Behind the Gare Saint-LazareHenri Cartier-Bresson, 1932 — the corpus's first photograph

Entry 01

Selection & condition: the handoff said mount, the pixels say none [stable]

The condition step opened with a correction. The handoff warned "mount border present — segment to the image-proper first." Measured, there is no mount in this file: content reaches all four pixel edges (left columns mean ~77, the dark hoarding; right ~174, grainy fence reflections; bottom ~88, dark silt; top 217 with std 3, flat near-saturated sky, not cream sheet). The measurement frame is the full native file, user-verified before any run. One hairline ~2px seam at the extreme right column (sky region): scan edge, covered by edge-aware mode, not worth a crop (crops change numbers).

Two condition facts worth carrying. First, the aspect ratio is 0.723 against the 35mm negative's 0.667: consistent with the historical record that HCB, famously anti-crop, cropped this one print. [Corrected in review round 3: the record says the crop removed intruding fence at the LEFT and a chunk of the BOTTOM. The bottom trim matters structurally — the bottom of this frame is the silt corner with the hoops, the instrument's top-leverage accent cluster, so HCB edited directly against the zone where the punctuation system lives. A small print in his Scrapbook at the Fondation HCB shows the full frame with his cropping marks: the crop-delta experiment is runnable if a reproduction is sourced.] Also absorbed at round 3: the frame we measure was authored in 1946–47, not 1932 — HCB found the image in his contact sheets while preparing the MoMA retrospective, and the first prints were made then at Chim Seymour's New York lab. The equilibrium this study measures was recognized and constructed at the enlarger, fourteen years after the shutter, exactly where "a precise organization of forms" gets decided. Second, film grain here is the medium, not contamination: the study runs native (Entry 02), and the null battery, not denoising, is the grain-guard. One naming repair: the posters advertise the pianist Alexander Brailowsky (the B occluded — hence RAILOWSKY), and HCB stated he had not noticed the posters until afterwards; that testimony is now cited in the scope table's Intent row.

Colour subsystem tabled on sight (monochrome print): β reads 0.220 and coupling 0.220, and the persistence pass reports fragile_to = grayscale — all of it the warm scan tint, not colour at work (the I-12 lesson in its purest form). Tabled by agreement, to return only if it earns its way back.

Entry 02

PM diagnostic: grain moves facture, not composition [stable]

One Perona–Malik read, diagnostic only (pm_check.py). Verdict: meaningful grain energy, with a clean split. Composition-tier metrics barely move; facture-tier metrics move exactly the way I-11 predicts when PM softens edges.

Decision held from the handoff: the study runs on the native image, pipeline fixed; the PM'd read is never compared against non-PM works. Coarse composition claims below are grain-robust by this table; fine-structure claims stay grain-limited (I-10).

Entry 03

Headline: a still, counterweighted frame [stable]

State counterweighted, regime gradient-field / controlled instability, field regime QUIET_MODERATE. The celebrated stillness of this photograph is not a mood, it is the number set: imbalance 0.040 (raw mode 0.061), offcenter_force 0.041, mass centroid (+0.031, +0.048) a whisker right-down of frame centre, radial verdict neutral (dRC +0.005, RC_f 0.897: strongly centre-compliant), DGI 53.1 "leaving the basin" (cf. Klimt 54.9, Bruegel 44.4). For a frame whose subject hangs mid-air, the composition is remarkably becalmed: the leap is suspended inside a resolved structure, not disturbing it.

Entry 04

The anchor is the fence band; the leaper hangs off the most resistant corridor [stable]

Island 1, the anchor: the entire upper dark band (hoarding + posters + fence + station roofs) fused into one density component, 15.4% of frame area, 24.3% of total attention, centroid (+0.043, -0.203). The dim standing figure behind the fence (the witness nobody writes about) stands at the anchor's centre of gravity. [Corrected in Entry 23: the first draft attributed μ 0.505 to this band — wrong subsystem. The kernel's largest GRADIENT-mask component is a different object entirely, pixel-verified: a welded web in the mirror world (reflected pickets, the leaper's rim, his reflection, the ladder ring), centroid (0.56, 0.63). The upper band is dark-on-dark, nearly gradient-invisible; it dominates the density/island system instead. Two subsystems, two biggest things; both true, now correctly attributed.]

The leaper is island 3: 2.3% of area, 5.9% of attention, centroid (+0.400, +0.059). His relation to the anchor is the frame's flagged resistant corridor (r=0.65), the costliest connection in the graph; the strong conduits run from the anchor to his reflection (island 20, r=0.57) and to the silt corner with the hoops (island 19, r=0.54). Read structurally: the composition's easy paths flow to the echoes; the man himself is held apart, the one region the field resists joining.

Entry 05

The accent system skips the celebrated triad entirely [stable, pixel-verified]

24 accents (lens-capped top-24: positions, never abundance on a grain surface). Where they land: the iron hoops in the silt corner (top leverage 1.47), the clocktower and roof ridge (1.47), ripple-ridges and reflected fence pickets at bottom-centre, the steam plume behind the fence (light filament 1.13), and a light patch at the puddle's far edge [corrected in Entry 20: first labelled "the ladder rungs" — the uniform audit shows the box sits at y≈0.52, above the ladder; the ladder carries no top-24 accent]. Where they do not land: the leaper carries zero accents; his reflection zero; the poster-dancer zero. Three boxes that plot near the man sit, on pixel inspection, on the reflected fence band and the witness's reflection, not on his body.

Entry 06

The subject is a mass hole [stable]

The field panels show the man's interior cold in both mass and support: a black silhouette is gradient-flat inside, so structurally he is a hole with a rim — his contour carries all his mass, his body carries none. The same holds for his reflection and the witness. His rim is visibly fainter than the static world's edges (motion blur vs the crisp reflected pickets beneath him): the one moving thing is the least committed figure in the frame. That claim in numbers is queued probe work (prediction 3).

Tonal economy: almost exactly equal thirds — shadow 0.354, midtone 0.332, highlight 0.314, τ 0.425, mean zone 4.34 — but bimodal underneath (zone 1 holds 0.210, zone 8 holds 0.243; the middle thins). A two-pole tonal system in balance: the dark inventory (band, silhouettes, silt, hoops) against the light one (sky, water sheet). The subject class holds a third of the image yet contributes neither the anchor's attention share nor a single accent from the triad.

Entry 07

Edge-mode and crop stability: mode-robust read, crop-sensitive frame [stable]

Both edge modes run (content touches all four edges, I-1). All headline metrics hold: kernel, tonal, radial, DGI unchanged; raw mode amplifies the rightward mass slightly (mass nx +0.031 → +0.061), promotes 4 → 6 counterweights, imbalance 0.040 → 0.061. No material disagreement; edge-aware quoted throughout.

The islands diagnostics flag crop stability "moderately sensitive" (score 0.246; a 4.5% inset moves mass-centre by 0.057). On most works this is a caution label. Here it reads as a fact about the work: this print is already a crop HCB chose (Entry 01), and the balance Entry 03 measures is tight enough that shaving 4.5% moves the centre visibly. The authored frame is load-bearing; the composition has no slack at its edges.

Entry 08

Edge-suppression audit: the left-edge stripes were suppressed, and they are real [stable, pixel-verified]

User review of Entries 01–07 flagged exactly the right thing: the readouts ran edge-aware only, and there are vertical white stripes at the left edge, mid-height, that suppression might hide. Rerun with suppression off: the accent lens is capped at 24, so raw mode churns the mid-leverage band, and among the newcomers are three accents at the extreme left edge (nx −0.47 to −0.50, x ≈ 5–37 px, leverage 0.83–0.93): on pixel inspection they sit precisely on the reflected white hoarding-stripes beside the reversed RAILOWSKY lettering. Real content, not scan artifact; edge-aware had them under the border ramp. The top of the hierarchy is mode-stable (hoops ≈ clocktower ≈ 1.4+ in both), and the raw newcomers that plot near the man land on the handcart behind the fence, the fence reflection, and a ripple ring. The triad stays accent-silent in both modes: Entry 05 survives with a stronger pedigree, and the accent inventory gains a left-edge member, the poster-reflection stripes.

Entry 09

Edge-commitment probes: strong but unresolved [stable, jitter-checked]

Prediction 3 said the leaper would be the frame's least edge-committed figure. Mass-locked rim probes (Otsu-segmented object inside each box, metrics on the rim band only, I-3/I-4 discipline):

Scored honestly: directionally right, wrong on uniqueness. The leaper is softer than the static built world (vs fence: jitter-clean, 0.259–0.274 against 0.339–0.342 across ±5% box perturbations, no overlap), but the hazy clocktower and the hoops sit in his commitment cohort. The sharper finding is the energy column: his rim carries the most gradient energy in the probe set (0.550) at bottom-cohort commitment — strong but unresolved, the motion-blur signature. And the witness is his exact inverse: weakest energy (0.230), highest commitment (0.450), faint but sharp. The two human figures occupy the two ends of the commitment axis: the mover burns bright and soft, the watcher faint and hard.

Entry 10

Persistence: blur-proof composition, tint-fragile anchor [narrowed]

Full read with persistence: 0.722, and the per-degradation split is the finding. Blur 0.928 — the island system survives blur almost untouched (anchor drift 0.012). This composition lives in big notan masses, not fine edges: the exact opposite of the Bruegel, which was fragile_to blur. Low-contrast 0.800, robust. Grayscale 0.437, anchor does not survive (drift 0.513) — suspicious on a monochrome print, so it was looked under rather than reported: stripping the warm scan tint directly moves the top island from the fence band (+0.14, -0.23) to the silt corner (-0.20, +0.25) and flattens the hierarchy (entropy 0.826 → 0.893).

Entry 11

Candidate battery + nulls: R_spatial lands where registered; address splits [stable]

C-3, R_spatial (logged to CANDIDATES.md, n→9, third domain): real 0.164–0.659, subadditive — prediction 4 called ~0.2–0.7, the Bruegel distributed-order camp, before the run. Under matched-stats nulls it collapses to additive/cut-sensitive (phase-scramble 0.215–3.558, patch-shuffle 0.977–1.726): the subadditivity needs the intact composition. Entry 03's balance is cancellation, measured: the whole is quieter than its loudest part.

C-4, address: central frontality +0.641 — far above the predicted 0.4–0.6, second only to the Mäda (+0.663) in the corpus (a comparison to flag: the Mäda's "figural" classification is itself single-draw and owes the same seed sweep this study ran; the raw scalar comparison stands, its classification is known-unreliable) — but not figural: the strongest symmetry axis sits far off-centre (offset -0.348, lock 0.304). [Corrected by the Entry-19 seed sweep — read this version.] The single scramble draw (+0.475) suggested a "spectral floor + composition margin" split, and the first version of this entry called that a new kind of result. Twenty seeds dissolve it: phase-scramble address is +0.522 ± 0.067 with max +0.638, so the real +0.641 sits at the top edge of the null distribution. Address on this work is spectral — the banding reproduces essentially all of it with zero arrangement — which is what prediction 5 said before the run. The pre-registration was right and the single-draw revision was the error. (Patch-shuffle: +0.076 ± 0.115 — destroyed, as expected.)

Spine scalars under nulls (prediction 8): mass_divergence 0.0342 → 0.0136 / 0.0261 — reproduced by patch-shuffle and tiny to begin with: attention and weight converge here, no compositional purchase, set aside (I-15). DGI 53.1 → 58.1 / 61.2 — both nulls score higher default-gravity, the corrected-Bruegel direction (the real arrangement resists the centred basin more than its own statistics would); margin small, reading aid only. [aid]

Entry 12

Same-coordinate compare (n=6): the photograph sits inside the painting corpus [stable, placement]

vs Bruegel / Degas / Pollock / Klimt / Caravaggio, one coordinate space (raw value + rank; at n=6 z-magnitudes are meaningless, the Study-7 lesson). Rank stability across modes and scales was added in Entry 18; this entry quotes the stable form. Where the photograph tops the group: theta 0.059 (rank 1/6 in both edge modes) — the strongest single-orientation field in the corpus, the fence/waterline banding as measured fact; centre weight 15.4% (1/6); TLG 1.000 (1/6, tied at the plane-differentiated top with Bruegel, Klimt, Caravaggio). Where it sits lowest: imbalance 0.040 — most balanced in the group edge-aware and at all sweep scales, second to Pollock (0.043 vs 0.061) in raw mode: quote as top-two, not first. Torque second-lowest beside the Bruegel (the subadditive-cancellation pair). Mid-pack and rank-stable: μ 0.505 (2/6 at every mode and scale), τ 0.425 (4/6 everywhere), DGI 53.1 (2/6; 3/6 at 768px), mass divergence 0.034 (4/6). gap_fraction is withdrawn from this table (Entry 18): Entry 02 classified it facture-tier and grain-dominated (0.534 → 0.732 under PM), so quoting it as a cross-media separator against painting reproductions measured emulsion-vs-canvas, not composition — the original version of this entry violated the study's own tiering. The medium-independence evidence: the photograph interleaves with the paintings on composition-tier axes, pairs with the Bruegel on distributed order, and separates on orientation stability — an axis a lens plausibly should top. (Its high coupling_index z is quoted nowhere: scan tint, Entry 01.)

Entry 13

Review corrections: the witness box, and the user's stripe reading [stable]

Two items from user review. First, the witness box was right-shifted (0.54–0.60 caught his right half plus fence). Corrected to the figure (x 562–630, y 432–585) and re-run: commit 0.450 → 0.450, energy 0.240 → 0.230. The numbers did not ride on the misplacement, and the honest caveat is now attached: the witness stands against co-dark pickets, so any tonal mask of him is a figure+picket blend. Second, a reading the instrument's Entry-08 recovery supports but cannot originate: the recovered left-edge white stripes are the value-inverse of the fence pillars — the vertical rhythm continues through the mirror world with polarity flipped, and those light verticals push the eye back up into the frame. Logged as a user-contributed compositional reading anchored to a measured accent group.

Entry 14

Adversarial: is the triad's silence a cap artifact? The heel enters at rank 77 [narrowed — rescopes Entry 05]

Entry 05 was measured under the lens's top-24 cap; the obvious adversarial question is whether the triad sits just below it. Rerun with max_accents=200: 125 candidates. A first box-membership test flagged many "hits," but every checked one sat on the reflected fence, steam, handcart, or ripples — the same co-dark confusion as before — so the test was upgraded to point-in-object-mask with pixel verification of survivors. The honest result: the man's body enters the candidate field in the bottom third of the extended list — two candidates at his front boot, leverage ~0.69, half the hoops' 1.47. Precision notes (Entry 18): leverage is cap-independent (top-6 values match across the 24- and 200-cap runs to ±0.005), so the "half" ratio is valid; the rank (~77 of 125) sits in the churn-prone mid-band documented in Entry 08 and is quoted as "bottom third," not as a point. His reflection and the poster-dancer have no clearly on-body candidate at any rank. Entry 05 rescopes from "silent" to: silent within the lens; open the lens five times wider and the first thing the instrument finds on the man is the heel — the decisive-moment pixel itself. The punctuation hierarchy stands; its bottom now has a name. [exploratory]

Entry 15

Adversarial: resolution sweep (I-10) — composition tier passes [stable]

Downsampled reads at 1536 / 1100 / 768 long-edge. State counterweighted at all three scales; μ 0.505 / 0.473 / 0.473; τ 0.425 / 0.423 / 0.423; imbalance 0.040–0.042; 20 islands at every scale; shadow_mass flat at 0.353–0.354. gap_fraction wobbles 0.534–0.577 (facture-adjacent, expected). Every claim quoted in Entries 03–12 survives resampling; nothing finer than grain scale is claimed anywhere in this study.

Entry 16

The doubling, measured: the water repeats the man most literally [stable]

The received reading's central fact, measured directly: flip each object strip vertically and scan it down the water for best normalized cross-correlation — now with a chance floor (Entry 19 fix): each strip is also block-shuffled (pattern destroyed, statistics kept) and re-scanned, 20 seeds. [Arithmetic corrected at round 3 — the producer's excess is computed against the null-best MEAN; the first prose paired it with the max.] Leaper ↔ reflection: NCC +0.642. Null-best mean +0.164 (excess +0.478); strictest comparison, the null-best MAXIMUM across 20 seeds, +0.288 (excess +0.354). Far above chance either way. Fence: +0.303 vs null max +0.290 — inside the noise at 20 seeds, indistinguishable from chance. Hoarding: +0.202 vs null max +0.233 — does not clear chance. The ranking sharpens under the floor: the man's mirror is the only doubling that beats a pattern-destroyed control. Caveat kept: the leaper strip's offset curve is high everywhere in the water (mean +0.516 — coarse dark-on-light layout), so localization is weak even though the pattern match is strong. The Entry-09 expectation (static doubles sharply, mover softly) inverts at pattern level; the mechanism is geometric and honest: he leaps close to the water plane, so his reflection survives at nearly full scale, while the fence and hoarding reflections are foreshortened and broken. And the axis framing is corrected (Entry 18): the mirror axis at y ≈ 0.667 is geometrically determined by the waterline, which is fixed by camera position — the probe did not discover a compositional axis; HCB placed the waterline on the lower third, and everything mirrored inherits it. [exploratory]

Entry 18

Hard-pass fixes: caps, tiers, ranks, edges [stable]

A user validation pass (ten items) drove this round; the four that bit. (1) gap_fraction withdrawn from cross-media claims: Entry 02 tiered it as facture/grain (0.534 → 0.732 under PM); Entry 12 then quoted it as a corpus separator — a violation of the study's own tiering, fixed in place. (2) 20 regions is a segmenter cap: max_islands=20; HCB, Bruegel, and Degas all sit at the ceiling, and active_islands is 10 for every work in the corpus — Entry 03's "island signature twin" claim is rescoped to the axes that genuinely pair the works, and island counts at 20 should never be quoted as convergence anywhere in the corpus. (3) Leverage is cap-independent; ranks are not: top-6 leverages match across the 24- and 200-cap runs to ±0.005 (Entry 14's ratio is valid); the heel's rank is quoted as bottom-third. (4) Rank stability swept (modes × scales): theta 1/6 everywhere; μ 2/6 everywhere; τ 4/6 everywhere; DGI 2/6 (3/6 at 768px); imbalance 1/6 edge-aware and at all scales but 2/6 raw — requoted as top-two. And the four-edge suppression audit the reactive Entry-08 fix owed: with suppression off, border-zone accents appear only at the left edge (the three stripes); top, bottom, and right border zones contribute zero. The one edge a human flagged was in fact the only edge with suppressed content.

Entry 19

Seed sweeps: the address "split" dissolves; R_spatial nulls are chaos [stable — corrects Entry 11]

The null battery ran on single draws; null_seed_sweep.py sweeps seeds. Address: 20 phase-scramble seeds give +0.522 ± 0.067, max +0.638. [Round-3 correction: at 20 seeds the real +0.641 topped every draw (p ≈ 0.048, ~1.8σ) — weak evidence FOR a small margin, and this entry's first "wholly spectral" conclusion over-corrected. The required 100-seed run settles it: scramble mean +0.549 ± 0.075, max +0.685, with 10 of 100 draws exceeding the real value — empirical p = 0.109, z = 1.22. By the rule stated before that run, spectral-dominated stands with no resolvable arrangement margin — the conclusion survives, now on evidence rather than overcorrection.] (Patch-shuffle at 100 seeds: p = 0.0099, z = 4.98 — destroyed, as expected.) R_spatial: 5 seeds per null show the null verdicts are seed-chaotic (phase-scramble r_max spans 0.22–10.97; patch-shuffle 0.14–3.10, including subadditive-looking draws). The corrected claim: not "dies under the null," but no null draw reproduces the work's stable verdict — the nulls produce a different verdict every seed, the composition produces the same one every time. Both corrections propagated to CANDIDATES.md, with a methods note that the corpus's earlier single-draw null rows owe the same sweep.

Entry 20

Uniform accent audit: the top-24, every box inspected [stable]

The verification-asymmetry item: pixel-level escalation had been applied only where the narrative was at stake (accents near the man). The audit gallery puts all 24 through the same inspection. Result: 24/24 sit on nameable structures — hoops and ripple arcs (~11), clocktower and roofline (6), reflected fence and the witness's reflection (4), the steam plume, a light patch at the puddle's far edge, a roofline finial. Zero accents on bare grain; one borderline (a light ridge adjacent to a hoop arc); and one analyst mislabel caught: Entry 05's "ladder rungs" accent is actually the waterline light patch — corrected in place. The ladder itself carries no top-24 accent, which Entry 21 makes structural sense of. Demotion count from uniform audit: 0 instrument, 1 label.

Entry 21

The second double, and the exception family [exploratory]

The poster-double (the most important gap). The received reading says the picture doubles twice: below (water) and behind (the RAILOWSKY dancer). Silhouette correlation, scale-normalized, mirror allowed, ranked against distractor shapes: reflection +0.47 (the probe selected the vertical flip unprompted — it independently identified the water double as a mirror) > clocktower +0.34 > dancer +0.24 > witness +0.19 > hoop +0.13 > ladder -0.08. The dancer does not out-score the best distractor: at this reproduction's scale, the lateral rhyme is not recoverable as pre-semantic silhouette correlation. Caveats before concluding "semantic, not structural": the dancer is ~57×65 px of grain (the limb geometry carrying the rhyme is mostly gone — I-10), the leaper mask fuses unavoidably with the co-dark band behind his head (stable across opening settings), and blob-NCC rewards compactness (the clocktower's solid rectangle scores against any centred blob). The honest finding: the two doublings are different in kind — the water double is structural at every level tested; the poster double lives at or below what this file plus a pre-semantic measure can see. Re-run on a higher-fidelity source before "semantic" is claimed.

The imminence gap. The not-yet interval between heel and reflection: 32 px = 0.021 of frame height (the boundary is motion-blurred on one side and rippled on the other, so allow ±a few pixels of definitional slack). The narrative charge of the picture hangs on a gap of two percent.

The heel against the master lines. Heel (mask's lowest point, the front boot) at (0.875, 0.659): its height rides the lower-third line (Δ -0.008) — the same line the waterline mirror axis sits on, necessarily nearby — but horizontally it is nowhere golden (Δ +0.208 to the 2/3 vertical). The Lhote-geometry folklore gets a split verdict: the vertical placement of the decisive event is on the third; the phi-vertical story finds no support at the heel.

The ladder hypothesis, refuted. Proposed as the frame's one committed diagonal, the orientation exception twinned with the leaper. Measured: the ladder is the most orientation-coherent, most elongated object probed (coherence 0.76 vs frame 0.06; elongation 6.3) — but its picture-plane tilt is +1.6° from horizontal. It is not a diagonal on the picture plane at all; it is a depth vector that reads diagonal only through foreshortening. The hypothesis dies, and usefully: the leaper remains the frame's only true exception, now in orientation as well as mass, commitment, and balance.

Edge clearance. The leaper's leading extremity stops 0.090 of frame width from the right edge; his head clears the fence band by ~0.04–0.06 of height. Tight, consistent with Entry 07's no-slack frame; quoted as description (no corpus baseline for clearances yet).

Entry 23

Round 3: the required fixes, and what the armor checks found [stable]

The two required items, resolved. (1) Address at 100 seeds (Entry 19, corrected in place): the 20-seed "tops every draw" was a small-n artifact; at n=100, 10 draws exceed the real value (p = 0.109) — spectral-dominated, no resolvable margin, by the rule stated before the run. (2) The doubling excess arithmetic traced to the producer (excess is against the null-best mean) and requoted both ways in Entry 16; the fence's doubling requoted as indistinguishable from chance.

The recommended armor checks, run. Mask-fusion threshold sweep: μ rises smoothly 0.425–0.639 across percentile 80–90 with no weld/split cliff, and the largest component's centroid stays put (cy 0.61–0.64) at every step. No bridge-pixel fragility — but the sweep exposed a misattribution: the welded half-of-the-mask is not the upper band. Rendered below, the largest gradient-mask component is a single web in the mirror world: the reflected pickets, the leaper's rim, his reflection's outline, the ladder ring. The upper band is dark-on-dark, nearly invisible to the gradient mask; it dominates the density/island system instead. Entry 04 corrected in place. The corrected fact is stronger than the original: the relation system holds the man apart, while the frame's largest connected structure physically runs through his rim — he is welded into the mirror lattice he is about to break.

Lettering polarity: the direct RAILOWSKY lettering carries 0 of 125 accent candidates; its reflection carries 4 (best at rank 43). The lettering speaks only in the mirror — dark-on-dark against the hoarding, it becomes leverage only against the brighter water. Same grammar as the stripes (Entry 08): in this tonal system the mirror world is where the upper band's contrast gets paid out.

Lhote rabatment test (snap threshold 0.02 stated before the run): the punctuation family ignores the 2:3 armature — the heel misses the bottom rabatment line by -0.0637, and Entry 21's thirds hit stands as thirds, not rabatment. But the two density poles snap: the reflection island's centroid sits on the bottom rabatment line (Δ -0.0017) and the anchor band's centroid on the top one (Δ +0.0197 — threshold-riding: it clears the pre-stated 0.02 by 0.0003 and would fail at 0.019). Baseline with its denominator: 9 points × 2 rabatment horizontals = 18 point-line tests at ±0.02 tolerance, expecting ≈ 0.7 snaps under uniform placement; observed 2 (the frame diagonals were also tested: zero snaps), landing on the two structurally central masses, not on random members. Suggestive at n=2, one threshold-riding, quoted as [aid]: if it holds on the HCB mini-corpus, the Lhote geometry lives in the masses, not the incidents.

Entry 22

Synthesis, revised after the validation rounds [reading]

The instrument's account, assembled from Entries 03–23 with all three validation rounds' corrections folded in: this photograph is a becalmed, banded, cancellation-balanced structure whose celebrated subject is present as an exception, not as an emphasis. The machinery of the composition is in the static world: the upper band dominating the density system (its "anchor" label tint-assisted and crop-sensitive — Entries 07 and 10 — with the witness at the leading component's measured centre), while the gradient system's largest welded piece is the mirror lattice that runs through the man's own rim (Entry 23); four counterweights; the strongest single-orientation field in the corpus (theta 1/6 in both modes — the fence rhythm); punctuation spent on hoops, clocktower, steam, and the value-inverted stripe rhythm of the mirror world; torque that cancels into the same stable subadditive verdict at every cut while its matched-stats nulls produce a different verdict every seed; big notan masses that shrug off blur (persistence-blur 0.928). Against all that, the leaper is 2.3% of area, 5.9% of attention, accent-silent within the lens, mass-hollow, connected to the anchor by the frame's most resistant corridor, and — after Entry 21 — the frame's only exception: the ladder, proposed as the orientation exception, measured out at +1.6° from horizontal. What the man owns is of a different kind: the highest rim energy at the lowest commitment (strong but unresolved — the blur of the still-happening), the frame's one chance-clearing mirror (+0.642, clearing even the strictest null comparison — the 20-seed null-best maximum — by +0.354, on the lower-third waterline HCB chose), and a two-percent interval of not-yet between his heel and its double. The "decisive moment" is measurable in this coordinate space, but not as sharpness or weight: as the one energetic, uncommitted, doubled event inside a fully resolved structure. The structure is the stillness; the subject is the exception the stillness exists to hold.

Two closing readings, clearly labelled as readings. First: the instrument, knowing nothing, spent one of its two top accents on a clocktower — the frame of the photograph that defined the decisive instant carries a clock in its punctuation system. Second, the user's reframe, which this study adopts: the ISR never saw the moment. It cannot; there is no time axis in a gradient field. What it reconstructed is the equilibrium of the rectangle — why this frame stays visually stable regardless of when it was taken. HCB's own definition of the decisive moment had two parts, the event and "a precise organization of forms," and nearly all writing stops at the event. The instrument reads only the second part, and finds it sufficient to account for the picture's stillness. On this evidence the phrase runs backwards: it is not that the moment made the picture; the organization was in place, and the moment is the one thing in it the organization cannot close over — the decisive frame, holding an undecided event.

Corpus placement makes the same point from outside: the photograph interleaves with the paintings on composition-tier axes (Entry 12, ranks swept in Entry 18), pairs with the Bruegel on distributed order and torque cancellation, and separates on orientation stability. The displaced-subject lineage now has three members in three media — Icarus in the corner of a panel, the abonné at the edge of a pastel, the leaper mid-air in a photograph — and in all three the instrument reads the same grammar: the subject is what the composition points away from, holds apart, or refuses to weight.

What This Does Not Prove

Scope, stated plainly

Scorecard

Pre-registration, read against (running)

Status

Lab book complete through Entry 23 (three validation rounds)

Round three resolved both required items (address at 100 seeds with empirical percentile; the doubling excess arithmetic traced to the producer), ran the recommended armor checks (mask-threshold sweep, which found and fixed the μ misattribution; lettering polarity; Lhote rabatment), and absorbed the historical corrections (the bottom crop, the 1946–47 printing date, the Brailowsky naming with HCB's own testimony on the posters). Awaiting user confirmation to start the critique. Round-3 artifacts: address_seed_100.py/.json, threshold_sweep.py/.json, lettering_rabatment.py/.json, mu_component.png.