Pieter Bruegel the Elder — Landscape with the Fall of Icarusoil on canvas, Brussels (RMFAB), attributed / contested as an early copy after a lost original c. 1558 · public domain

Entry 01 — Selection & condition check

A construct the corpus did not have

Six studies in, the library had never read a work that is distributed and ordered at once. That is a real gap, because it hides a distinction the instrument had never had to make: distributed order versus dissolution. The two best prior studies (the Degas abonné, the Matisse "the eye goes where the structure isn't") were both the instrument finding a displaced or buried subject, and Icarus is the canonical buried subject in Western painting: the nominal hero is a pair of legs vanishing in a corner while a ploughman and a ship own the frame. There is a measurable, falsifiable proposition here (does the composition structurally marginalize Icarus, and by how much?), tested against a real critical claim (Auden's Musée des Beaux Arts).

Entry 02 — The headline

Counterweighted, not dissolved: distributed order

Full spine read (isr read, ~33s no-persistence). The state classifier returns counterweighted (flag counterweight_present: "a region reads as a load-bearing balance term more than a construction outlier"), and the regime block reads gradient-field / controlled instability, distributed, construction gradient-built. This is the crux of the selection: the picture is as busy as a dissolution work (20 regions, 10 active islands, island entropy 0.85, near Degas's), but it is ordered, not dissolved.

Entry 03 — The island system

The anchor is the ship; the balance is a triangle across a void

Read against the reproduction (edge-aware). The single anchor is island 8, the galleon (orange node, right), not any human figure. The resistance nodes (red) are the upper-left cliff/trees/city (15), the ploughman/horse/dark-earth foreground (14), and the sheep band (17). The load-bearing balance is a triangle: cliff (15) – ship (8) – plough (14), spanning a 72% void of sea and sky. The Icarus corner registers only as satellite nodes (12, 5, violet): peripheral, non-load-bearing.

Entry 04 — What the fields see

Mass on the foreground, colour in the void

The field panels corroborate the island read and split the labour by channel. Mass is hot across the entire foreground band (trees, ploughman, horse, sheep) plus the ship's rigging; sky and sea are cold. Colour evidence is brightest in the sky and water and dark on the figures: colour operates in the void (β 0.234). Edge lives in the left trees, the figures, and the rigging. Header: persistence 0.65, fragile_to = blur, i.e. the composition depends on its fine edge detail (a distributed-order signature; Klimt by contrast was fragile_to grayscale).

Entry 05 — The marginalization battery, leg 1

Icarus is NOT marginalized by intensity (the false null)

icarus_probes.py reads every named object over its own mass (the I-4 honest zone read: loose boxes stop lying) on four channels: structural mass, saliency (recomputed as a map, mirroring the spine's spectral-residual saliency), colour, and edge. Boxes verified against the pixels first (scout_probe_boxes.png). The naive intensity comparison returns a false null: per pixel, Icarus is not marginalized at all.

Ratio convention, here and in the jitter table below: Icarus ÷ ploughman throughout, so above 1 always means Icarus reads higher (a review caught the first version switching direction mid-entry).

Jitter check (added after review, re-run normalized, jitter_probe.py). The comparison rests on small differences, so a fair objection is that the sign might flip if the hand-placed boxes move. First version jittered both boxes ±20px absolute; a second review noted that is ~19% of the small Icarus box's width but only ~5% of the big ploughman box's (the two objects received non-comparable perturbations), so the sweep was re-run with jitter normalized to each box's own dimensions (±7.5% and ±15% of box width/height, 25 positions each). Ratios below are Icarus ÷ ploughman:

The load-bearing claim survives box-proportional perturbation: Icarus is the sharper mark in 24 of 25 positions (median 1.09×), at least the ploughman's equal in salience (20/25, median 1.02×), and never dimmed on mass (the ploughman leads by 1.1–1.2×, an order of magnitude short of the extent ratios in Entry 06). The edge result carries the "not by intensity" claim; the saliency margin is directional [stable, jitter-checked, normalized].

Entry 06 — The marginalization battery, leg 2

Extent and placement: the protagonist at background-incident scale

Integrate the same channels over the frame instead of per pixel. Icarus's legs hold 0.18% of the frame's total structural mass and 0.22% of its saliency. The obvious objection, raised in review: a pair of half-submerged legs is smaller than a man plus a horse plus a plough in any painting containing both, so the raw ratio to the ploughman (25×) is partly a fact about legs, not a compositional decision. The honest control is the corpus within the painting: is Icarus rendered at foreground-actor scale or background-incident scale?

Taxonomy scoping (same status as Entry 19's gaze). The three-tier assignment is a hand-made semantic judgment, not an instrument output: the instrument produces the frame-shares, and a person sorted them into actors, incidents and details. The consequential edge case is the fisherman: a worked human figure in the lower band, classifiable as a foreground actor. Reclassify him upward and "below every foreground actor" has a floor of 2.1×, not 6.8×. The claim that does not depend on the taxonomy at all: Icarus's frame-share is 6.8–35× below the five large named objects and 2.1× below the smallest named human figure, at full local intensity (Entry 05), with the mass centroid on the opposite side (Entry 07).

Entry 07 — Mass grid & tonal planes

Weight bottom-left, hero-corner empty, deep-space recession

The 3×3 mass grid confirms the displacement quantitatively: weight pools bottom-left (bot-center 14.5%, mid-left 14.6%, bot-left 14.2%) and the Icarus corner (bottom-right) is a low-mass zone (10.2%, void ratio 0.47, one of the emptier foreground cells). Left 41% vs right 28%, bottom 39% vs top 25%. And the tonal-plane independence is maximal (TLG = 1.0): the quadrants are fully differentiated in value, bright sun/sky (zone 5.4) against dark plough-earth (zone 2.4). That is the deep-space tonal recession the corpus had not seen: order carried by planes, not by a single tonal hierarchy.

Entry 08 — Two accent systems (I-5 / I-7)

The default lens floors at zero; the colour lens finds the incident field

As predicted in the condition check, the default structural accent lens returns accents = 0 ("no high-leverage accents survived the current threshold"). This is the Degas I-5 pattern in a new key: a densely and uniformly worked surface where no single mark clears the leverage floor. Switching channels (I-7) recovers the real system: the coalesced chromatic lens finds 17 accents (top leverage 1.06), distributed across the ploughman, shepherd, sheep, sails, harbour and ships. Reported as it is: a distributed accent field, not a few loud marks. Accent counts are lens-capped and are never abundance measures (Study 5 lesson); what is claimed is the channel segregation and the distribution.

Entry 09 — Attention vs weight, and the null-control (I-15)

The eye/mass split is real composition; β is not

Saliency and structural mass diverge (mass_divergence 0.077, Klimt-like 0.080; the Degas coincided at 0.02). Both pull down-left toward the ploughman, neither toward Icarus. Before trusting that as a fact about the arrangement (I-15), it was run through the matched-stats null battery (null_control.py: phase-scramble = same histogram + power spectrum, zero arrangement; patch-shuffle = same histogram + local texture, no long-range composition):

Entry 10 — R_spatial (candidate C-3)

Subadditive: the calm whole is a cancellation of louder locals

The candidate battery (AGENTS step 7, appended to the register in the same commit): R_spatial 0.23–0.52, subadditive, cut-stable. The global torque is a cancellation: many local torques (each incident tilts its own patch) that partially cancel into a calm global balance. This is the opposite of Klimt's emergent superadditivity (2.91–5.04), and it fits the compare (Entry 11), where Bruegel has the lowest global torque of the five.

Entry 11 — Same-coordinate compare

The deep-space plane-stacker, placed against the poles

isr compare against four poles (Degas dissolution, Pollock all-over, Klimt address, Caravaggio tenebrism) in one coordinate space, on each study's pinned reproduction. A note on how to read this, prompted by a reviewer: at n=5 with a sample SD, the maximum possible |z| is √(n−1) = 2.000 exactly, so a "+2.00" is not "two sigma," it is the arithmetic ceiling, and it means only that Bruegel is the sole outlier while the other four are effectively tied. At this n the recoverable signal is the rank, not the magnitude. So this table reports Bruegel's raw value, its rank of 5, and the spread of the other four, and drops z entirely:

What the rank presentation demoted, said plainly. The switch from z to raw-plus-rank quietly downgraded one first-pass "finding": grid_center_pct was reported as "−1.96, alone," which sounds like an extreme; in raw values it is 10.03 against 11.85–12.32, real but small, nothing like the 13–59× separation on lg_score. That is the rank presentation working as intended: the ceiling-z hid the difference between a dramatic outlier and a mild one, and the raw spread shows it. Logged as a point in favour of the change.

Entry 12 — Edge-mode check (step 4)

Paint to all edges, so raw is the honest comparison

The capture has paint to all four edges (no frame or mat), so the safe zone protects nothing and edge-aware down-weighting could suppress real figures near the edges (I-1). Both modes were run; island diagnostics report edge_artifact_risk low, crop_stability stable, so contamination is minimal either way. The counterweighted verdict and the cliff–ship–plough triangle hold in both modes (the raw readout is kept as islands_raw.png for the record).

Entry 13 — Synthesis

Distributed order vs dissolution: the distinction made

The selection question is answered. Given a work that is busy in depth, the instrument does not collapse it into the dissolution regime. It reads the busyness (entropy 0.85, 20 regions) but organizes it: two load-bearing counterweights, a mass gradient pooled bottom-left, fully differentiated tonal planes (TLG 1.0), a low, cut-stable global torque, and colour coupled to form. The same-coordinate compare places it far from Degas and Pollock and beside Caravaggio on the pre-modern colour-structure coupling. Distributed order is a distinct regime, and the instrument finds it.

On the art-historical proposition: yes, the composition structurally marginalizes Icarus, and the measurement says how. Not by dimming (the splash is full-intensity, above the ploughman on edge and saliency) but by scale and placement: 0.18% of the frame's mass, the ploughman 25× more, the mass centroid on the opposite side of the picture. The instrument that located the Degas abonné and the Matisse displaced subject measures, here, exactly how the protagonist is disappeared: the mark is kept and the room is taken. [superseded → Entry 20: the corrected form of this sentence lives there]

First-pass close 2026-07-09. Open threads carried to the critique and to OBSERVATIONS: the I-10 resolution sweep for structure-D (native craquelure); zone probes on the harbour, sun and repoussoir for the C-1 depth blind-spot re-test; the delta/NCD authorship pass against the Van Buuren panel version (the copy question); and a repoussoir-depth probe (does the dark foreground read as 'in front of' the deep space, or does the instrument flatten it?).

Entry 14 — Deepening pass: I-10 resolution sweep

The window holds, and the break is at the BOTTOM this time

resolution_sweep.py, 512→5200px resampled from the 12265px source (native 98MP is memory-prohibitive and the craquelure that motivates the sweep is already present at 5200; the claim window is stated as 512–5200). dark-mask D 1.823–1.862 across the whole sweep: sampling-robust, comparable across works, and it places Bruegel in the same dense-fill family as Degas (1.86) and Pollock (1.88). structure-mask D plateaus at 1.64–1.66 across 1536–5200px; the pinned 2600px value (1.648) sits mid-window.

Entry 15 — The C-1 re-test: depth in the friendly case

Even global aerial perspective is not one gradient

C-1 (Study 2, four attempts) established the instrument cannot detect a LOCAL depth event (the Degas repoussoir nexus) without semantics. Bruegel is the friendly case: GLOBAL aerial recession across a deep landscape. The narrower question (depth_probe.py): is the depth ordering legible in the fields as coordinated monotone gradients? Four verified clean-water bands, near → far (boxes audited against zoomed crops; the near-mid band was tightened off the shore bank and off Icarus's flailing limb):

Entry 16 — The versions: layered delta with an internal control

Different pixels, same spatial skeleton

The Brussels canvas is contested as an early copy after a lost original; a second period version (the Van Buuren panel, with Daedalus still in the sky) survives. During sourcing, a third reproduction turned out to be a museum phone photo of the Brussels canvas in its frame (samsung EXIF, glare, slight skew), kept deliberately, because it turns the pair into a controlled three-way: Brussels-GAC ↔ Brussels-phone measures pure reproduction-chain distance on the SAME object; Brussels ↔ Van Buuren adds the object difference on top. isr delta (substrate NCD / tone / 16-axis form):

Review passes (2026-07-09)

A reviewer's challenges, run

Close readings of the first passes flagged the corrections folded in above (the DGI verdict bug and the N=100 percentile upgrade, Entry 09; the normalized jitter check and ratio convention, Entry 05; the tier-claim correction and taxonomy scoping, Entry 06; ranks over z-scores, Entry 11; the matched-resolution delta, the dissolved control, and the dendrochronology, Entry 16) and prompted three additions the study had missed or under-served: the compositional armature the house runs on every study, the sun, and the witnesses' orientation. (The former "deepening synthesis" entry is merged into Entry 20; a review noted it duplicated the final synthesis.)

Entry 17 — The compositional armature, with a positive control

Long lines, no convergence, and a control that had to be corrected

A genuine gap: the house runs a line-convergence read on every study (Degas's tight 22-line node, Matisse's grid, Klimt's short-line node), and Study 7 skipped it, for a landscape whose reputation rests on structure. Run (armature_probe.py, probabilistic Hough on committed edges, multiple runs for I-8): ~20 genuinely long lines per run (≥12% frame height), diagonal-dominated. Orientation bins, disclosed after a reviewer asked: horizontal = within 20° of level, diagonal = 25–65° or 115–155°, vertical = 70–110°, with the border zones (20–25°, 65–70°, 110–115°, 155–160°) left unbinned. The orientation share is itself an I-8 quantity, so it is read from a frozen 40-run measurement (armature_stats.json, generated once by armature_stats.py) rather than quoted as a decimal. This retires two single numbers from earlier passes: the "49%" of the first three revisions and the "39.5%" of the first exact-count audit were each one invocation of a stochastic estimator, and a fresh 40-run sample shifts even the range itself (this record: diagonal spanning 10–71%), which is exactly why the bounds cannot be the claim. What is stable across resampling, and what the entry rests on, is the median and the ordering: diagonal median 36% against horizontal median 5% (7×), the horizontal share never exceeding 17% in any run, and nearest-bin reassignment of the border zone unable to flip the ordering (most border lines are near-vertical, 65–70°, not near-horizontal). The control is the clean inverse and invariant: strict-diagonal 0% in all 40 runs, horizontal 88–94%. The intersection-centroid of the long lines locks onto nothing: across 10 runs per threshold (armature_node_spread.json) it wanders over y 0.24–0.61 of the frame (range 0.37).

Entry 18 — The sun: scoped by its own baseline

The agent registers on colour and nowhere else, which is partly a fact about disks

In the myth the sun is the agent: it melts the wax. In the object battery it reads last on saliency (0.026), edge (0.115) and mass (0.319), registering only on colour (0.684). The first version of this entry called that "the buried agent" and stopped. A reviewer pointed out the confound, which is Entry 06's in a new suit: the sun is a smooth low-contrast disk on a smooth hazy ground, and low edge/saliency are properties of that geometry, obtainable from any painted sun. So the baseline was run (sun_baseline.py): what do the picture's other smooth regions score on the same channels?

Entry 19 — The witnesses' orientation (scoped, not measured)

Where the instrument stops

The essay wants to say the shepherd gazes up at an empty patch of sky and the fisherman does not turn toward the splash: the human core of the "everything turns away" reading. It is worth stating explicitly that the instrument does not measure this. Gaze direction and figural facing are exactly the pre-semantic ceiling C-4 named on the Klimt: mirror-symmetry can register that a figure is bilaterally frontal, but it cannot read where a head is turned or where eyes point without knowing a face is a face. So the witnesses' inattention is carried in the critique as visual observation, flagged as not instrument-measured. What the instrument does contribute to that reading is one register up, and it is measured: the saliency field, the picture's own distribution of visual pull, points away from the Icarus corner (Entry 09, 100th percentile against both nulls). The figures' inattention is a human reading; the composition's is a number.

The scholarship adds a sharper version of the shepherd observation, logged as an open thread because it is measurable without gaze detection: the literature reads the shepherd's upward stare as aimed at the Daedalus that survives in the Van Buuren version, i.e. the Brussels canvas contains a gaze vector whose target was deleted. Taking the scholars' gaze line as given semantics, the instrument could ask what sits at its terminus in each version (nothing vs Daedalus): a registered cross-version probe of a compositional orphan.

Entry 20 — Synthesis

What the passes add to the study

The deepening pass certified the numbers (both D values from stable windows), converted the C-1 blind spot from anecdote to mechanism (depth carriers are plural, their binding semantic), and gave the copy question its first instrument statement (one skeleton, two objects, upper-edge same-work, scale-robust, with the layering caveat). The review passes hardened the spine: the DGI verdict corrected and then settled at N=100 (composition-dependent, percentile 1.0/0.0), the intensity comparison defended under normalized jitter, the tier claim restated at its own numbers (6.8–35×, bottom of the incident tier) with the taxonomy scoped as hand-made, z-scores demoted to ranks, the armature control corrected by its own intersection record (grid-crossing stability validated, VP detection retracted, the no-rectilinear-structure finding kept as recovery of the documented Weltlandschaft mode), and the sun scoped by its own smooth-region baseline to a colour-only, Brussels-only fact. The headline survives every correction applied to it, in a form that no longer overstates its own table: distributed order; a protagonist rendered at full local intensity and given an incident's frame-share, at the bottom of the incident tier, with the weight and the picture's own attention placed on the opposite side; in a composition whose skeleton also survives in its sibling version.

What This Does Not Prove

• Not intention. The displacement of weight from Icarus to the ploughman is measured; its purpose is a human reading.
• Not authorship. The delta pass (Entry 16) reads one shared skeleton at same-work distance; it attributes no hand (the Van Buuren panel is dendro-dated 1577+ and definitionally not by the Elder). The versions' local differences (Daedalus, the sun, canvas vs panel) are below the form layer's granularity, and the Van Buuren reproduction is 1772px on an unknown chain. The chain-vs-object layering holds only at the pinned scale (Entry 16's matched-scale caveat).
• Condition chain. The Brussels surface is, by the museum's own account, heavily overpainted after transfer questions and at least two relinings (the transfer-from-panel reading itself is disputed: Currie & Allart 2012 discount it). Facture-tier numbers (structure-D, edge means, fragile_to=blur) are therefore partly measurements of restoration campaigns; skeleton-tier metrics (islands, mass grid, TLG, counterweights) are predicted restoration-invariant, and the phone-photo delta (form 0.057–0.060 across a wholly different chain) is consistent with that prediction, but it has not been tested against documented restoration maps.
• Size is confounded; the claim does not rest on it. A pair of half-submerged legs is small in any painting. The corrected tier claim (Entry 06): Icarus is 6.8–35× below every foreground actor, at the bottom of the incident tier with the fisherman (2.1×), an order of magnitude above the true background details, at full local intensity (Entry 05), with the mass centroid opposite.
• The tier taxonomy is hand-made (Entry 06): the instrument produced the frame-shares, a person sorted them. Reclassify the fisherman as a foreground actor and the "below every foreground actor" floor is 2.1×, not 6.8×.
• One object's reproductions, pinned scales. Numbers are pinned to bruegel_working_2600.jpg. D values are quoted from their swept stable windows (Entry 14): dark-D 1.82–1.86 (512–5200px), structure-D 1.64–1.66 (1536–5200px); native 12265px remains unswept (memory), stated as the window's edge, not a claim.
• The depth probe is a proxy battery (Entry 15). Aerial fade, plane stacking and diminution are each measured; "depth" as a bound percept is not, and the monotonicity statistics ride on four hand-verified bands (n=4, Spearman on ranks: directional, not significance).
• Null verdicts are now percentiles, at 1300px (Entry 09, N=100 per null): mass_divergence 100th/100th; DGI 1st/0th (composition-dependent); β below both nulls but by ~0.6% relative (detectable, compositionally negligible; the colour-in-void reading stays the map's, Entry 04). The heuristic's tol=0.25 label is superseded by the percentiles.
• The armature control validates grid detection, not VP detection (Entry 17): all the control's intersections are horizontal × vertical facade crossings, so "the estimator can find convergence" is retracted; what stands is that a stable rectilinear intersection structure exists on the Albertian control (y locked to width 0.003 over 10 runs) and nothing locks on the Bruegel (y range 0.37). One control; the orientation share is an I-8 quantity read from a frozen 40-run measurement (diagonal median 36% vs the control's 0% in every run), not the single decimals ("49%", "39.5%") quoted in earlier passes; the finding is framed as recovering the documented Weltlandschaft mode, not discovering it.
• The sun finding is scoped (Entry 18): its last-place saliency/edge match the picture's other smooth regions (geometry, not act), and the sun differs between versions (Brussels-only claim).
• R_spatial is a candidate (C-3), not a shipped metric. The range, not the point, is the finding, and the subadditive verdict does not classify composition type (Entry 10).
• Accent counts are lens-capped and never abundance measures; the claim is channel segregation and distribution, not a count.
• The compare is placement, not measurement (n=5): at this n the max |z| is √(n−1)=2.000, so magnitudes are meaningless and only the rank is read (Entry 11).
• The witnesses' gaze is not measured (Entry 19): figural facing and gaze are the C-4 pre-semantic ceiling. The composition's attention is measured (saliency, Entry 09); the figures' is a human reading.

Sources & References

Primary object: Pieter Bruegel the Elder (attributed; contested as an early copy after a lost original of c. 1558), Landscape with the Fall of Icarus, oil on canvas, Royal Museums of Fine Arts of Belgium, Brussels; source image the Google Arts & Culture / RMFAB high-resolution capture (12265×8000), public domain via Wikimedia Commons. Delta pass (Entry 16): the Van Buuren Museum version (oil on panel, with Daedalus; Wikimedia reproduction, 1772px, chain unknown) and a museum phone photograph of the Brussels canvas (Wikimedia, frame-cropped here) as the same-object chain control. Comparison sources: the library's pinned study reproductions of Degas Dancers, Pink and Green, Pollock Convergence, Klimt Mäda Primavesi (Met CC0), and Caravaggio The Calling of Saint Matthew.

• W. H. Auden, Musée des Beaux Arts (1938): the attention claim the study's saliency finding is set beside (the poem is about orientation, not rendering; Entry 06's framing was corrected accordingly).
• Rudolf Arnheim, Art and Visual Perception (California, 1954/1974): centric/eccentric composition and perceptual weight, the frame the mass-vs-attention reading borrows.
• Royal Museums of Fine Arts of Belgium: the object, its condition history (support anomaly, overpaint, relinings, the disputed transfer-from-panel reading), and the copy/attribution debate.
• Christina Currie & Dominique Allart, The Brueg(h)el Phenomenon (Brussels, 2012): the technical study treating both versions as copies of a lost original and discounting the panel-transfer reading for the Brussels canvas; dendrochronology dating the Van Buuren panel 1577 at the earliest.
• Positive-control image (Entry 17): Perugino, Christ Handing the Keys to St Peter (Sistine Chapel wall fresco), public-domain Wikimedia reproduction, run at the matched 2600px working scale.
• Instrument lineage and prior studies: the ISR library lab books for Matisse, Degas, Pollock, Caravaggio, Soejima and Klimt; instrument findings I-1…I-15 and methodology in OBSERVATIONS.md; candidate register (R_spatial C-3) in toolkit/CANDIDATES.md; the program's published papers at parallaxmetrology.com.

Computational results in this lab book are reproducible from the scripts named in each entry (environment: python3.14; instrument installed editable from instrument/; Hough-based results are stochastic and reported as multi-run ranges per I-8). Record manifest: report_nopersist.json · icarus_probes.py/.json · jitter_probe.py/.json · null_percentiles.py/.json · null_control.txt (toolkit/patterns/null_control.py) · depth_probe.py/.json · resolution_sweep.py/.json · delta_versions.json + delta_versions_matched.json · sun_baseline.py/.json · armature_probe.py + armature_intersections.json + armature_node_spread.json (the retraction's evidence) · make_armature_figure.py · armature_stats.json (frozen stochastic stats) · make_charts.py / make_charts_2.py · compare_out/. Shared-number integrity (review 7): every figure quoted in both this lab book and the critique is interpolated at assembly time from these JSON artifacts by study_metrics.py (a double-brace token substitution, fail-loud on any unknown token); build.py rebuilds both documents and asserts no token survives unresolved. A correction to a probe JSON therefore propagates to both documents by rebuild, not by memory — the study's trace-to-record rule enforced mechanically for numbers, after hand-sync proved the residual failure mode. Core instrument untouched (one shared-tool fix: the null_control.py verdict string, Entry 09).