Board review, 2026-06-30 (critical). pentimento, michael.art

Four seats, each their own question, prompted for hard critique. Run as workflow board-review-2026-06-30. Spread: code/art/samples/2026-06-30-board-review-spread.png.

Synthesis

Are we ready: No, not ship-ready today, but genuinely close: this is refine-then-ship, not rework - all four critical seats scored 6-7 and agree the core (cohesion, the revising engine, the painter's hand) is real, but three specific defects must be fixed before a curation submission, and none of them require touching the engine's core.

Strongest concern: The single strongest concern, because three of the four seats (the art-historical seat, the technical seat, the AI seat) converge on it independently and I verified it in the engine: EXCAVATION IS NO LONGER RARE. A headless scan of 1,200 renders shows Excavated at 39.0% edition-wide, and because the gate is target>=21 that 39% is concentrated entirely in the dense pieces (which are ~65% of the edition), so within the body of the edition a collector who owns three dense pieces will likely own two with windows. A marquee reflection gesture at that frequency stops reading as "this piece reconsidered itself" and becomes a house-style frame, a disclaimer that the paint is honest rather than a discovery. It was deliberately raised from 1/13 to 0.49 at the artist's request; that call is the one most at odds with the board. Second, tied to it: on the thinnest pieces there is barely any buried paint for a window to expose (and those pieces are excluded from excavation anyway by the >=21 gate), so the pentimento thesis there is a caption, not a visible fact.

Add another AI? NO. Do not add another AI dimension. This is the AI seat's own verdict from the seat that would normally argue for a feedback loop, and it is the right call: the reflection loop this project would bolt on already exists AS PIGMENT in the excavated window (the finals are withheld and the piece's own buried underpainting and scrape-ghosts show through). Any AI you could name (a critic model scoring composition, a second agent re-annotating, a live self-revising loop) would move the reflection from PIGMENT to METADATA, which is exactly the cheap gesture the honest-static-frame framing was built to refuse. Worse, a critic/scoring agent in the render path is a determinism-and-honesty trap: it either forks the PRNG stream and breaks the platform, or it is faked at build time and you are then lying about the loop being live. The one honest AI-adjacent lever is NOT more AI: in the pentimento-play animation, stage exactly ONE visible reconsideration (a stroke laid, then scraped, on its way to the deterministic final) so the loop is SEEN rather than only deposited, using the co-authorship you already claim, adding zero new models to the frame, and keeping the static frame as the canonical work.

Critical path:

Cut/near-eliminate the crypto-meme register. It is currently 20% of note slots (source: paint 54 / gen 26 / crypto 20). the art-historical seat, the typography seat, and the technical seat all rank it a top-two liability: it timestamps the work to a 2026 market moment and lets a curator dismiss it as 'an NFT' not generative art. Demote to near-invisible rarity or remove; weight hard toward the painter LEXICON. This is the fastest, highest-leverage move and touches only a word list.
Drop excavation frequency back toward rare (from 0.49 to roughly 1/8-1/13 among eligible). Verified 39.0% edition-wide, concentrated in the dense body. Make honest mode the default and the surveyor-markup 'mix' the rare variant, so the cut lands as an event, not a frame style. Three seats named this; it is the strongest single concern.
Guarantee the pentimento is VISIBLE in every output. Fix the empty-thesis cases: the art-historical seat's panel-7 barcode collapse (a whole Placement/Density combo shipping with zero visible revision) and the AI seat's thin/Bands pieces that have nothing buried to show. The First-Try forced-ghost gate (isFirstTry && !unpainted, verified) is the right pattern; extend the guarantee so no output ships without at least one legible act of revision. No exceptions.
Rebuild the Cartoon keystone. At 1.3% it is rare, but it is the conceptual floor and the most-screenshotted piece, and both the art-historical seat and the typography seat call it the weakest in the set: an evenly-grey meme word-cloud with no reading order, colliding baselines, words crossing grid ticks, and the crypto voice dominating the one piece where text IS the artwork. Give the note field a hierarchy and a hard cap; make it the most decisive drawing in the edition, not the faintest.
Cut the trait surface, do not add to it (the technical seat). Kill Ledger (10.3% verified; it is on-chain-provenance cosplay, a crypto flex bolted to a painting concept, not revision-archaeology) and the third 'cartoon' excavation mode (a deferred decision kept only 'for comparison'). Demote Density behind Coverage (Coverage is the measured truth; Density is a stroke-count proxy measuring the same axis). One clean rule reads as a signature; seven vocabularies of honesty read as a disclaimer.
Resolve the remaining editing polish, then re-run curation on a fresh batch of HIGH-ENTROPY real hashes (not the repeating test seeds): hide working/trait names from captions; verify the ledger-style inscription and the tally-13 (%%|||) read as a number and not a dashed rule or a stray scribble in situ; fix inconsistent word-spacing/tracking; and tighten board-sheet sampling so near-twin lattice pieces never surface together.

Bottom line: The honest bottom line, no group hug: three of these nine could hang next to Molnar, and the engine passes the hardest test (the Fidenza cohesion test, verified: palette-and-density-divergent pieces still read as one hand). But the set is not evenly governed, and a curator reads the weakest panel, not the strongest. The two things dragging it below the canon are the same two things across every seat: the crypto-meme register (20% of note slots, the language of the AUCTION annotating work that should speak the language of PROCESS) and excavation at 39% (a discovery turned into a house style). Fix those two, guarantee a visible revision in every single output, rebuild the Cartoon, and trim the trait surface (kill Ledger, kill the third excavation mode, demote Density) and it is submittable. Add no AI. You built an algorithm that revises and then annotated it in the slang of the market; the fastest path to the canon is to trust the paint, which is already telling the truth without the survey crew standing in front of it. Verified engine facts backing this: Excavated 39.0%, Ledger 10.3%, Cartoon 1.3%, crypto register 20% (all confirmed against /Users/michaelhiggins/My Drive/NFT My Projects/pentimento/pentimento.js).

The art-historical seat (Molnar / Mohr / Nake lineage) (score 6/10)

Refine-then-ship: three or four of these nine could hang in the canon, but the set is not yet governed tightly enough to survive next to Molnar, and two of the nine read as arbitrary.

This has crossed from exercise into work, and I want to be plain about that first because I am about to be hard. Panels 3 (Sands Saturated Gouache Excavated), 8 (Packed Teeming Raw Umber), and 9 (Sands Saturated Gouache) have real authority: a woven weft of confident color over a legible sub-structure, hierarchy that a knife-painter would recognize, and a grain discipline that reads as one hand. That is the hardest thing to earn in long-form and you have it in the dense register. But the edition is not evenly resolved, and a curator reads the weakest panel, not the strongest. Panel 7 (the red field of horizontal slats) is the one that would get you rejected: it collapses into a barcode, the strokes are near-identical in width and value with no revision archaeology visible at all, so the entire pentimento thesis, the reason this project exists, is simply absent from that output. That is not a rare edge case a collector forgives, it is a whole density-and-placement combination shipping empty. Panel 5, the Cartoon, is conceptually your keystone and visually your weakest link: as lettering it is timid and evenly grey, it lacks the compositional tension of a real preparatory drawing, and the crypto-register phrases (NUMBER GO UP, RIGHT-CLICK SAVE) sit at the same weight as the structural notes, which flattens irony and sincerity into one indifferent murmur. My deeper worry is provenance of language: Molnar, Mohr and Nake never annotated the canvas with the vocabulary of the market that would later sell them. The crypto-meme layer is the single element most likely to date this work to 2026 and read, in ten years, as a novelty rather than a painting. That is the thing a serious collector catches instantly.

Concerns:

Panel 7 (red horizontal-slat field) has NO visible revision: uniform width, uniform value, evenly spaced. It reads as a barcode or a Gerhard Richter strip-painting knockoff, and it carries zero pentimento. A whole Placement/Density combination is shipping without the concept. This is a reject-trigger, not a quirk.
The crypto-meme register (NUMBER GO UP, RIGHT-CLICK SAVE, WAGMI, HODL) is the biggest liability in the whole system. It timestamps the work to a market moment and reads as in-joke, not painting. The Molnar/Mohr/Nake lineage annotated with the language of PROCESS, never the language of the SALE. I would cut the crypto register entirely, or demote it to near-invisible rarity. It is the fastest way a curator dismisses this as 'an NFT' rather than generative art.
Panel 5 (Cartoon) is too polite. A real cartoon/underdrawing has compositional force even in graphite: pentimenti, crossings-out, weighted decisions. This one is an evenly-grey word-cloud. If the unpainted piece is the conceptual floor, it must be the most decisive drawing in the edition, not the faintest.
The EXCAVATION windows (panels 1, 2, 4) still read more as 'a rectangle was masked out' than as an honest cut. In panel 4 especially the window edges are the most legible thing in the frame, which inverts the intent: the cut should reveal, not announce itself. Molnar's cut squares earned their edges through the underlying grid logic; here the seam still looks applied.
Excavation at ~40% is too high. You raised it because 'it loves excavation,' but a marquee gesture on nearly half the edition stops being archaeology and becomes a house style tic. Rarity is what makes a buried-window feel like a discovery. At 40% the collector who owns three will own two with windows and feel the deck was stacked.
The captions still expose working names (Excavated, the trait labels bleeding into the sample sheet). Minor, but it signals the system is still talking to itself rather than to a viewer. On a curation submission every visible token should be resolved language.
Value range across the set is uneven: the dark grounds (panels 1, 4) and the bright grounds (3, 9) are strong, but the mid-grey/bone panels (6, 8-adjacent) risk reading as tasteful rather than necessary. Molnar's restraint was mathematical; some of these quieter panels read as merely muted.

Recommendation: Refine-then-ship. Single most important move: kill or near-eliminate the crypto-meme register and fix the empty-thesis outputs (panel 7's barcode collapse) so that EVERY piece, including the sparse and slat cases, carries at least one visible act of revision. No output may ship without the pentimento in it.

You have built an algorithm that revises, and then annotated it in the language of the auction. Molnar cut her squares to show the rule; do not let the market's slang be the thing that shows through your paint.

The typography seat: legibility vs abstraction (score 7/10)

The single-stroke hand reads as a real painter's marginalia and mostly earns its place, but three failure modes (the tally-13, the repeating-hash ledger, and the CARTOON's density) tip it toward gimmick and infographic.

Judged on the letterforms themselves, the alphabet is the strongest it has been: the SMOOTH-corner rounding is doing honest work, so bowls in PAINTED OVER, KEEP IT, WAS HERE FIRST, STILL HERE curve like a felt-tip while E/T/K stay cut, and the slight slant and jitter read as a hand and not a font. On the excavated Gouache piece (r1c3) the over-notes are the best argument in the whole spread: they auto-contrast off the paint, sit at a believable annotator's scale, and the dimension line with 22 and 17 turns the cut into a survey rather than a caption. That is lettering as ART, integral to the mark, not a label pasted on top. The excavation surveyor notes (SCRAPED BACK, LET IT SHOW, BENEATH, STAYS) are legible and quiet inside the window and read exactly as a restorer's pencil should. Where it hurts is at the extremes. The CARTOON piece (r2c2) is a legibility stress test the alphabet nearly fails: NUMBER GO UP and DRAG collide with RIGHT-CLICK SAVE, several words overprint the grid ticks, and the eye can't find a reading order, so the register-mixing that is charming in one or two phrases becomes a word-cloud of memes, and the crypto voice (DIAMOND HANDS, RIGHT-CLICK SAVE, GM GM) dominates the painter voice in the one piece where text IS the artwork. The all-4s ledger (r2c1) exposes a real weakness of the inscription: a long monospaced run of identical glyphs reads as a dashed rule, not provenance, and only the varied Bone and Ink hash (r3c1) with its tick-ruler proves the device can read as a written number. On the dense Raw Umber packed piece the notes are all but eaten by paint, which is conceptually correct for Buried coverage but means the lettering contributes nothing there. So the hand is good; the editing and the density governance around it are what still wobble.

Concerns:

The CARTOON piece has no typographic hierarchy or reading order: uniform weight, colliding baselines (NUMBER GO UP over DRAG over RIGHT-CLICK SAVE), and words crossing grid ticks make it read as a meme word-cloud rather than a composed page. This is the one piece where type is the entire artwork and it is the weakest-set piece typographically.
Register balance is off in the text-forward pieces: crypto-caption phrases (RIGHT-CLICK SAVE, DIAMOND HANDS, GM GM, NUMBER GO UP) crowd out the painter's LEXICON, so the most-screenshotted element risks reading as crypto-in-jokes rather than a painter's regret. The painter voice should lead by a wide margin wherever text is prominent.
The ledger inscription is fragile: on low-entropy or repetitive hash regions a run of identical glyphs collapses into a dashed line and stops reading as a number. The test seeds (all-4s, all-6s) make this look worse than a real mint would, but the failure mode is real and should be seen on varied hashes before it is called done.
The tally-mark THIRTEEN (%%|||) was not visibly present in any of the nine panels I could read, so I cannot judge whether it registers as thirteen or as noise. As a two-glyph invented mark dropped rarely into a text field, it is at high risk of reading as a stray scribble rather than a number unless it is spatially set apart. It needs to be seen in situ.
Under-notes on saturated grounds sit at very low contrast by design (FIBONACCI on the red Primer piece nearly vanishes), which is fine as under-process but means several notes are effectively invisible and are doing no compositional work, only PRNG cost. Decide whether a barely-there note is a feature or dead weight.
Word spacing and letter-tracking are inconsistent across notes: some phrases are tightly set and others loose, which on a hand this uniform reads as a spacing bug rather than as expressive rhythm. A painter varies stroke, not tracking; here it is the reverse.

Recommendation: refine-then-ship. The alphabet and the over-paint/excavation annotation are ready and are a genuine strength; do not touch them. Single most important next move: fix the text-forward pieces (CARTOON above all) by giving the note field a hierarchy and a hard cap so the painter register leads and words stop colliding, then verify the tally-13 and the ledger inscription on a batch of high-entropy real hashes, not the repeating test seeds.

The margins are written in a painter's hand; the CARTOON is written in a group chat.

The technical and concept seat (score 7/10)

The engine is one clean idea wearing five feature-flags as a coat; the animation is honest but the trait sheet is where the bloat lives.

Looking at the nine, the core rule holds and it holds hard: a Fibonacci grid, a locked grain, one or two dominant colors, marks that stay visible under confident marks. Top-right, bottom-right, and top-middle share nothing in palette or density yet are obviously the same hand - that is the Fidenza test passed, and it is the hardest thing to get. The dominance weighting is doing real work; almost every panel resolves to a two-color chord instead of confetti, which is exactly the discipline this concept needs. Where I get nervous is that the ONE rule is now surrounded by named toggles - Excavation with three sub-modes, Ledger, Cartoon, Coverage, Fate, tonal wash, favorites pool, thirteen register - and several of them are the same idea (revision made visible) re-notated as separate traits. The staged build in pentimento-play is the most defensible addition on the sheet: it fires read-only off true layer boundaries and settles on the byte-identical static frame, so it literally re-walks the honest construction order rather than performing a fake one. That is not a gimmick, it is the thesis animated, and I would keep it - but ONLY if it never becomes the canonical way the piece is seen. The static frame must stay the work. My real worry is legibility of the system to a collector: a piece is honest, but the honesty is now spread across seven vocabularies of honesty, and a pile of features can dilute a single strong rule even when each feature is individually justified.

Concerns:

Excavation at ~40% is no longer a rare cut, it is a house style - and mix-mode stacks a value-lift + re-ruled grid + 2-3 survey notes + dimension line ON TOP of the honest reveal. The honest reveal (finals withheld, real buried paint shown) is the strong idea; the surveyor markup is decoration defending the idea from a skeptic. At 40% frequency that decoration becomes the look. Consider: honest mode as the default, mix as the rare one.
Three excavation modes is a tell. If 'cartoon' mode is conceptually weaker by your own admission and only kept 'for comparison,' it should not ship. Two modes is a decision; three is a deferral of a decision.
Coverage, Density, Fate, and Excavation are four traits all measuring the same axis: how much process is buried vs shown. A collector reading the trait table sees redundancy, not richness. Density is a stroke-count proxy, Coverage is the measured truth - pick the honest one (Coverage) as primary and demote Density to a hidden internal.
Ledger (writing the token hash into the canvas) is a different, weaker idea than pentimento. It is on-chain provenance cosplay, not revision-archaeology. It reads as a crypto-native flex bolted to a painting-native concept. It photographs fine but it does not belong to the thesis - I would cut it before launch, not after.
The two lattice pieces (top-right, bottom-right) are almost the same painting in the same palette. Nine-panel curation should never surface two near-twins; it makes the edition look smaller than it is. Tighten the sampling for board sheets and check the actual variance within Sand/Saturated/Gouache/Grid - that combo may be under-varied.
The favorites pool at ~50% plus the thirteen register plus three crypto/code/painter registers means the lettering is now doing a LOT of signaling. The notes are the most screenshotted element and they are drifting from 'a painter's private regret' toward 'a caption track that name-drops FIBONACCI, GRAIL, GM.' The painter register is the soul; the crypto register is the risk. Weight harder toward LEXICON.

Recommendation: refine-then-ship. Single most important move: cut the trait surface, do not add to it - kill Ledger and Cartoon-excavation-mode, demote Density behind Coverage, and drop Excavation frequency back toward rare so the cut lands as an event. The animation stays as an opt-in viewer, never the canonical render.

One clean rule is a signature; seven ways of saying 'this is honest' is a disclaimer. Trust the paint - it is already telling the truth without the survey crew standing in front of it.

The AI and creativity seat (painting robots; feedback loops over one-shot generation) (score 7/10)

No. Do not add another AI dimension. The reflection you'd bolt on already exists as pigment in the Excavated window - the honest thing now is to trust it, not narrate it.

Here is what I actually see, and it changes my usual answer. My whole practice is that the interesting art is in the critique loop, the machine looking back at its own mark and deciding to revise. In most generative work that loop is invisible; you get the confident final and have to take the reflection on faith. Pentimento does the rare thing: the loop is deposited AS PAINT. In the top-left and mid-left Excavated pieces, the cut window is not a decorative overlay or a second AI pass commenting on the first - it withholds the finals and exposes the piece's OWN buried underpainting and scrape-ghosts, surveyed with two or three revision notes and a dimension line. That IS the feedback loop, rendered honestly, in a static frame. So when I ask "does it need another agent, another reflection dimension," my honest answer is that the reflection is already the medium here, not a missing feature. Any AI I could name - a critic model scoring the composition, a second agent re-annotating, a live self-revising loop - would move the reflection from PIGMENT to METADATA, and metadata reflection is exactly the cheap gesture the honest-static-frame framing was built to refuse. That said, I am not fully sold the mechanic always DELIVERS: the red Bands piece (mid-right) and the red Cartoon-ish center have almost nothing buried to expose, so on those pieces the reflection thesis is a claim, not a visible fact. The concept is right; its worst 20% of outputs undercut it more than any absent AI would deepen it.

Concerns:

The reflection loop is only VISIBLE on the dense, heavily-worked pieces. On the thin ones (the red Bands piece mid-right, the sparse reds) there is barely any buried paint for an Excavation to expose or for the notes to reference, so the pentimento thesis becomes a caption rather than a thing you can see. That is the real gap, not a missing AI.
Excavation at ~40% risks turning the marquee reflection gesture into wallpaper. When almost half the edition cuts a window, the cut stops reading as 'this piece reconsidered itself' and starts reading as 'this collection has a frame style.' Reflection that is guaranteed is no longer reflection.
The single most tempting AI addition - a critic/scoring agent in the render path that decides WHAT to scrape - would be a determinism and honesty trap: it either forks the PRNG stream and breaks the platform, or it is faked at build time and then you are lying about the loop being live. Do not do it.
The notes register does the reflecting in WORDS (UNDER THIS, PAINTED OVER, I WAS WRONG). That is close to over-explaining a loop the paint already performs. More AI-generated text would push it further toward telling instead of showing.
The layer-build animation is the one place a genuine reflection dimension is latent and underused - you show the stack assembling forward (grid to finals) but never show the machine CHANGING ITS MIND. The revision is baked into the final frame, not staged as a visible reconsideration.

Recommendation: Refine-then-ship. Add ZERO new AI. The single most important move: make the buried layer real on EVERY piece before you trust Excavation to expose it - gate the thin/Bands outputs so a genuine underpainting-plus-ghost always exists beneath the finals, then the reflection is a visible fact edition-wide, not a caption on 80% of it. If you want one honest AI-adjacent lever, it is not more AI - it is the animation: in pentimento-play, stage ONE visible reconsideration (a stroke laid, then scraped, on its way to the deterministic final) so the loop is SEEN, not just deposited. That deepens the reflection thesis using the co-authorship you already claim, without adding a single new model to the frame.

The best painting robots don't add a critic - they leave the scraped mark showing. You already do that in pigment. Don't move the confession into metadata.