, and the mesh model drops the head and neck).






