The Calling of Saint MatthewCaravaggio, 1599–1600: tenebrism, measured as a relation system

01 Condition check & setup

Clean reproduction, extreme dark key

The source photograph shows paint running to all four edges; no frame, mat, or bare-canvas border; so island detection runs in edge-mode raw per I-1; the edge-aware safe zone would protect nothing here and risks suppressing the standing figures at the right margin. The image is also the darkest key of the four works now in this library: a single diagonal beam of daylight against an otherwise near-black interior.

The native scan is 8792×8304 (73 MP); roughly 20× the pixel count of the Matisse/Degas/Pollock working files. A first full read at native resolution did not complete in three minutes; the instrument's own performance envelope (≈45s without persistence) is calibrated on reproductions in the 1200–3000px range. The working copy for all core-spine and relational reads is a Lanczos downsample to 2600px long edge, matching the Matisse scale. The native file is kept only for the resolution sweep in Entry 11 (I-10: resolution is part of the measurement, and a claim that hasn't survived a sweep is a reading aid, not a result).

02 Core spine

Evidence fields, kernel, tonal/colour numbers

03 The anchor is the wall, not a body

The largest connected mass, and what the method bakes in

The island pass (raw mode) finds 20 islands, entropy 0.7716 (raw) / 0.792 (edge-aware); mid-range between Matisse's single dominant anchor and Degas's near-even dissolution. One island dwarfs the rest: island 18, area 16.07% of the whole image, tagged anchor. The next-largest is less than a third its size (island 10, 5.55%).

Per house rule (verify placement against pixels, not the eye), the anchor's centroid (nx +0.267, ny −0.295) was cropped directly:

Independent corroboration: the 3×3 mass_grid finds the top-right zone carries both the highest mass share (14.18% of 9 zones) and the highest edge density (0.4001) of any zone; the same region, reached by an entirely separate 3×3 binning method. The remaining large islands were also checked against pixels: #10 (5.55%) is the seated youth's leg and breeches on the storage box; #16 (4.85%) is the standing companion's cloak; #2 (3.99%) is the coin-counting hands and tabletop; #1 (2.09%) is wall behind Matthew's group. Of the five largest islands, two (18 and 1, together 18.2% of the frame) are bare wall; more image-mass than any single figure.

04 Two masks, two answers

A methodology note, not a contradiction

Entry 02's kernel block reports mu (cohesion) = 0.2476; a fragmented reading (TEXTURE_FIELD mode, 23 connected components); while Entry 03 reports one island at 16.07% of the whole image. These are not the same measurement. kernel.mu is the largest connected component's share of a binary structural mask (itself only 6.48% of the frame; mass_fraction); the largest component within that mask is 24.76% of the mask, not of the image (≈1.6% of the whole frame). The island system instead segments a continuous density field by watershed, and finds a single 16%-of-image region. Both are correct measurements of different quantities; citing mu and an island's area_frac interchangeably would be a category error, in the spirit of I-2's two denominators. Practice going forward: state which mass pipeline a number comes from.

05 Tonal extremity: measured chiaroscuro

The dark is the field, not the backdrop

For comparison, Degas (Dancers, Pink and Green) reads 60% shadow / 0.01% highlight, and Matisse (La Leçon de piano) reads 90% midtone with a compressed, non-shadow-keyed distribution. Caravaggio is the most shadow-extreme of the three by a wide margin, and, counter-intuitively for the most famous tenebrist of the sample, carries proportionally more highlight mass than Degas (0.44% vs 0.01%): the beam of daylight, though narrow, is a real highlight register that Degas's shadow-keyed dance scene never reaches.

field_regime = QUIET_DISSOLVED, with the instrument's own gloss: "high rv reflects absorbed edges, not empty space." That is the right caution to hold onto reading this painting; large areas measure as void not because nothing is painted there, but because form is modelled into the dark below the edge/contrast floor. vtl.spectral.gamma = 0.01 (Matisse 0.415, Degas 0.108) says there is almost no gap between the confirmed edge spine and the widest gradient activation; construction_label LUMINANCE DOMINANT, same family as Degas, but with almost none of Degas's diffuse penumbra left over.

The reframe worth stating, because the obvious reading is a trap. "90.4% shadow" invites the reply that Caravaggio painted dark, which is not news. What the measurement changes is the unit of analysis. The instrument holds no figure/ground prior; it reads the whole frame as one field, and three of its readings sit oddly with the "lit figures on a dark backdrop" account. The largest, most cohesive single mass in the painting is a passage a viewer calls background (the lit wall, Entry 03; though that reading's direction is partly baked into the segmentation method, so treat it as corroborating, not load-bearing). The regime gloss calls the dark "absorbed edges, not empty space" (above). And the load-bearing reading, the one that depends on no segmentation choice at all: the two fields split diagnostically, with the dark dominating the tonal field (90.4%) while carrying almost none of the structural mass (mass_fraction 6.5%, void 72%). It holds the picture without weighing in it. That combination is what "container" means literally: the dark is not what the figures stand against, it is the field they are set into, and the lit passages are localized activations within it. Two honest limits: the container framing is interpretation (the tonal/structural split under it is measurement), and none of this corrects a viewer's perception, which is a different quantity from connected-component area (Entry 03). The claim is only that the instrument, reading without a figure/ground prior, finds the dark load-bearing rather than empty.

06 Colour rides the tone, not the void

The outlier of the four-work comparison

Both axes order cleanly and in the same direction: Caravaggio has the strongest positive coupling between colour and edge of the four works measured so far, and the least colour living in the structural void. Where Matisse and Degas each build a painting on a colour/structure split (Matisse weaponising it into instability, Degas dissolving everything through it), Caravaggio runs the opposite construction: wherever there is colour, there is already drawing. Colour is not an independent register here; it is bound to the same chiaroscuro modelling that builds the tonal structure.

Zone probes (I-4 mask-to-mass reads, box mean vs the object's own connected mass inside the box):

The wall's own mass-anchored colour reading (0.666) is the highest of the four probes; higher even than the calling hand's red sleeve. This does not contradict the low whole-image beta: beta counts colour specifically where edge activity is absent (below the 28th percentile), and the wall, flat as it looks, carries real edge activity (window mullions, the cast shadow line); so its warm colour does not register as "colour in the void." It registers as exactly what Entry 03 already found: a coherent, actively-structured plane, not a blank one.

One tension worth flagging as open, not resolved: chromatic_boundary_independence (cbi) = 0.236 is the highest of the four works (Degas 0.136, Matisse 0.100, Pollock 0.037); meaning the top-15% strongest chroma boundaries most often have no underlying luminance edge, which sits oddly next to the highest coupling_index of the four. The likely site is costume ornament; the parti-colour red/gold banding on the sleeves visible in the calling-hand and Matthew's-boy crops, where a dye-pattern edge can sit inside one continuous tonal fold. That reading is plausible but not yet pixel-verified against a boundary map the way the anchor and accent findings were; it is carried to the claim ledger (Entry 16) as open.

07 The accent system, corrected

A border artifact caught before it became a claim

The default (structural) gap lens returns 24 accents; real accents exist here, unlike the Degas's fine-brushwork "0 accents" null (I-5). But 11 of the 24 (46%) sit exactly on the frame border (|nx| or |ny| ≥ 0.48), clustered as a suspiciously uniform run along the bottom edge.

This reads as a reproduction-edge artifact (a canvas-mount or scan edge), not a painted mark, and was excluded before drawing any conclusion from it. The remaining 13 interior accents concentrate in exactly two places: the extended calling hand, and the faces turning to react to it.

Once the border band is discounted, the accent system tracks the narrative action with real precision: the calling hand carries the two highest-leverage interior accents, and the next tier is the chain of faces turning toward it, plus one small, specific still-life detail on the table. This is a different kind of accent grammar than either calibration study: not decorative punctuation (Degas's hair-flowers/bodices) and not a texture correction (Pollock's chromatic bridges) but a measured trace of the gesture-and-reaction the painting is built around.

08 Chromatic & combined channels: no split here

Contrast with the Degas I-7 finding

Following house practice (I-5/I-7: check a second accent channel before concluding), the chromatic and combined channels were run at the same edge-mode. Both return the same accent count (24) and, more importantly, their top-leverage accents land in the same two places as the structural channel: the wall/window border zone and the calling-gesture group. Unlike the Degas, where the chromatic channel revealed an entirely separate accent system (warm flecks down each dancer's body, invisible to the structural channel), Caravaggio's punctuation does not split by channel; consistent with Entry 06's finding that colour here is not an independent register from structure. Where colour and edge move together throughout the image, there is no second, colour-only accent system left to find.

09 Saliency vs mass

The eye and the weight, again in tension

Saliency centroid (spectral residual, no model, no training) sits at (0.029, 0.042); close to frame-center; the structural mass centroid sits at (0.087, −0.024); pulled toward the wall, upper right. The resulting divergence, mass_divergence = 0.0877, is larger than the Matisse calibration figure (0.072, "the eye goes where the structure isn't") and far larger than Degas's 0.02 (weight and gaze coincide). The saliency algorithm favours small, high-frequency, singular regions over large flat planes almost by construction; consistent with a read where the eye is pulled toward the small lit faces and hands even though the largest cohesive mass sits on the wall behind them.

10 Stance: a discrepancy worth logging, not resolving

Torque on this working file vs the single earlier datapoint

isr stance on this study's working file reads torque = 0.1269 → standing. OBSERVATIONS.md's C-2 corpus (2026-07-02, n=10, 768px working resolution) recorded this same painting at torque = 0.24 → mixed, from a different source reproduction. Per the established "crops change numbers" principle; here extended to stance, which was not previously tested against this caveat; both readings are logged rather than treated as a contradiction: they are measurements of two different photographs of the same painting, and the gap (0.13 vs 0.24, standing vs mixed) is large enough that it should not be silently averaged away. Future stance work on this painting should record which source file it used.

11 Self-similarity: a resolution sweep, per I-10

D is stable; its own stability metric is not

Dark-mask D is saturated and uninformative here regardless of resolution; at 86.5% dark fill, the instrument's own reading string states the Jones-Smith caution directly: "dense fill: dark-mask D saturates toward 2; read the structure mask." Structure-mask D is genuinely resolution-robust for this painting (1.558–1.571 across a 2.8× sweep, unlike Pollock's 1.08→1.49 drift); but its slope_stability is not: it falls monotonically from 0.963 at 1300px to 0.703 at 3600px. This is a distinct nuance from the Pollock case (where D drifted and stability held fixed at each single measurement): here the fractal dimension itself travels well, but confidence in the fit (how well one slope holds across scale windows) degrades as more of the fine brushwork enters the mask at higher resolution. Any claim built on this stability number must carry its resolution.

12 Zone probes, revisited

What a tight box vs. a loose one tells you

The money-table probe (Entry 06's table) is the tightest of the four; only 4% mass coverage, the highest dilution (0.207); because the hands-and-coins gesture is a small, precise cluster inside a box that is mostly dark tabletop and background. Matthew's-boy probe is the opposite case: 36% coverage, the loosest dilution (0.149 is still non-trivial, but lowest of the four); his lit head and torso fill most of the box outright. Read together with Entry 03, the pattern is consistent: the painting alternates between a few large, coherent lit masses (the wall, the boy) and several small, precise, high-contrast incidents (the hand, the coins); exactly what a single directional light source picking out selected objects would produce.

13 Cross-scale consistency (LG / TLG)

Gradient-consistent, tonally at the ceiling

lg_score = 0.0196 ("globally consistent") says the gradient field's logic holds across quadrants; unsurprising, since the same chiaroscuro-modelling technique is applied uniformly to every figure regardless of position. tlg_score = 1.0, the ceiling value, says the opposite is true of the tonal field: the four quadrants occupy four essentially independent value registers.

METRICS.md notes that an image can be gradient-consistent and tonally divergent simultaneously, and names the Matisse (tlg 1.0) as the example. Caravaggio ties that ceiling here, but for a different structural reason: not a decoupled colour/structure grammar (Entry 06 found the opposite for Caravaggio) but four genuinely different lighting registers occupying the four corners of one directionally-lit room.

14 Radial compliance & default gravity

An authored, off-center composition

The reading classifier agrees: structural_state = "counterweighted", flags counterweight_present + offcenter_mass, with the note that a region "reads as a load-bearing balance term more than a construction outlier"; here, that region is the money-counting group. Remove it, structurally, and the composition's balance measurably worsens; it is not incidental staffage but a counterweight to the mass pulled right by the calling group.

15 Same coordinate space: the four studies together

isr compare: Caravaggio, Matisse, Degas, Pollock

Sorted by coefficient of variation, the top of the dispersion table is exactly where the four works are expected to differ: radial compliance direction (dRC), highlight_mass (Pollock's all-over field vs the three figural works), and (the clean headline) coupling_index, where Caravaggio (+0.363) anchors one end and Pollock (−0.088) the other, with Matisse (+0.169) and Degas (+0.009) in between: a genuine four-point spread from bound to fully decoupled colour/structure, matching the read in Entry 06 with an explicit rank across the whole small corpus rather than a pairwise comparison to one calibration study.

16 Claim ledger

What survived, what narrowed, what stays open

17 Open threads

A dedicated chromatic-boundary-independence location map (to test the costume-trim hypothesis directly, the way the accent-lens crops tested the gesture hypothesis). A named structural probe on the two men at the head of the table; distinguishing which figure carries which role is an art-historical question out of this instrument's reach, but their two lit faces could still be probed and compared structurally without resolving it. A full run with persistence enabled (this session used --no-persistence throughout for iteration speed) to get a fragility reading akin to the other three studies'.

Status at close of study: the cbi location map was run and closed the hypothesis (Entry 29, refuted); persistence was run on both pipelines (Entry 30); the two-men probe is absorbed by Entry 25's face-photometry battery (old man 0.515 vs bearded man 0.456, both frontally lit, 1.13× apart; structurally near-equals; which is which remains a semantic question). An isr position completeness run places the painting in the absolute coordinate frame: displacement 1.944 from structural origin, loudest axes orientation entropy (0.997), shadow mass (0.904), island evenness (0.80), void ratio (0.72) (caravaggio_position.json). Still open after everything: the stance discrepancy against the C-2 datapoint (Entry 10), which needs a second source reproduction; the seed of a reproduction-robustness study, a project of its own.

18 Addendum: the noise check (appended after first assembly)

PM verdict "meaningful": one claim narrows, two strengthen

The dark passages of this reproduction carry a visible speck field; dust, varnish scratches, and grain, plainly visible in any zoomed crop of the lower shadows. The first assembly of this lab book did not test whether that noise mattered. The instrument has a built-in answer: re-reading with Perona-Malik preprocessing (preprocess_pm=True) reports a noise-energy verdict. The verdict here is pm_noise_energy 0.0164 → "meaningful"; above the 0.010 threshold. The noise is real and measurable, so every headline claim was re-checked against the denoised read (caravaggio_report_pm.json):

Correction to this entry's own first draft (same session): the draft asserted that the 3×3 mass-grid corroboration was noise-independent, without checking. Checked in Entry 21: it is not fully; under PM the top-right cell keeps edge-density rank 1 but slips from mass rank 1 to rank 2 (13.90%, behind the center cell's 16.44%). The corroboration weakens from "highest on both axes" to "highest edge density, second-highest mass." The pixel crop at the anchor centroid remains noise-independent evidence of what the anchor is. What changes is the strength of the magnitude language: "2.8× the next-largest region" is a fact about this reproduction's noise as much as about the painting; "the largest, and still the wall, under denoising" is the version that survives. The claim ledger (Entry 16) is superseded on that row by this entry. Practice reinforced: for museum photography of dark paintings, run the PM check before first assembly, not after.

19 Accents, re-run denoised

All 13 survive; the noise was hiding six more

The gap lens re-run on the PM-denoised image (structural channel, edge-mode raw): all 13 interior accents from Entry 07 reappear; every one position-matched within 0.03 frame units, most within 0.01. The border-artifact count drops from 11 to 5 (the artifact band partially smooths away), and 6 new interior accents surface that the speck field had been masking. Each was verified against pixels before acceptance:

Net effect on Entry 07's claim: strengthened. The accent system still concentrates on the calling gesture and its reaction chain, now including the gesture's own fingertip, plus the window/beam edges that carry the light into the room. Denoising did not relocate the punctuation; it sharpened it.

20 Self-similarity sweep, re-run denoised

Entry 11's "stability degrades with resolution" was mostly the noise talking

Two revisions. (1) Entry 11 read the monotonic stability collapse (0.96 → 0.70) as a fact about the painting's fine brushwork entering the mask; denoised, the collapse largely disappears (0.87 → 0.82, nearly flat). The degradation was speck noise entering the structure mask at higher working resolutions, not paint. Entry 11's "new nuance" (D travels, confidence degrades) is retracted as mostly a noise artifact. (2) The denoised D drifts slightly more across the sweep than the noisy D did (1.560–1.607 vs 1.558–1.571); the noise had been flattening the apparent drift by adding a resolution-independent speck floor to the mask. The honest summary number for this painting is structure D ≈ 1.56–1.61 across a 2.8× sweep and both noise pipelines; cite the range, not a third decimal. Dark-mask D remains saturated and uninformative in both pipelines (86.5% dark fill, both → ~1.94).

21 Balance, grid, probes, stance: denoised

One attribution retracted; the counterweighted reading survives

Entry 14 and the reading classifier's note ("the coin-counting group is the load-bearing balance term") are superseded by the two-counterweight version above. The composition remains measurably counterweighted and right-skewed in both pipelines; what the noise was distorting is which single region gets the credit; a distinction the instrument can only make reliably on a clean signal.

22 Chromatic channel, re-run denoised

The no-split claim survives: and the border band turns out to be achromatic

Entry 08's claim that Caravaggio has no second, colour-only accent system was an absence claim measured on the noisy image; exactly the kind of null I-5 warns can be fake. Re-run denoised: the chromatic channel returns 24 accents whose top positions co-locate with the structural zones already pixel-verified; the stylus cluster on the money table (its top accent, leverage 1.13, sits at the same coordinates as the structural closure-delay pair), the hand cluster, the window/beam edges, and the right-hand group. No Degas-style second system appears once the noise is gone; the claim stands on a clean signal now, not a noisy one.

One incidental confirmation: the denoised chromatic accent list contains zero border-band accents (the structural channel still finds 5). The bottom-edge band is an achromatic, tonal-only event; consistent with a physical mount/scan edge rather than anything painted, closing the loop on Entry 07's exclusion.

23 The critical literature: what is actually contested

Mapping the debates before measuring them

A survey of the scholarship (sources below) yields three live, measurable disputes and several settled claims worth quantifying on the official reproduction. All probes in Entries 24–27 are run by caravaggio_deep_dive.py (this folder), PM-denoised pipeline primary with noisy sensitivity throughout, deterministic (no Hough. I-8; line fits are least-squares on detected edge points, gesture axes are intensity-weighted PCA).

Sources: Stork & Nagy, "Inferring Caravaggio's studio lighting and praxis in The calling of St. Matthew by computer graphics modeling," SPIE 7531 (2010), read in full (diatrope.com/stork/StorkNagy.pdf); Wikipedia, "The Calling of Saint Matthew" (Varriano 2006 p. 111; Lavin 1993 p. 95); Italian Insider, "Who is Saint Matthew for Caravaggio?" (Magister 2018 vs Sullivan; Lev 2012); caravaggio.org commentary (oilskin window, three-source reading). The Magister and Sullivan arguments are reported at second hand from the Italian Insider account; her book was not consulted directly.

24 The wall falloff and the window, measured

Stork & Nagy's key premise gets its numbers

The falloff premise is confirmed, and now has values: along the clean wall segment (x 0.79–0.99, ≈68 cm of painted wall) mean luminance rises monotonically toward the light, right/left ratio 2.32 (linear slope +0.76 luminance per frame-width; identical in both pipelines). Across the full width the story is starker: the wall left of the window reads 0.076 against 0.329 right of it; a 4.3× drop across one wall. Stork & Nagy matched this gradient by eye in a renderer; the profile itself, from the museum reproduction, is what these curves record.

The window: its brightest pane reads 0.278, its dimmest 0.133; both below the mean of the wall they are set in (0.329) and far below the lit wall above the shadow boundary (0.439). The window is measurably an object the light falls past, not a source: a luminous opening would have to out-read its own wall. The oilskin-window reading survives quantification.

25 The faces: Stork's praxis premise, refined

"Nearly equally bright" is true of exactly three faces

Two refinements to the published claim. First, "the faces are nearly equally bright" holds for the three frontally-lit faces only (0.456–0.597, a 1.31× spread); across all five table figures the spread is 2.56×, driven by pose, not position. Second, and this is the part that carries Stork & Nagy's praxis inference: brightness has no correlation with x-position (r = −0.08 across the table group; the brightest face is in the middle, the leftmost face is second-brightest). Under a local source at the right, 1/r² predicts a left-dim gradient; there is none. So either the light was effectively distant (solar), or the faces were individually corrected; exactly the fork they propose, now resting on a measured flat profile rather than an impression. Incidentally measured: Christ and Peter (0.284 / 0.271) carry about half the luminance of the tax-collector group's lit faces; the caller stands in his own shadow zone below the boundary (Entry 26), while the light he "brings" lands on the called.

26 Two shadow edges, two occluder distances

The penumbra numbers Stork & Nagy's hypothesis needed

The wall carries two diagonal shadow features, and they are physically different kinds of edge. The lower shadow boundary, the crisp diagonal separating the lit upper wall from the darkness the figures stand in, runs at 15.1° above horizontal (PM) / 17.5° (noisy; the systematic difference between pipelines exceeds the ±0.2° statistical error, so quote 15–17.5°), with penumbra widths of 0.9–1.8 cm that broaden mildly leftward, away from the frame edge, as an off-right occluder predicts. The upper darkening (Stork & Nagy's "unnoted upward-slanting cast shadow") resolves on the official reproduction into a soft triangular patch centered at (0.88, 0.07), long axis 25.8° (roughly parallel to the lower boundary), whose edge is ≈6 cm soft: four to fifteen times more diffuse than the crisp boundary below it.

The arithmetic these widths permit (reading aid, stated fully): a shadow edge's penumbra width ≈ occluder-to-wall distance × the source's angular size. For the sun (≥0.5° effective, more near a horizon), a 0.9–1.8 cm penumbra puts the occluder roughly 1–2 m from the wall (window-frame territory, not a distant roof); the ≈6 cm patch edge puts its occluder at roughly 7 m (architecturally right for a roofline). Reproduction blur and varnish can only widen a painted edge, so the measured widths are upper bounds and the distances upper bounds with them. Read together: the two edges are mutually consistent with Stork & Nagy's composite geometry; strong light entering past a near occluder (window edge), plus a far, soft occluder (their cornice) darkening the top of the wall. What the measurement adds is that the two features genuinely carry two different occluder distances on one wall: the quantitative content their renderer matched by eye. Caveat kept in view: these are 1600-era painted edges read through a 2026 photograph; the arithmetic assumes Caravaggio rendered penumbrae faithfully; which is exactly the premise of their paper, not a finding of ours.

Also for the record: left of Christ, the lit/dark boundary is a different, near-horizontal edge; the window-sill shadow at y ≈ 0.35; not the continuation of the diagonal. A first-pass fit that mixed the two (plus two stations that caught the top of Peter's head) read 22° with scattered residuals; the pixel check caught it. The diagonal proper is only measurable on the clean segment right of Christ's hat.

27 The Matthew debate, measured three ways

The light illuminates one candidate and points between both

Read together, the three measurements give the debate a sharper shape than either camp's prose: the light's illumination unambiguously favors the bearded man (1.9×, robust), while the light's geometry and the bearded man's gesture are both precisely poised between the two candidates; the gesture axis threading his own chest and the slumped youth's head within centimeters of each other. On these measurements, "deliberately ambiguous" is not a critic's shrug; it is a described property of the picture's geometry, coexisting with a photometric vote for the tradition. What the instrument cannot do (and nothing in this entry does) is say which figure is Matthew: that is a question about narrative, series context, and gospel text (Sullivan's strongest arguments live there, as do Magister's), not about pixels.

28 The "empty" upper half, closed out

Plus the deep-dive ledger

The top third of the canvas reads brighter than the rest (mean 0.156 vs 0.144) and carries 2.5× the share of bright pixels (21.7% vs 8.6% above 0.25 luminance). With Entry 03's anchor finding (the largest coherent mass in the painting is the lit wall in this band), the didactic commonplace inverts cleanly: the upper half is not empty space above the action; it is where the painting's light, its largest structural mass, and its structural anchor all live. The action happens under a loaded band, not under a void.

29 The cbi location map: the costume-trim hypothesis is wrong

The open ledger row closes as refuted, and yields a metric caveat

Entry 06 left one row open: the painting's high chromatic-boundary-independence (0.236, highest of the four studies) was hypothesized to be the costume trim; dye-pattern edges riding inside continuous tonal folds. Mapped (caravaggio_closeouts.py, replicating the core metric's mask exactly at the call site), the hypothesis fails: the costume zone holds 14.8% of cbi pixels at a density of only 1.1× its area share (PM), while 48.9% of all cbi pixels sit in the dark left-and-floor zone (density 1.22×), with the lit window/wall zone at 0.62×. The picture, not the numbers alone: the magenta lace above threads the shadow field. In a 90%-shadow painting, "no luminance edge" is trivially true almost everywhere dark, so any faint warm/cool modulation in the darks counts as an independent chroma boundary. Two consequences. (1) The costume-trim row closes as refuted. (2) The cross-study claim narrows: denoised, the cbi value itself drops from 0.236 to 0.190; below the metric's own ≥0.20 autonomy line; and what remains is dark-field micro-modulation, not colour drawing lines light doesn't draw. Metric caveat earned: cbi is inflated on dark-keyed paintings and should be read against a location map there, never as a scalar.

The general warning is worth stating, because cbi is unlikely to be the only case. Any metric built on a threshold relative to this painting's own distribution risks measuring the dark key rather than the composition: in a field that is 90% shadow, "below the luminance-edge floor" or "in the darkest zones" describes most of the canvas, so a metric gated that way can report the tenebrism back to itself and look like a finding. cbi is the one this study caught because its location map made the artifact visible; the honest position is that every distribution-relative metric here would need the same test ("what would this read on any very dark painting?") before its number is trusted as a fact about this composition rather than about its key. That test is a cross-corpus study, not run here; until it is, the tonal, mass, and coupling readings that survive denoising should still be read with that possibility live.

30 Persistence: the standard spine completes

The composition survives degradation; what it is fragile to depends on the pipeline

The full read with persistence (the one step Entry 01's iteration skipped) ran on both pipelines: 0.844 (noisy, fragile_to grayscale) / 0.809 (PM, fragile_to blur). Both are high; the island structure survives blur, grayscale, and low-contrast degradation better than the Pollock (0.718); a composition built from a few large lit masses in darkness has little to lose to any single channel. The fragile_to flip is consistent with I-11's grain findings: on the noisy copy the weakest link is colour removal (grayscale collapses the warm-cool separation that helps segment the darks), while once the grain is smoothed away, blur (which erodes the few crisp boundaries that remain) becomes the binding constraint.

31 The spark gap

The interval Christ's hand reaches across, measured: and Michelangelo's, for scale

Three structural facts. The hand halves the crossing: group-to-group directly is ≈109 cm; through the hand it is 31 + 33; the hand is a stepping stone placed mid-channel, slightly nearer its own group (reaching, not arrived). The seated youth's nearest lit neighbor is the hand (33.4 cm); not any member of his own group (his closest is the feathered boy at 36.0): the figure most often read as poised to rise is, structurally, already half-claimed by the gesture. Accent 41 sits on the corridor: 20.8 cm from the hand, 15.9 cm from the table group; the sum (36.7) is the span itself; the gap-lens bridge accent lies on the crossing path, 57% of the way from fingertip to group. The interval is not empty; it carries the painting's punctuation.

Michelangelo, for scale. The quotation claim (Entry 23) invites one number: the Sistine Creation of Adam's fingertip gap, measured by the same landmark protocol on the Wikimedia Commons reproduction (3524px, fingertips pixel-placed at 10× zoom), is ≈16 px = 0.44% of the composition's width (≈2.5 cm at the ≈570 cm fresco; the file is cropped, so the cm figure is approximate; the width-fraction is the reproduction-safe unit). Caravaggio's spark gap is 11.0% of his canvas width: he quotes the hand and opens the gap by a factor of ≈25. Michelangelo paints the instant before contact; Caravaggio moves the gesture to the far side of the room and makes the distance itself; a dark channel one-tenth the width of the picture, punctuated by one bridge accent; the thing the picture is about. This comparison is geometric only: two reproductions, two media, one shared unit (fraction of composition width); it proves scale, not intent.

32 Where the Void actually hangs

"The exact center of the canvas" is occupied; by the brightest face in the painting

The didactic claim; a dark pocket "suspended in the exact center of the canvas between Christ's hand and the table"; fails in both halves on the official reproduction, and both failures are pipeline-stable to the millimeter. The deepest dark pocket in the inter-group channel sits at (0.73, 0.62), 86 cm from the frame center (18% of the diagonal); the painting's genuinely deepest void is the upper-left corner (146 cm radius), nowhere near the gesture. And the exact center of the canvas is not void at all: the distance-transform reads 1.2 cm to lit structure there; the frame center lands on the feathered boy, the brightest face in the painting (0.597, Entry 25). The honest replacement for the claim: the void between the groups is real (Entry 31 measured the crossing it forces), but it hangs right of center; what Caravaggio put at the exact center is the one face turned toward Christ (a human observation, not a measurement), at the painting's maximum luminance. The center is not empty; it looks.

33 Is the light "physically impossible"?

Four witnesses, one off-stage source: and one witness that failed honestly

Every usable witness; two independent cast-shadow edges, two independent frontal-face shading vectors; places one source up-right, outside the frame, above the window's sill line. The two crisp shadow edges agree to within a few degrees of each other; the face pair agrees internally to ≈1°. So the debate's framing needs splitting: the window-disconnect is real and was already quantified (the panes are dimmer than their own wall. Entry 24), but "physically impossible" is unsupported internally: the depicted room's light is geometrically coherent, one consistent source that simply is not the visible window. What the measurements support is a narrower, better claim: not impossible light; a coherently constructed light whose origin is deliberately kept off-stage. Whether that origin is divine is a question about meaning, not geometry, and nothing here touches it.

34 The Peter counterfactual

The "formal anchor" account, in the one sense the instrument can test

The reported X-ray finding (Peter absent from the first state, added in a second campaign; reported across the survey literature; primary radiograph plate in Friedlaender's Caravaggio Studies per the published record; not independently verified here) makes one of the two competing explanations testable. If Peter was added because "Christ's figure was too isolated on the right edge and needed formal anchoring," his presence should improve the composition's measured balance. Simulation: Peter's figure masked to the surrounding darkness (feathered blend, no artificial seams), full re-read, raw mode, PM pipeline:

Read strictly, the numbers run opposite to the balance version of the formal-anchor account: in centroid terms Peter does not balance the canvas, he loads it further right. Removing him improves every balance measure the instrument has, by two independent methods (the counterfactual re-read and the native per-region dependency on the unmodified painting). But the honest headline is the scope, not the result. Centroid balance is almost certainly not what an art historian means by "anchor." The usual senses (repoussoir, depth staging, the closing of the composition's open right edge, the diagonal thrust of the two entering figures) are spatial and perceptual claims this instrument does not measure, and nothing here touches them. So the finding is narrow and should be read narrowly: if anyone argued Peter was added to even out the picture's left-right weight, the mass numbers contradict that specific claim; the readings people actually hold are untested. Two further boundaries. (1) The simulation is a lower bound: the real first state had Christ's whole figure where Peter stands, so the true before/after difference is smaller than measured here, though the sign is what carries the point. (2) Eliminating one motivation does not establish another; the theological-mediation reading gains nothing here except the removal of a formal rival it may never have had.