Gustav Klimt — Mäda Primavesioil on canvas, 1912–13 · Met object 436819, accession 64.148 · Open Access CC0

Entry 01 — Selection & condition check

An edge case, before any numbers

The brief: pick a well-known work that genuinely stresses the instrument, modern, and safely usable — life + 70 in 2026 means artists dead by 1955 [corrected, Entry 42; originally misstated as “the late 1910s”], and in practice the Met's Open Access program (CC0, full-resolution). Klimt (d. 1918) qualifies comfortably; Mäda Primavesi was confirmed public-domain via the collection API and fetched at full resolution. The sitter is Mäda Primavesi (1903–2000), aged about nine, daughter of a principal Wiener Werkstätte patron — a portrait staged against decorated grounds. That is exactly the predicted stress: ornament competing with the figure as structure, a regime the library had not touched (closest prior: the Matisse's flat planes, but here the pattern is the ground).

Condition check, per AGENTS Step 1, looking before measuring:

Predictions logged before the read, all later confirmed: three registers (wall / floor / dress bridging them); the smooth white dress likely reading as a void; the rose belt as the one concentrated chromatic event; ornament as a self-standing field.

Entry 02 — The headline

A portrait the instrument reads as a diffuse decorative field

Full spine read (isr read, ~67s no-persistence; persistence rerun in Entry 09 leaves every number below unchanged). The state classifier returns diffuse_field with flags low_packing_density and peripheral_pull; the regime block reads gradient-field / controlled instability, distributed. Despite an unmistakable central subject:

Entry 03 — Mechanism, leg 1

The highest colour-in-void in the corpus

β 0.378 — colour operating where drawing is absent — against Degas 0.24 and Matisse 0.14. The field panels corroborate: the colour-evidence field is brightest in the wall, near-zero on the figure. Colour is a self-standing layer displaced off the subject and into the ground.

Entry 04 — Mechanism, leg 2

Total decoupling with a spatial signature

coupling_index −0.022: colour and edge do not co-occur pixelwise — the Degas family of total decoupling (at this magnitude the sign is noise; the zero is the point). But where the Degas's decoupling is dissolved evenly through one weave, Klimt's has a spatial signature: it runs register by register. Four of the five construction families point at the figure; the colour centroid alone is displaced up into the wall — ω 0.106, the cleanest such split in the library. And colour does not just fill the void, it draws: chromatic-boundary independence 0.191 (corpus high; Degas 0.14, Matisse 0.10, provisional autonomy threshold 0.20) — colour lines light does not draw.

Entry 05 — Mechanism, leg 3

The dress is a high-luminance void — in the corpus's first high key

Tonal mass: 37.2% highlight, 1.6% shadow, τ 0.598. The painting lives in its lights — the first work in the library to do so (Matisse is midtone-keyed, Degas and Caravaggio shadow-keyed). And it spends that key on disappearance: the white dress — the largest, brightest thing in the picture — reads low on mass, low on chroma, low on the support field, and is held together only by its contour edge. This is the Matisse-woman signature — an edge enclosing a void — inverted from midtone gray to high-key white. The subject does not anchor the structure because, structurally, there is almost nothing inside her outline.

Entry 06 — Attention vs weight

The eye goes where the structure isn't

Saliency and structural mass diverge: mass_divergence 0.080 — Matisse-like (0.072), and the opposite of the Degas, where weight and gaze rest together (0.02). The eye is pulled to the figure and the belt; the structural weight sits up in the flecked wall. Where you look and where the picture's weight lies are two different places — the portrait apparatus points one way, the field pulls the other. The centering itself is hollow: Default Gravity Index 52, "leaving the basin" — centering high (0.79) but inner-mass low (0.176). A symmetric, frame-compliant subject (RC 0.87, grid the most symmetric of the three compared works) that the structure declines to weight.

Entry 07 — Two accent systems

The Degas pattern, in a new key (I-7 applied, not tripped)

The condition check predicted the default accent lens would miss the pure-colour flecks — the I-7 channel bias, learned the hard way on the Degas. So both channels were run from the start:

Entry 08 — R_spatial (candidate C-3)

The strongest, widest superadditive at n=6

The candidate battery (AGENTS step 7 — mandatory, appended to the register in the same commit): R_spatial 2.91–5.04, superadditive. The highest and by far the widest range in the corpus (width 2.13; the previous widest, Matisse, was 0.36). The reading is consistent with everything above: every local patch is a balanced piece of decorative field carrying little torque, so no part is loud — yet the whole holds a strong up-and-left asymmetry. The global torque is almost entirely emergent, an arrangement effect no region owns.

Entry 09 — Persistence (deepening pass)

The anchor is chromatic, not tonal

Full persistence read (report_full.json; all Entry-02–08 metrics unchanged). The composition survives blur (0.82) and low-contrast (0.85) with its anchor essentially fixed — and drops to 0.455 under grayscale, where the anchor does not survive (drift 0.323: it relocates entirely).

Entry 10 — I-4 zone probes

The object-level chroma ordering: wall 12× dress

deepen_probes.py → zone_probes.json: nine named zones, each read over the object's own mass inside its box (the I-4 mask-to-mass read — loose boxes stop lying), mask field chosen per object (colour for the chromatic zones, mass/tone/edge for figure and zigzag). The colour ordering is blunt:

The dress is a chromatic-and-mass void defined by contour (colour 0.069, edge 0.383) — the object-level proof of Entry 05. And the subject as a whole (dress + face + hair) is the least chromatic system in the painting, out-chroma'd by the empty wall roughly twelve to one.

Entry 11 — Resolution sweep (I-10)

The sweep corrects our own first pass

Worth logging as a nuance to I-10 itself: the finding says thin-structure measures don't survive resolution sweeps. Here one did — within a bounded window, breaking only at native res (the speck-entry effect, kin to the Caravaggio's I-11 story). Bounded-window stability is a usable middle state between "robust" and "reading aid."

Entry 12 — Edge-mode comparison (step 4)

The verdict is mode-robust; the magnitude is not

The read CLI does not expose edge-mode (the per-zone reads are never border-weighted — I-1); the island system does, so both modes were run through the Python API and as readouts. Both report edge_artifact_risk low:

Practice going forward in this study's records: entropy is reported as the range 0.63–0.82. The diffuse_field verdict — no dominant island, no hierarchy — holds in both modes; the degree of diffuseness is an edge-mode choice and is labeled as such.

Entry 13 — Same-coordinate compare

Placement against the poles

isr compare against the two calibration poles, on the studies' pinned source reproductions (Matisse source jpg; Degas Study-2 crop). n=3 — z-scores are "who sits where," never significance; and the three images are three different reproduction chains, so this is placement, not measurement of the paintings. Top-CV rows (where the group genuinely differs):

Entry 14 — Critique & close-out

What shipped, what stays open

The critique essay (critique.md / critique_standalone.html, per skills/isr-critique) is titled "The subject is the least chromatic thing in the painting." Its optional-interpretation section says the one interpretive sentence the measurements support: Klimt gives the child the whole compositional apparatus of importance — center, symmetry, scale, frontality — and withholds the structural weight that would make her the engine of the picture, handing it to an ornamental ground instead. The apparatus of centrality and the distribution of weight have been pulled apart; the instrument can locate the seam, not the intent.

Study closed 2026-07-07. Open threads passed to OBSERVATIONS: a second reproduction of the subject to bound the photography chain; a third R_spatial domain (photo / scientific / AI) — the register's promotion bar now most needs domains, not works; a Wiener-Werkstätte-adjacent comparator (another ornament-ground portrait) to test whether the chromatic-anchor signature is Klimt's or the genre's.

Deep-dive pass — external pressure joined

The criticism becomes the experiment

Same day, after study close: two reader comments and the published criticism were brought against the lab book — and each became a counter-test. The consulted sources (the Met's record and its Google Arts & Culture story, the Artists Network "Anatomy of a Painting" piece, HuffPost, Sartle) are remarkably unanimous: they read this painting through pose and gaze — the arm akimbo, the unsmiling direct stare, the Velázquez-infanta comparisons, the ~200 preparatory sketches for the stance — and subordinate the ground to "decorative surface." The instrument read it the other way around. Entries 16–23 put both readings under measurement.

Entry 16 — Audit of the “12×” figure

The ordering is unbreakable; the multiplier is not

The comment: the wall-vs-dress ratio is opaque without segmentation detail and easy to over-quote. The audit (audit_zones.py) shows the boxes as run, shows the I-4 masks themselves, and sweeps the configuration space: mass_quantile {50, 60, 70} × box jitter ±2% × an adversarial mask-field swap (the dress masked by its own most-chromatic pixels — the most favorable segmentation the dress can be given).

Entry 17 — Metric-design relativity, made operational

Instead of caveating the other narrative, build its instrument

The second comment: the "structure declines to weight her" reading is an interpretation dressed in numbers — a different metric design could produce a different, equally defensible narrative. Accepted, fully. The response is not a caveat: Entries 18 and 19 build the other instrument — a geometry channel (the critic's pyramid) and an address channel (the critics' pose/gaze) — and run them against the mass channel on the same pixels. The finding then stops being either narrative and becomes their measured divergence.

Entry 18 — The pyramid counter-test (geometry channel)

Both parties right, in different channels

The one formalist source (Artists Network) claims: that the carpet pattern and the figure together form a pyramid (in their words, the most stable design a portrait can have), with the composition anchored at the bottom edge by the three horizontal blue zigzag lines. That is a geometry claim; the lab book's diffuse_field / DGI-52 reading is a mass claim. pyramid_probe.py tests the geometry claim on its own terms:

Entry 19 — The address channel

The thesis, reframed on measurements from both sides

Every consulted source locates Mäda's command in stance and gaze — a channel the ISR cannot see. address_probe.py operationalizes the nearest pre-semantic proxies (the geometry of address, not gaze psychology): mirror-symmetry NCC about the best vertical axis (a frontal body is bilaterally symmetric; a turned one is not), axis position vs the frame axis, figure scale, face elevation.

Entry 20 — Wet-over-dry confound: refuted and refined

The D-break is the ground's facture, not session seams

The Artists Network technique note reports Klimt worked wet-over-dry in sessions, visible at the neck. Hypothesis: the structure-D native-resolution break (Entry 11) tracks those physical boundaries. neck_probe.py, matched bands at two scales: the neck/collar band has the lowest edge content of the four zones (0.197 native; cheek 0.265, dress shoulder 0.274, wall 0.242) and loses relative edge content at native resolution (ratio 0.93). The zone that balloons at native res is the wall control: 1.505×. The crop settles it visually — the neck/chest is the smoothest passage in the painting; the visible impasto is in the hair bow and the flecked wall.

Entry 21 — The Eugenia comparator: the inversion

Same painter, same family, opposite chroma policy

Klimt painted Mäda's mother Eugenia Primavesi in 1913–14 (Toyota Municipal Museum of Art) — the controlled comparator the study flagged as missing: same artist, same patron family, same two-year window, both ornament-saturated grounds, varying only the sitter. Reproduction: Wikimedia, "copied from an art book" — a weak chain, so claims here are directional; within-image ratios are chain-robust to first order (both sides of each ratio share the chain).

Entry 22 — The Flöge seam: authorship of chroma

Whose decorative labor got the colour

Per the Met's Google Arts & Culture account, the white dress was made by Emilie Flöge — Klimt's lifelong collaborator, chronically reduced in the literature to "muse" — at Klimt's suggestion; and Klimt added the band of coloured roses at chest height "to make the dress more colorful." Mapped onto the measurements: the fabric attributed to Flöge reads as the chromatic void (0.034–0.076, fabric-locked); the rose belt (0.256) and the scattered skirt ornaments the q70 mask finds (0.39–0.54) — the only chroma the figure carries — are on this account the painter's insertions into the designer's garment.

Scoped honestly: the account does not settle whether the added colour was sewn or painted, and the ground pattern is as much Werkstätte ornament vocabulary as the dress. What the pixels support is narrower and still sharp: within the figure, chroma lives only in discrete ornament points attributed to Klimt's intervention, while the garment attributed to Flöge is the void. The authorship-of-chroma question, whose decorative labor got structural priority, is thereby measurable in principle. [scope cut → Entry 42: a category error — who sewed the garment has no bearing on the chroma of paint; the narrow documented point above stands, the labor framing is retracted.]

Entry 23 — Loos, both readings, and the record

The numbers enter a fight that was actually happening

The Loos crossfire. Adolf Loos's Ornament und Verbrechen (written c. 1908, delivered as a lecture 1910, published in French 1913) attacked — in the same Vienna, in the same years — exactly the Werkstätte aesthetic this painting embodies: ornament as regression, a covering-over of the honest modern individual. The instrument can now put numbers on both sides of that period argument, and they refuse to pick a winner: in the mass/chroma channel the painting structurally enacts Loos's charge — the person is out-weighed and out-coloured by ornament, an order of magnitude in the fabric-locked read. In the address channel it refutes him — the person commands the frame axis to 0.1%, at 92% of frame height, frontality 0.71–0.86: a geometry ornament never touches. And the Eugenia comparator shows the efface/not-efface is a dial Klimt turned per portrait, not a doctrine of the circle. Strictly formal, per the house rule: this is about channels in a picture, never about the sitter.

Both readings stay open — deliberately. The measured seam (address maximal, chroma-mass displaced) supports "she is visually sidelined despite formal prominence" and "her presence transcends the decorative context" with equal ease. The instrument cannot adjudicate between them. That is stated here as the clearest available demonstration of the study's own limit — not buried in a caveat list.

Provenance — the source's own record. Reconstructed from the search index of the Met's record and corroborated against the restitution literature: Otto & Eugenia Primavesi → Eugenia alone (from 1926) → Hugo Bernatzik, Vienna (1928) → Neue Galerie, Vienna (1928, sold to Steiner) → Jenny Pulitzer Steiner (by 1937; fled Vienna after the Anschluss; the painting seized 1938; the family collection partly auctioned at the Dorotheum from 1940; restituted to her 1951) → her daughter Clara Steiner Mertens & André Mertens (1958–63) → Clara alone (1963–64) → gift to the Met 1964, "in memory of her mother, Jenny Pulitzer Steiner." The flagged timeline "discrepancy" (family sale vs 1938 seizure) dissolves — two different owners: the Primavesis had sold by 1928; the Nazis seized from Steiner. Unlike Steiner's Wasserschlangen II (never returned in her lifetime; 2013 heirs' settlement), Mäda's chain closes cleanly: seized, restituted to the victim herself, gifted by her heir in her name — the CC0 status sits on a contested-then-resolved chain, and the credit line carries the memory. Open: the Met page rate-limited automated access this session; the verbatim record should be confirmed when reachable (lostart.de also blocked; both Jenny Steiner and Clara Mertens have Lost Art person entries — other family objects remain listed as lost).

Entry 24 — Correction: the rose_belt box missed the garland

A reader box-check, and a number that gets stronger when fixed

A reader checked the probe overlay. 8 of 9 boxes were on target; the rose_belt box was placed too high — at y-frac 0.21–0.275, above the rose garland, whose chroma peaks at y-frac 0.317. It was reading the neckline trim and upper bodice, not the roses. With the box on the garland, colour reads 0.42–0.91 across segmentation (reaching the wall's level only at the favorable end) — not the 0.256 first reported. Per the house rule, the original value is kept in Entry 10's table (corrections are appended, not erased) and superseded here.

Deep dive 2 — the compositional maneuvers

Six theses, run in argument order; two tests refined in the doing

A second reader pass proposed six claims, each about a deliberate compositional move rather than a metric. They are run premise-first: a flat stage (6), then how figure and ground are built differently (1–3), then how each does a job the other won't (4–5). The figure is a void in the mass field, so its mask is dress-by-colour + head/legs-by-geometry — a reproduction-scale aid, verified against the pixels below. Two theses' proposed tests did not isolate what they meant; they are refined here, not buried.

Entry 26 — Thesis 6 (premise): flattened space

Does the floor recede, or stand as a pattern-plane?

If the floor recedes, the ground sits behind the figure and loses the structural fight automatically; the whole diffuse-field result would be downstream of one spatial decision. Banded front→back, the floor holds chroma cv 0.13, edge cv 0.14 — no monotone fade toward the back: consistent with a flat pattern-plane, corroborated by the zigzag being a frontal horizontal repeat with no vanishing convergence (Entry 18). The mark-scale proxy came back non-monotone and noisy — inconclusive, reported not relied on. The premise holds, weakest-supported of the six.

Entry 27 — Thesis 1: the void as a made object

Advances by light, recedes by mass — in absolute terms

The claim: the dress is engineered to pop and dissolve at once. The proposed test, within-silhouette r(luminance, mass), does not isolate it — r is negative everywhere (dress −0.24, figure −0.34, ground −0.44, whole −0.48), because smooth bright paint carries little mass anywhere. That is a global property. The real evidence is absolute: the dress sits at luminance percentile 0.75 (among the brightest in the frame) but mass percentile 0.28 (low) — an advance/recede gap of 0.47. The brightest major region and a below-average-mass region at the same instant.

Entry 28 — Thesis 2: two drawing grammars, one per region

Figure drawn in value, ground drawn in colour

Committed-content share (fraction above the frame's 85th percentile) per region×channel:

Entry 29 — Thesis 3: the rivet

One chromatic weld, at the waist

The figure-column chroma profile is a flat near-neutral line (median 0.087) with one dominant spike at the belt, y-frac 0.314 (mask-locked 0.42–0.91 across segmentation, reaching the wall's level at the favorable end [restated → Entry 42]). It sits 0.068 from the nearest golden section and well above the figure's centre of mass — at the empire waist, the pivot of a 92%-tall standing form. Uniqueness, honest: only 5 of 229 figure rows reach ground-level chroma, but the figure still reaches ~0.48 at the scattered skirt ornaments without the belt — so the belt is the dominant banded rivet with sparse secondary point ornaments (the "white void + sparse chromatic ornament points" of Entry 24). A full pull-test stays open.

Entry 30 — Thesis 4: the head turns inside the planted stance

Shoulders squarer than the face — the 200-sketch signature

Two independent methods agree in direction: the address-probe axis-sweep (Entry 19) read shoulders 0.856 > face 0.714; the fixed-centreline band sweep here reads shoulders 0.726 ≫ face 0.240. The square stance with the head rotated off the mask-axis is the oldest way to put life into a frontal figure. The full-body monotone gradient is not clean — below the waist the skirt is a featureless white field, so mirror-symmetry falls to ~0 for lack of bilateral structure (not asymmetry). Scope the claim to the upper body, where it is strong.

Entry 31 — Thesis 5: static figure, dynamic field

Remove the figure — the torque stays

The cleanest of the six. Subtract the figure (mask + fill) and recompute stance torque:

Entry 32 — The six as one argument

And a promotion candidate

On a flat pattern-plane (6) that lets the ground compete at all, Klimt builds the figure as a made void (1: brightest region, lowest mass), drawn in value (2), with one chromatic rivet at the waist (3), turns the head inside a square stance (4), and offloads all the dynamism to the ground (5). Five deliberate moves on one flattened stage. Strongest-supported: 5, 2, 1; refined in the doing: 1 (correlation→absolute), 3 (dominant not unique), 4 (upper body only), 6 (uniform, mark-scale inconclusive).

Promotion candidate — the address channel (C-4). Frontality + axis-lock + scale is a pre-semantic proxy for the "command" every critic cites — the one measure that reads the figure the way viewers do, the axis orthogonal to mass. On the Klimt it is at maximum while the mass channel is displaced to the ground. Filed to toolkit/CANDIDATES.md (n=1 here; run corpus-wide in Entry 33, promotion declined in Entry 34 — the masked and head-turn refinements did not close, a pre-semantic ceiling).

Entry 33 — The address channel across the corpus

Does C-4 discriminate? Cleanly — and it separates three ways of not facing the viewer

The address channel was the study's one measure that reads a portrait the way a viewer does, so it was taken across the whole corpus to see whether it discriminates. A corpus-general form (address_corpus.py, no per-work masks, all seven at 800px): central-column mirror-symmetry frontality, the strongest symmetry axis and its offset from centre, axis-lock. The result separates the works cleanly:

Entry 34 — C-4 closeout: the refinement bars fail, honestly

Both tested; neither closes; promotion declined — and the boundary is the finding

Two bars stood between C-4 and the spine: a figure-masked form (to resolve the Eugenia confound) and a corpus head-turn sub-signal. Both were built and run on all seven works with verified figure/face/shoulder boxes (address_masked.py). Both fail to close, for one principled reason — and the honest outcome is no promotion.

Bar A — the masked form adds nothing. Frontality on a verified figure box (box-centre, the honest metric; the best-axis variant is rejected — it handed the small Matisse face box 0.85 by cherry-picking a symmetric strip) barely changes the corpus ordering (Mäda 0.61, Matisse 0.27, Degas 0.13, Eugenia 0.11, Caravaggio 0.00). The confound posited last session was smaller than feared: Eugenia's low score is largely real — her figure, about its own axis, genuinely reads less bilaterally symmetric than Mäda's. The central-column form was substantially right, so masking earns no promotion.

Bar B — the head-turn is confounded. Shoulder-minus-face frontality is positive for all three portraits, but by different causes: Mäda a genuine mild rotation (validated twice, Entry 30); Eugenia an anti-symmetric face box (hair/lighting, face frontality −0.14); Matisse the deliberate face shadow (−0.27). Mirror-symmetry cannot tell "head turned" from "face asymmetric for any other reason" — a real signal on Mäda, but not a clean corpus discriminator.

Maneuvers 2 — six more theses on the figure/ground weld

One strong confirmation, one partial, two refuted-into-refinements, one history-limited, one declined

A further reader pass proposed six moves. Run honestly, they split: the akimbo aperture is a strong new finding; two refutations sharpen Entry 28 rather than complicate it; the two-rivet thesis half-holds and reframes; the commission-as-unit idea is real in the measurements but undocumented in the history; and the pad-and-rerun proposal is declined as confounded.

Entry 35 — #2 The akimbo aperture: the ground breaches the figure

The gesture of command is where the decorative field gets inside the outline

Every critic reads authority in the arm akimbo — which encloses a triangle of negative space between forearm and torso. Verified against the pixels first (the pose could have closed it; it does not — the lavender wall and floor blooms are plainly inside the gap). Measured (convex hull of the figure minus the figure, at the akimbo region): the aperture reads chroma 0.539 = 86% of the wall, and 13.3× the dress. The single gesture the literature reads as command is the one place the decorative field passes inside the figure's convex hull — at the elbow, the instrument's own densest edge-knot (Entry 02). The competitor is admitted through the pose that supposedly dominates it.

Entry 36 — #3 The zigzag is not bilingual — which sharpens Entry 28

Only the wall is chroma-drawn; everything else is value

The claim was that the blue zigzags translate between the two grammars, committed in value and chroma at once. Refuted — the zigzag is the most value-committed region and among the least chroma:

Entry 37 — #4 The currents co-flow, not counter-flow

Chroma and darkness both concentrate in the upper wall

The proposed motor — chroma ascending while dark descends, crossing at the figure — is refuted at the mass level. Chroma is higher at the top (0.535 vs 0.312) but so is darkness (0.469 vs 0.349); the two are positively correlated, r +0.85 over n = 18 frame bands [disclosed → Entry 42: with both variables monotone-ish in height, the r is weak evidence alone — the directional statement carries]. One gradient (upper wall dark + chromatic, lower floor bright + pale), not two opposed ones. The "dark descends" intuition was about Entry 07's sparse floor accents, not tonal mass.

Entry 38 — #6 Two facture welds, but not borrowed from the wall

Belt and bow are raised events; the wall is smooth

Facture (fine high-frequency energy): the bow at the crown reads 0.043, the belt 0.048 — both about 1.5× the smooth dress (0.028). So the figure carries two raised-facture welds, at waist and crown, and the belt is a double rivet (highest chroma AND facture). But the reader's mechanism half-fails: the wall is smooth too (0.029 ≈ the dress), so the figure does not "borrow the ground's facture" — the welds are its own raised events, not imports.

Entry 39 — #1 the commission as unit, and #5 declined

A measured opposition that history does not (yet) confirm as designed

#1 — the pair as the object. The opposition is real and double-supported: Entry 21 (mother most-chromatic / daughter least) and Entry 33 (Eugenia lands next to Pollock on address, nowhere near her daughter) — two independent channels, the same maximal split, one painter, one family, a two-year window. Secondary sources call Eugenia an "elegant counterbalance," and the commission was sequential same-family (Eugenia commissioned Mäda's portrait in Feb 1912; Otto invited Klimt to paint Eugenia in 1913; both later at the Winkelsdorf Hoffmann villa). But no source documents an intentional pendant or a shared hanging — the "counterbalance" is a retrospective framing, and the canvases are now separated (Met + Toyota). Honest verdict: the opposition is measured and striking; the designed diptych is unconfirmed, and intent is the instrument's stated limit. The strongest defensible claim is that the two canvases are measured poles — neither alone is "the Klimt formula."

#5 — pad the frame and re-run: declined as confounded. Synthetic padding invents pixels; whatever fills them sets the result, and adding non-figure area trivially lowers the figure's mass fraction and pulls every centroid outward — the verdict would flip for any painting, telling us about the padding, not Klimt (the I-13 fill-invariance problem generalized). The honest insight underneath — framing is compositional; the 92%-height crop denies the ground room to recede — is already covered by I-10 (crops are different measurements) and is a measured term in the address channel (Mäda's 0.92 height fraction). No confounded experiment added.

Entry 40 — Critique rewritten at study close (v2, essay form)

The painting-first essay ships; the lab book stays the instrument's record

Per house convention (Studies 4 and 5 both rewrote at close), the critique was rewritten painting-first as klimt_critique_standalone.html (14 sections, 10 figures, 3 tables): cold-reader orientation (Klimt, the Primavesi commission, the Flöge dress), the painting as the first figure, the analysis carrying the story — the diffuse-field paradox, the made void, the stance quantified against the corpus, who holds the colour and the grayscale collapse, the two grammars with the zigzag correction, the rivets and the Flöge seam, the akimbo aperture, the division of labor and the pyramid adjudication, the Eugenia pair — closing through Optional interpretation (Loos, both readings held open), Public frame, What This Does Not Prove (including the corrections record), and Sources. Corrected values throughout; claim grades carried inline; Met dimensions verified verbatim against the object record. v1 marked superseded and kept as record.

Entry 41 — Adversarial pass on the critique (v2.1)

Nine overclaims fixed, three omissions built — the Study-4 precedent applied

The essay was re-read hostile, claim-by-claim against the measurement record. Fixed: the thesis line "the movement lives in the floor" (contradicted our own Entry 37 — it lives in the ground, peaking in the upper wall); "wallpaper" ×4 (an identification the instrument cannot make — now "wall/field," with the limit stated); a garbled five-rows sentence (all five ground-chroma rows are belt rows; the ornaments' ceiling is 0.48); the Flöge claim scoped to the belt (the only insertion the record documents); "welded at three points" corrected to welded at two, breached at a third; the aperture/edge-knot conflation ("the same arm draws both"); the golden-section proximity quantified (0.07); "his only major child portrait" attributed to the literature; and a components/islands conflation separated. Built: the missing punctuation section (the two accent systems as two mark-languages, lens-cap caveat carried); a poles placement paragraph (Degas dissolution by the opposite route, Matisse's eye/weight split, first high key); and R_spatial in the division of labor ("no region owns the movement — the arrangement does"), graded [aid] with its candidate status declared. The critique is v2.1: 15 sections, 11 figures, 3 tables.

Entry 42 — Second reader review: six challenges, run

Conceded where wrong, tested where testable — the strongest numbers survived

A second review hit the lab book itself (2026-07-08). Every point was conceded and fixed, or turned into a measurement (review_fixes.py). The inline [superseded / restated → Entry NN] markers now scattered through earlier entries are this review's mechanical companion to the append-don't-erase rule.

1. The belt double standard — conceded. Entry 24 quoted the top of a 2× range and matched it to the wall, exactly what Entry 16's discipline forbids. Restated everywhere: the belt reads 0.42–0.91 across segmentation, reaching the wall's level only at the favorable end, roughly half at the worst. The rivet thesis survives in that form.

2. Entry 35's ratios used unnamed denominators — conceded and re-based. The 86%/13.3× were the probe's own local box-means (wall 0.626, dress 0.041); against Entry 10's mass-locked values the same read is 62% and 7–16×. The v1 overlay also missed the elbow crook and spilled below the arm (reader catch). Re-run on a geometry-only mask, tightened:

Expanded aperture reading (reader observations): both hands are hidden — the akimbo hand disappears behind the hip, the other arm behind the back; the pose read as command shows no hands, an iconographic tension the instrument cannot adjudicate. The "darker violet behind the wrist/hip" localizes to the aperture interior; around the outer silhouette the wall actually pales (ring +0.03 brighter, −0.05 less chromatic than the distal wall) — a light desaturated halo, same push-forward effect, opposite mechanism [aid]. The sleeve rosette with its two green leaves sits exactly on the elbow — an ornament planted on the instrument's own anchor knot; the reader's larger chain (elbow → belt garland → skirt ornaments → the floor's strewn islands) is consistent with the island map and logged as a compositional reading [aid].

3. The deep-dive-2 mask was circular — conceded and re-run. A dress mask built from low chroma cannot testify to low chroma. Rebuilt as a hand-drawn geometry-only polygon (no colour input; two variants; overlay verified; known slack at hair/shoulders makes these conservative bounds):

4. Was Entry 27 tautological? Answered with the corpus. Disclosure first: mass = 0.42·edge + 0.32·tone + 0.26·chroma, so a chroma void is mechanically a quarter of the way to a mass void; the edge+tone 74% carries the reading. Then the comparator test — the same bright-vs-massless measure on the corpus's other smooth bright passages:

Bright-but-massless is not a general property of smooth bright paint: the corpus's other candidates sit at zero or negative gaps (Caravaggio's bright smooth wall is that painting's anchor). The Klimt gap is unique in sign and size; the informative claim is restored.

5. Correlations without n — conceded. Entry 37's r = +0.85 was over 18 frame bands with both variables monotone-ish in height: retired to the directional statement, n disclosed. Entry 18's single template NCC swept over apex height × base width (nine configs): mass~pyramid −0.29 to −0.32, robustly negative; edge~pyramid −0.00 to +0.04 — not robust, withdrawn. The geometry claim's only surviving support is the base-securing zigzag seam; the mass refutation strengthens.

6. The Flöge framing was a category error — conceded and cut. Who sewed the garment has no bearing on the chroma of paint. The "whose decorative labor got structural priority" framing is retracted (marker at Entry 22); the narrow documented point stands.

Errors fixed everywhere: "the library's first modern work" was wrong by our own corpus (Matisse 1916, Pollock 1952 precede it) — now "first high-key work"; 436819 is the object ID, the accession is 64.148 (verified; it encodes the 1964 gift, as the reviewer inferred); the public-domain rule was misstated (life + 70 in 2026 = dead by 1955, not "the late 1910s"; Klimt qualifies either way); Entry 18's ~40 verbatim words from the Artists Network piece paraphrased; stale captions now carry inline supersession markers. The per-entry epigram habit is accepted as a discipline going forward. And the largest clarity gap is closed below: a glossary, including what "colour" actually is.

Entry 43 — Third review: two experiments, one concession, one clean pass

The converging channels were mostly one channel; the mass formula was doing part of the work

A third review found the sharpest structural problem yet, and it is correct on the code: mass = 0.42·edge + 0.32·tone + 0.26·chroma, with edge and tone from luminance; the grayscale persistence variant replaces every channel with luminance, so it zeroes only the 0.26 chroma term. And stance torque runs on luminance, not the mass field. Three experiments followed (experiments_r3.py). The contested values are now collected in claims.json as a reconciliation ledger, hand-synced against both documents this pass; auto-rendering the two documents from it (so a number cannot drift between surfaces) is the outstanding engineering item the review flagged.

A · Grayscale persistence across all seven works. The claim “most portraits are value structures that survive loss of colour” was asserted on no data. The corpus:

B · mass′ = 0.57·edge + 0.43·tone (chroma removed) — the concession. Does “mass migrates up into the wall” survive removing the chroma term? No. The raw mass-field centroid moves from ny −0.074 (up) to +0.011 (centred), and the inner-disk mass fraction rises 0.176 → 0.215 (edge and tone concentrate on the central figure). The mass-displacement headline is chroma-driven — “mass up into the wall” is substantially “chroma is in the wall” restated, and it is estimator-dependent besides (island ny −0.174 vs kernel δy +0.093, Entry 02). Demoted to a colour claim. What is robust stands unchanged: colour is displaced into the wall (β 0.378, the zone probes, the colour centroid) — that was always the real finding. [conceded]

C · Ground ablation, same frame, biharmonic inpaint. Entry 31 deleted the figure but never the ground in the same frame (the figure-crop was a different frame; torque is frame-dependent). Now, inpainted, no crop, no flat fill:

The architecture reframe (the review’s central point, accepted). β, CBI, coupling, ω, and the mass-displacement metrics are algebraic transforms of the same two fields (luminance, chroma) over the same regions — one observation in several costumes: the wall is smooth and chromatic; the dress is smooth and neutral. “Independent channels converge” (Entry 09) is withdrawn. The study’s genuinely independent lines are five: (1) the address channel (luminance mirror-symmetry), (2) facture (high-frequency luminance residual), (3) ground-dynamism / torque (luminance, ablation-confirmed), (4) the Eugenia comparator (a different painting), (5) grayscale persistence (corpus-relative). The chroma cluster is one of these, stated once, not four. Downstream corrections (Entry 09 callout, Entry 18 caption/verdict, Entry 24 body, Entry 33 callout, Entry 32 marker, the limits list, DGI weights, “said plainly”) are all folded in via claims.json.

Glossary — every number, operationally

All fields are computed from the reproduction decoded as sRGB and are robust-normalized per image (1st–99th percentile → 0–1) unless noted. Consequence: every value is a position within this reproduction's distribution, and every chroma ratio inherits the museum photograph's white balance and ICC handling — a named, uncontrolled error source on top of Study 4's reproduction-variance bar.

What This Does Not Prove

• Not intention. The split between centrality and weight is measured; its purpose is not.
• Not plural construction. construction_diversity 0.030 is below the 0.12 flag: the figure/ground split is register-level, not a region-by-region difference in how paint is built. Matisse, not Klimt, owns that axis in the compare.
• One reproduction. Crisp decimals carry an unmeasured reproduction-variance bar (Study 4); the compare mixes three reproduction chains and one crop, n=3.
• Mode- and magnitude-sensitive values are labeled where they occur: coupling's sign is noise at |0.02|; island entropy is an edge-mode range (0.63–0.82); the elbow anchor is the instrument's densest edge-knot, partly method-shaped (the Study-4 anchor lesson); structure-mask D is quoted only in its stable window (~1.53 at 512–2048px).
• R_spatial is a candidate (C-3), not a shipped metric — the range, not the point, is the finding, and the measure has not yet earned the spine.
• The "12×" is retired as a bare soundbite (Entry 16): quote the ordering (unbreakable, worst case 1.3×) and the fabric-locked range (9–27×), with the coverage signature.
• The address channel is a proxy (Entry 19): mirror symmetry, axis lock, and scale measure the geometry of address, not gaze or psychology — "command" remains a human reading.
• The Eugenia comparison is directional (Entry 21): art-book reproduction chain, 1072px; within-image ratios only. A museum-grade reproduction would be needed for decimals.
• The Flöge framing was retracted (Entry 42) as a category error — who sewed the garment has no bearing on paint chroma; only the narrow point stands.
• Mass is 26% chroma by weight (0.42·edge + 0.32·tone + 0.26·chroma), so grayscale persistence and mass-displacement are partly formula-mediated — Entry 43 gives the corpus baseline and the mass′ recompute that demoted mass-displacement to a colour claim; torque, by contrast, is luminance-based and chroma-independent.
• The provenance chain (Entry 23) awaits verbatim confirmation against the Met record (rate-limited this session) and the Lost Art entries.

Sources & References

Primary object: The Metropolitan Museum of Art, object 436819, accession 64.148 (Mäda Primavesi, Gustav Klimt, 1912–13, oil on canvas), Open Access (CC0); object record, image DP243354, and public-domain status confirmed via the collection API at read time. Comparison sources: the library's pinned study reproductions of Matisse La Leçon de piano (Study 1) and Degas Dancers, Pink and Green (Study 2, cropped source).

• The Metropolitan Museum of Art. Object record 436819 (accession 64.148), incl. sitter identity (Mäda Primavesi, 1903–2000), dating, credit line, and provenance. Open Access, metmuseum.org. (Verbatim provenance confirmation pending — automated access rate-limited this session; chain corroborated via the record's search index and the sources below.)
• The Met / Google Arts & Culture, "Klimt's Portrait of Mäda Primavesi" — the Flöge dress account (made by Emilie Flöge at Klimt's suggestion; the rose band Klimt's addition) and the preparatory-sketch history.
• Artists Network, "Anatomy of a Painting: Gustav Klimt Captures Mäda Primavesi" — the pyramidal-composition and zigzag base-securing claims tested in Entry 18, and the wet-over-dry technique note tested in Entry 20.
• Adolf Loos, Ornament und Verbrechen (written c. 1908; lectured 1910; French publication 1913) — the period argument Entry 23 puts the measurements into.
• Jenny Steiner: Wikipedia biographical record (née Eugenie Pulitzer, 1863–1958; collection confiscated 1938, Dorotheum auctions from 1940); the Wasserschlangen II restitution literature (2013 heirs' settlement); Lost Art database person entries for Jenny Steiner (541155) and Clara Mertens (541626), automated access blocked.
• HuffPost and Sartle commentary — the pose/defiance readings triangulated in Entries 17 and 23.
• Rudolf Arnheim, Art and Visual Perception (California, 1954/1974) — perceptual weight and the centric/eccentric composition frame the mass-vs-attention reading borrows.
• Instrument lineage and prior studies: the ISR library lab books for Matisse, Degas, Pollock, Caravaggio, and Soejima; instrument findings I-1…I-14 and methodology in OBSERVATIONS.md; the program's published papers at parallaxmetrology.com.

Computational results in this lab book are deterministic and reproducible from the scripts named in each entry (environment: /usr/local/bin/python3, Python 3.14; instrument installed editable from instrument/). Charts are generated by make_charts.py from the study's JSON artifacts. Art-historical context draws on the sources listed here. Core instrument untouched.