Paul Gauguin — Vision of the Sermon(Jacob Wrestling with the Angel) · oil on canvas, 1888 · National Galleries of Scotland, NG 1643

Entry 01 — Selection, source & condition check

Look at the paint before any number

The brief: a work that stresses the Round-17 additions — the texture-robust Notan mass, the bottom-up scanpath, the tonal-vs-chromatic construction signature, and the opt-in epistemic layer (assumption / indispensability / singularity), which was lifted from a multi-mass instrument and wants several distinct masses split by a void. Vision of the Sermon is almost purpose-built for it: two territories — a tonally-built crowd of Breton women, and Jacob wrestling the angel on a vermilion field — divided by the diagonal apple-tree. Gauguin (d. 1903) is comfortably public-domain; the source was reconstructed from the National Galleries of Scotland OpenSeadragon deep-zoom pyramid to a single 6617×5285 file (aspect 1.252 vs the 91/72.2 = 1.260 canvas, so undistorted).

Condition check, per AGENTS Step 0/1 — looking before measuring, with attention to what will inject false structure (the study's opening instruction from RP):

Entry 02 — Pre-registration

Predictions before the read, so numbers test a hypothesis

Written before running the spine (the full table is in labbook.md §3). The bets that matter for the new areas: construction would read chromatic (fragile_to=grayscale) — a cloisonnist colour field that collapses without hue; scanpath eccentric and restless; assumption a real depth spread (red field shallow, crowd deep); singularity crowning a small mass (the wrestling pair). Two were flagged as genuinely contested at the time — the construction verdict (the tonal crowd could dominate the Notan mass and flip it), and whether bottom-up saliency would confirm the critics' "the red catches the eye first."

Entry 03 — Run 1: the spine & the first-run trust gate

The mass is centred, distributed, and counterweighted

Raw read, no persistence, 2000px working copy. State counterweighted; kernel delta (−0.067,+0.038) near-centre; mask_mode REGION_FIELD (not the TEXTURE_FIELD predicted — the weave did not fragment the edge mask at 2000px); island entropy 0.960 (distributed, Degas-like) with high bridge_cost 0.613 — masses isolated by the red ground. The single densest island is the right-foreground woman's coif (+0.39, 0.00), not the left crowd — the first correction to the eye's "lower-left is the mass" bias.

Entry 04 — The accent surface: a flipped prediction

Craquelure SUPPRESSES accents; a second system is chromatic

Step 0 predicted the craquelure would fake accents. The opposite is true. The noisy surface raises the leverage floor, so the raw count under-reads at 6; denoise (Perona–Malik, pm_verdict=meaningful) and ~41 marks emerge — the dark cloisonnist contours plus the light coif highlights. The two raw dark-accents survive PM (so they are real), and the rendered PM field shows the craquelure gone from the darks with the accents sitting on genuine contours — PM reveals, it does not manufacture. Meanwhile the mass/imbalance system is robust to PM (delta moves ~0.05, state unchanged), unlike the Caravaggio.

Entry 05 — Audit: resolution, the nose, the cow, the red

Three of RP's observations, chased to numbers

Resolution (I-10). Re-read from the full source at 1400/2000/2800px. Tone, coupling, β, Notan and the construction verdict are resolution-stable; delta and mu are not — delta_y drifts −0.006 → +0.038 → +0.113 and mu 0.36 → 0.30 → 0.54 as fine crowd detail enters the edge mask. Kernel geometry is therefore reported as a resolution range, not a point.

The cow (RP). Genuinely absent from islands and accents — and a tuned chromatic pass at three coalesce settings recovers nothing, so it is a true structural null, not a threshold miss. The zone-read says why: the cow is the most colour-built, least edge/tone object in the painting, more chroma-pure than the red field itself.

A footnote on where error enters. An earlier draft of this study called this animal a goat; scholarship is consistent that it is a small bucking cow. The correction touches not one number — the pre-semantic instrument only ever found "a chroma-only form on the red." The one unforced error sat in the human naming layer, the layer the study is otherwise most disciplined about: a live demonstration of C-1 (the tool points; the naming is external and can be wrong).

Entry 06 — The epistemic layer (the headline)

Weight and uniqueness live in different places

Final read with include_consensus=True. Scope, stated up front: the layer segments only 4 masses, ~29% of the canvas — the dense dark praying crowd is almost entirely unlabeled (edge/tone-quiet, like the cow and red). So this reads the segmented masses, not the whole composition. Within them, the two halves of Dependency disagree, and the disagreement is the finding:

Entry 07 — Null calibration & construction

β is real but spatially inert; the build is not chromatic

Null sweep (900px, n=24 seeds, phase-scramble + patch-shuffle). β is high (0.371) and real, but identical under patch-shuffle (0.371±0.000, z=−0.7): it measures the statistic "colour occupies the drawing-void," not a spatial composition — the "colour where drawing is absent" reading rests on the location map (we can see it is the red field), not on β. DGI grips and sits below both nulls (z=−3 to −8): the composition sits below the centred-gravity null — that is the coordinate. "Authored against gravity" is the frame one brings to it; default_gravity presumes purpose and cannot falsify it from inside (the preamble's caveat), so the finding is the position, not the intention. cbi and mass_divergence both grip.

Construction — the contested bet, lost and instructive. fragile_to = gamma_down, built = tonal, grayscale-survival 0.912 — identical across 1400/2000/2800px. But adversarially: the mass survives every degradation at 0.87–0.93 (spread 0.061), so it is robustly over-determined, and gamma wins by only 0.046. The claim that survives that scrutiny is the narrow one — the structure is not chromatic-built (grayscale barely dents it), not the flat "tonal" label. And that narrow claim is pipeline-stable, not a reading aid: the 0.046 margin is identical across 1400/2000/2800px and under achromatic and chromatic noise at σ 0.008 and 0.02 (Entry 09) — so it is not inside seed noise. It still partially disputes the art-historical "colour is the engine" emphasis: the load-bearing structure is tone and contour (Notan field_void 0.026 reads the red as mass, not void), while the vermilion is a chromatic plane, and a second colour-accent system, riding on top. (Only the finer "tonal vs over-determined" lean is soft; the "not chromatic" sign is firm.)

Entry 08 — Adversarial audit

Made to attack its own findings, and told not to Goldilock

The story came out clean — every finding reinforcing the last — which is exactly when to distrust it. A dedicated pass re-ran the segmentation, checked placements against pixels, tested for degeneracy, and triangulated the art-historical claims to their actual sources. Four claims survived; three were overstated and are narrowed above (with pointers left where they sit, house rule 3).

Entry 09 — Critique-prep hardening

RP's pointed critique, chased to measurements

A second external critique (RP) aimed the critique pass at the study's soft joints. Each was resolved with a measurement rather than a concession; the deliverable above already carries the fixes.

Entry 10 — Corpus placement & the C-10 test

Where Gauguin sits, and whether the headline generalises

Same-coordinate compare (isr compare vs Matisse / Degas / Caravaggio, all 2000px; z-scores are "who sits where," never significance at n=4, and the four are different reproduction chains). Two standing findings get independent, cross-corpus support: tau z −1.72 — Gauguin has the flattest tonal hierarchy of the four, the cloisonnist flatness confirmed apart from the depth-ladder; and offcenter_force z −1.30 — the most near-centred mass, holding the Entry-05 correction against corpus. Gauguin is also the outlier on highlight_mass (+1.73, the coifs), x_p (+1.67, figures at the frame), and omega (+1.46, the most family-centroid spread).

C-10 test — does the indispensability↔singularity inversion generalise? Ran the epistemic layer on three more works. The result reframes the candidate rather than confirming it:

Entry 11 — The Bernard test: appropriation, measured

What Gauguin did to a canvas he owned

The marquee productive-disagreement test (RP). This is not "influence": in Aug 1888 the 20-year-old Émile Bernard painted Breton Women in the Meadow at Pont-Aven with a fully-formed Cloisonnism (Japanese woodblock + medieval stained glass), and exchanged the canvas to Gauguin (40), who had been painting impressionistically that summer. Gauguin absorbed the flat-colour method overnight, immediately made Vision, then marketed himself as the Synthetist "master" and cast Bernard as a disciple — a permanent rupture. So the dispute is near-appropriation of a method Gauguin physically possessed. Source: the Musée d'Orsay oil (RF MO P 2019-1), matched to Gauguin at 850px — the d'Orsay IIIF ceiling, confirmed.

The instrument refuses the lazy verdict in both directions — Gauguin's is neither a flatter copy nor uniformly more radical. And an adversarial pass killed the one claim that would have been too convenient:

Entry 12 — Triangulation & close-out

Where the instrument meets, and disputes, the critics

External readings were consulted after the independent eye-first read, for grounding not steering (full notes in labbook.md §7). The consensus is stable: the diagonal tree organises the canvas into reality vs vision; the vermilion is deliberately non-naturalistic — a symbolic, dreamlike field, not a depicted lit space; flat cloisonnist colour under strong dark contours; an abrupt scale shift; Hokusai's sumo wrestlers behind the Jacob/angel pair. Where the instrument meets this it agrees and refines: its luminance-salience reads the red as a flat, non-attracting plane — top-3% saliency hotspots are 87% on the left coif contours, near-zero on the red — which is exactly what a symbolic, non-naturalistic field (rather than a lit surface) should be in luminance terms. Stated as a caveat, not a coup (Entry 08): the ISR saliency runs on luminance and is blind to chromatic pop-out, so it says nothing about the red's colour pull; the red's force is chromatic and symbolic, and the luminance channel correctly does not register it. Gauguin's stated deliberate abstraction ("Art is an abstraction… dream before nature") is the clean anchor: the presumption of purpose is unusually safe here, but only for the flatness being authored, not for its quality.

Study lab book in a solid place, 2026-07-27. Open to the critique pass (skills/isr-critique) on RP's confirmation. The one-line thesis to lead a critique with is the chronology one: the composition delays the reveal of its own subject; the tonal-skeleton / chromatic-plane split is the mechanism under it.

Pre-registration scorecard

What was bet before the read, and how it landed

Ten bets: four clean hits, one hit-on-components, five predicted-against misses — and the misses carried the study (the construction flip, the near-centre correction, the indispensability inversion, the accent reversal). A reading fitted to numbers has no misses; a reading that tests hypotheses does.

What this does not prove

• Not "the subject is not the structural anchor" as a fact about the painting. The epistemic layer saw ~29% of the canvas (4 masses); the claim is a dissociation within those masses, and "which mass is the subject" is the C-1 label map, not ablation. The crowd is read by other tools, not the epistemic layer.
• Not that colour is inert. β, void_preference and the colour-family centroid are all spatially inert or tautological here; the colour channel shows distributed boundary-drawing, not that colour does nothing. The claim is only that neither channel shows colour building hierarchy.
• Not human viewing order. The chronology (step 17/20) is the instrument's saliency-ordered traversal, calibrated against one control (Klimt step 4). It is bottom-up and luminance-driven; it does not claim where a person looks, or when.
• Not who invented cloisonnism. The Bernard comparison shows Gauguin's canvas is a transformation (colour-autonomy, tonal drama), not a copy — it does not adjudicate priority, and the 850px cross-chain comparison excludes cbi and fine-structure (I-25/I-10).
• Not quality, and not choice. Every reading is analysis of structure, not a verdict on merit; the instrument presumes purpose and cannot falsify it from inside (the preamble). "Authored against gravity" is the coordinate (below the null), the intention is the frame brought to it.
• Not resolution-independent everywhere. Kernel geometry (delta, mu) drifts with resolution and is quoted as a range; the tonal / colour / Notan / construction findings are resolution-stable.

Future / owed items (bookmarks — do not lose these)

• The subject-weight corpus test (replaces the fragile C-10): locate the subject (C-1), measure structural weight there (island / Notan / saliency), ask if subject-location anti-correlates with weight, across every prior study with an identifiable subject. Robust form of the thesis; no epistemic layer, so the segmentation blocker (I-29) does not bite. The single highest-value next build.
• Structural chronology as the fifth Round-17 lens. Subject-reach-step, calibrated against subject-is-anchor controls (Gauguin 17 vs Klimt 4 is the first pair). Generalise + give it a proper control distribution.
• C-10 (indispensability↔singularity inversion): gated on the consensus layer resolving ≥3 differentiated masses (I-29); not promotable until the segmentation granularity improves — a spine change.
• I-31 as a Step-0 rule: name the work's dominant construction mode before running, expect the default lens to be blind there, reach for the matching bolt-on first.
• Higher-res Bernard: the d'Orsay IIIF ceiling is 850px (confirmed); a better plate would let cbi and fine-structure into the appropriation comparison.
• Shelf tools not yet run that bear on standing claims: isr stance, tonal_gestural_offset, local_contrast_ratio render panels.

Status

Lab book open, in a solid place — 2026-07-27. Twelve entries; three adversarial passes and two external critiques absorbed; ten pre-registered bets scored; four fail-boxes and every superseded claim logged where it fell (house rule 3). Instrument findings harvested: I-29 (epistemic-layer degeneracy), I-30 (capped-scalar guard), I-31 (default-lens blind spot). Candidate register updated: hue-tension (n=3), C-10 (n=1, not promoted). Owed to close-out: OBSERVATIONS.md Study-14 gather, the critique essay on confirmation, and a commit (this all lives in the working tree). No number retracted without its replacement measured.