Soejima Taneomi — Two-Line Calligraphy八朶青蓮馬前堆 / 一杯滄海李中寒 · hanging scroll, ink on satin · Met 829410

Entry 01 — Corpus, condition check, design

A claim that happens to be measurable

Soejima Taneomi (1828–1905), Meiji statesman-calligrapher. Two columns of Chinese verse — "Ma faced eight blue lotuses standing erect in front of him. / Emperor Li faced a vast ocean in the midst of deep cold" (tr. adapted from Sadako Ohki). The curatorial label asserts odd proportions, intentional distortion, stroke order and direction departing from the norm, shattering the expectation of regularity and symmetry. Every noun in that sentence is structural — irregularity, asymmetry, proportion variance, axis drift — measurable without reading a single character. But "departure from a norm" needs a norm, and the instrument has none. So the norm was sourced as a comparator corpus (Met Open Access throughout, matched documentation):

Domain shift named before any numbers: ink on woven satin is near-binary tone, no chroma except seal vermilion; content is line, gesture, negative space. Predicted misbehavior — β and coupling near-degenerate (the only chroma is the seals), radial compliance likely uninformative, regime/profile labels forced Western, thin-stroke measures unclaimable without a resolution sweep (I-10), satin weave as a texture floor. Each prediction was checked, not assumed; two died (see Entries 05–06).

Entry 02 — Where does the image start?

The honshi boundary, measured not eyeballed

A scroll has no frame. It has a nested stack of textured zones — photographic ground, roller, outer mounting silk, patterned cloud-damask, then the satin sheet (the honshi). Every transition is a real edge to the instrument, so "which pixels are the artwork" is itself the first measurement.

v2: per-pixel bright mask → largest connected component → fill the ink holes morphologically (strokes become interior holes of the bright substrate and are absorbed) → bounding box. Both photographs, segmented independently, agree with each other and with the physical object: detected aspect 3.288 (full mount) and 3.270 (tight) against the Met's catalogued sheet, 146.6/44 cm = 3.332 — within ~2%.

Source of record: the honshi crop of the tight photo (1095×3578). The mounted whole remains a separate, labeled read about presentation; the raw photographs are diagnostics only.

Entry 03 — The corpus gets the same treatment

And the cross-check catches a leak

The segmenter ran on all six works, each checked against an external ground truth it cannot see: the Met's catalogued physical dimensions. Five passed at 0.02–2%. The Gochiku Banzai failed at 19% — its mounting silk is pale blue, luminance 0.671, brighter than the fixed threshold, and saturation cannot separate it (mount 0.082, honshi 0.082 — identical). The threshold is now adaptive (Otsu on above-ink pixels), and the fix pulled every work under 3%:

Entry 04 — First read of the subject honshi

What the instrument sees, checked against the sheet

Full read at the library's working scale (2600px, Lanczos, edge-aware, --no-persistence this pass). The islands are character clusters; the corridors run down the two text columns; nothing fires on mount or damask — the crop did its job.

Entry 05 — One coordinate space; the torque signal

The story arrives

Before the compare: the seal check. The only chroma on the sheet is vermilion seal paste — the purest I-7 case the library has seen. The chromatic accent channel finds the seals (4 accents, three on the upper-right seal) — but they are not the headline. The top chromatic accents sit inside brushstrokes, on dry-brush passages where warm satin shows through split bristles: kasure, "flying white," the ink-load and speed evidence a calligraphy connoisseur reads first.

Then the compare — six works, matched honshi crops, one coordinate space. The dispersion table's top rows say where this corpus genuinely differs: dRC, coupling, μ, torque, construction diversity. The torque column (stance lineage — the library's dynamism axis, C-2 "artists torque, AI stands"):

Also banked: subject τ 0.796 with 79.7% highlight mass — the anti-Caravaggio pole (Study 4: 90.4% shadow); the library now brackets both extremes. One hypothesis died on contact: gap_fraction does NOT separate satin from paper (subject 0.614 vs Duzhan-on-paper 0.618) — the satin-bleed story of Entry 04 is not supported at whole-image level. Island entropy is high across all six (0.80–0.95): dissolution is the register's ground state, not a distinction.

Entry 06 — Three verifications before any critique

The claims earn their strength labels

1. Torque survives everything. Resolution sweep 1300→3300px, three works; subject drift ±0.6%; determinism exact (path curvature has no stochastic step, unlike Hough — I-8 does not apply); subject/norm ratio 2.96–3.16× at every scale. The cursive sibling wobbles at 3300 (0.34→0.17) — noted; the direction only widens the attribution contrast. Claimable at headline strength.

3. β is measuring the substrate — quantified, not just suspected. Real image 0.726; chroma-stripped copy 0.000 (the measure responds to chroma only — it is honest); bare-satin patch 0.837, higher than the whole image. On this material the "structural void" is by definition the warm cream substrate: β = substrate tint, not "color working where drawing is absent." The number is real; the painting-culture interpretation does not transfer. Caveat class: any substrate-dominated image — calligraphy, drawings, prints.

Entry 07 — Full spine; the grayscale surprise

The one chroma relation is load-bearing

Full read with persistence: 0.813 overall — blur 0.877, low-contrast 0.970, grayscale 0.593, fragile_to: grayscale; the anchor does not survive chroma-stripping (drift 0.159, +3 islands, entropy +0.13). A near-monochrome ink work fragile to grayscale is real, and it is the β lesson's inverse face: ink on toned satin is cool against warm, and that single opponent relation carries structural weight in the mass construction. Instrument-relative — not a claim about how humans see B&W photographs.

Rounding out the spine: saliency÷mass divergence 0.0135 (lowest in corpus — the eye is sent where the weight is); the top two counterweights are the two column-heads (八朶 0.047, 一抔 0.046) — the sheet hangs from a beam across its top; mass grid center 8.8% "distributed", asym 0.068 "symmetric"; DGI 55.7. The static map is quiet; all the wildness is cross-scale.

Entry 08 — Methods note: why this corpus, and the n question

Built, not sampled

This corpus is six works because the question is not "where does Soejima sit in the population of calligraphy?" — it is "does THIS scroll depart from ITS norms?" That is answered by placed comparators, each holding one variable still: the orthodox running-script norm (the departure claim's reference point), the artist's own other cursive (the attribution control — separates the piece from the habit), and the seal-script calibration poles (which verify the axes track regularity before we lean on them). Scaling n does not sharpen this design; it dilutes it — forty more scrolls regress every axis toward a mushy corpus-center and bury the matched contrasts that carry the inference. The z-scores are descriptive placement, never significance; the inferential work is done by the structure of the comparisons (norm / self / pole), not by sample size. Where a cell is thin we say so in What-This-Does-Not-Prove ("the norm is one hand") — the honest form of the n caveat, and the only form this design owes.

Source discipline: all six are museum-catalogued works from one collection (Met Open Access), one photographic regime, published physical dimensions — which is what made the Entry-03 cross-check possible at all. Canon status is not an aesthetic endorsement; it buys documented provenance, stable reproductions, and citable ground truth.

Entry 09 — Substrate audit; the torque decomposition

The headline goes on trial and comes back transformed

β was caught measuring the substrate; the honest follow-up is to run the same three-way test (real / chroma-stripped / bare-satin) over the whole spine — find the next β before trusting anything. One catch: cbi joins β in the substrate class (bare satin 0.275 vs whole-image 0.136 — weave sheen draws chroma boundaries light doesn't). Side-finding: only ONE 480×190 window on the entire sheet qualifies as ink-free — the "empty" satin is haunted by the ghost stroke and dry traces.

Then the torque stress: real 1.992, chroma-stripped 1.992, bare-satin 0.051 — but binarized ink-mask 0.057. Apparent collapse: neither substrate alone nor committed strokes alone produce the torque. The decomposition (three works × four variants) resolves it:

Entry 10 — Census and map: amount vs deployment

Not more breath — breath placed to carry force

Prediction going in: the sibling is dryness-poor. Three measures, three refutations of amount: accent counts are capped by lens design (gap_lens returns top-24 for every work — logged as a lens fact); dry-interior area ratio is confounded by stroke spacing (closing bridges the gaps in fine spidery writing — Duzhan reads highest); width-independent midtone-per-ink puts the subject mid-pack again (0.249 vs Duzhan's 0.426). Dryness abundance is register physics — everyone writing cursive has breath. Combined with Entry 09 — only the subject's torque is carried by that layer — the finding sharpens: not more breath, breath deployed. Amount is common; function is the choice.

Entry 11 — Stratigraphy built; the carrier story corrected

The tool's first act was to break its builder's story

toolkit/patterns/stratigraphy.py: stratify → ablate (delete / flatten / isolate, explicit fill) → attribute — "which layer carries this number" as a one-liner. Its first carrier table contradicted Entries 09–10 three ways, and chasing the contradiction found the real result: ablation attribution is fill-dependent for whole-field relational metrics — the same mask read 0.08 with a median fill and 2.05 with a white fill, because the fill level redefines every remaining pixel's deviation. The fill-invariance matrix separates what survives:

Entry 12 — Implied time: the depletion cycle (PRELIMINARY)

Reading process from a still image

The end-state encodes process: a brush depletes as it writes, re-inking resets the cycle, and the canonical reading order is known. Per-character dryness vs reading position, first probe:

Entry 13 — The verification battery refutes the arc

Built to refute; it succeeded

The area confound turned out to be the wrong one (it passed: r(area,pos) = −0.09, the late characters are not larger). The stroke-width confound is the live one and it lands: breath_share is strongly width-dependent (r = −0.66; thin strokes carry more marginal breath per unit ink regardless of load), so the "rising arc" tracks the left column's thinner cursive characters, not an emptier brush. The mapping tiles confirmed one real error — a merged strip of thin characters + seal — past which the per-position order was never reliable.

The decisive test is a measure that is width-independent by construction: substrate-bright fraction of the DEEP stroke interior (flying-white you cannot have in a hairline).

Entry 14 — Non-decomposability, adversarial

Proving the claim against partitions in general

Entry 11 proved the torque is owned by no tonal stratum — but that was one partition. Challenged to prove it against partitions in general, the battery (halves, quadrants, thirds, spectral DoG bands, a 200-window random null; each part read as its own image; scale confound controlled by running the same battery on norm + sibling) corrected Entry 11 — into something stronger.

Entry 15 — Attacking the null: it was a boundary artifact

The retraction gets partly retracted

Challenged on whether the ink-load path was truly dry or whether the pipeline manufactured the null: Entry 13 declared it using raw hull components, in which the thin left column FUSES into a merged strip (Entry 13's own flagged mapping error) — an I-5-shaped risk. Re-segmenting with watershed on the distance transform and re-asking:

Self-attack — is +0.25 a watershed min_distance tuning artifact? Sweep it: +0.18 / +0.13 / +0.28 / +0.25 / +0.31 / +0.30 / +0.25 across min_distance 60–120. All positive, stable, not tuned.

The clean positive finding: the distortion is not scarcity

The subject's stroke cores are the most solidly inked of the three (lowest interior flying-white, robust in 3 of 4 parameterizations). The scroll is not running on scarcity. Even granting the provisional late-drift, the baseline is a fully-loaded brush — the distortion is a loaded brush deployed into superadditive structure (Entry 14), not a hand fighting depletion. And the gap_lens cap is neutralized: uncapped chromatic accents run subject 131 / cursive 132 / duzhan 148 — the subject is not more (slightly fewer); the abundance refutation no longer rests on the saturated top-24 count.

Entry 16 — The hull-bloat audit and the library-wide R_spatial sweep

A user-caught artifact, and the new axis it cleared the way for

16a. Caught from the Entry-9 figure: the flat_ink blobs are far fatter than the strokes. Quantified: the closing-hull is stroke territory, not stroke body — 28–39% non-ink, 1.4–1.6× bloat. Entry 9's "substrate deleted" caption was overstated (those cells were already withdrawn as fill-dependent — the bloat is the visible why), and worse, the core-flying- white of Entries 13–15 measured depth into the bloated hull. Re-run on an honest tight mask (r5, bridges intra-stroke streaks only): both live findings survive, cleaner — loaded-brush sharpens (subject 0.066 vs 0.112 / 0.114, separation widened) and the provisional drift holds (+0.26 vs −0.09 / +0.15). Ledger: never measure interior properties on a closing-hull; tight mask for measuring, hull for grouping.

16b. R_spatial pushed across the whole library — is superadditivity general?

Entry 17 — The superadditive pole has two mechanisms

Coarse assembly vs fine assembly — mirror-imaged spectra

Entry 18 — The enclosed-ground check: third masking refinement

An artifact weakens under cleaner measurement; these signals sharpen

User-caught from the boundary figure (the 蓮 highlight): the tight mask's blanket fill_holes swallows large enclosed voids — 蓮's loops, 馬's loop, the lasso circle — as "stroke body." Enclosed ground is not kasure; the discriminator is hole size (kasure streaks thin, ground loops big). Quantified: 4.2% of the subject's old mask was swallowed ground (7.1% cursive, 1.9% duzhan) — real, modest, now excluded via size-limited hole filling.

Entry 19 — R_spatial earns its keep, mostly: the Degas demotion

The youngest load-bearing claim, put through the battery it owed

Also adopted from the same review, for the critique rewrite: (1) the loaded-brush / late-drift tension is real and unresolvable at this resolution and n — the critique will say exactly that; (2) Entry 17's "matches its study's independent conclusion" carries an infrastructure caveat — same-instrument agreement is consistency, not independent corroboration.

Entry 20 — Close-out: reproduction, goldilocks, consolidation

8/8 numbers reproduce; 7/7 parameter configs hold; the study closes

All inline analyses are now local scripts. reproduce_and_goldilocks.py re-derives every headline number from scratch: 8/8 checks pass against the logged values. The goldilocks audit varies every mask parameter around base (tight radius 4/5/7, hole_max 1000/1500/2500, core depth 3/4/6): both live verdicts (subject lowest core flying-white; subject-only positive drift) hold at all 7 configurations. No parameter tuning carries either verdict. Consolidation shipped: critique, final form v3 essay (Context-first narrative, byline, Public Frame, sources, with a visible Record of withdrawn and surviving claims), OBSERVATIONS Study 5 with instrument findings I-12, I-13, I-14 and candidate C-3, the study README, and library doc updates (README, TOOLS, METRICS, CATALOG).

Status

The study's durable spine, after five adversarial passes: the subject's torque is spatially superadditive and spectrally coarse (Entry 14) — a whole-composition property no partition preserves. Every hunt for the distinction in a local/material stratum failed the same way twice (dryness abundance, Entry 10; ink-load depletion, Entries 12–13) — EXCEPT that attacking the second null exposed a boundary artifact and reopened a modest, provisional, subject-specific late-drift signal (Entry 15). And a clean positive stands: the strokes are fully-loaded, not scarce — the distortion is deployment, not depletion.

Study closed 2026-07-07. Everything owed is shipped: critique (final form v3, essay), OBSERVATIONS Study 5 (I-12 / I-13 / I-14 / C-3 promoted), study README, library docs. Open threads passed to OBSERVATIONS: a larger calligraphy corpus for the provisional drift; R_spatial nulls on paintings, a second superadditive exemplar, AI-image behavior; a second reproduction of the subject to bound the photography chain.

Sources & References

Primary objects, all Metropolitan Museum of Art Open Access (CC0): accession 829410 (Two-Line Calligraphy, Soejima Taneomi; poem translation adapted from Sadako Ohki, per the Met label) and the five comparators 897142, 897143 (Soejima), 845130 (Duzhan Xingying / Dokutan Shōkei), 916602, 897146 (Nakabayashi Gochiku). Object records, label text, and catalogued physical dimensions (used for the Entry-03 segmentation cross-check) from metmuseum.org.

• The Metropolitan Museum of Art. Object records and label texts for accessions 829410, 897142, 897143, 845130, 916602, 897146. Open Access, metmuseum.org.
• Biographical accounts of Soejima Taneomi 副島種臣 (1828–1905): Saga-domain scholarly-samurai origins, Foreign Minister 1871–73, the 1873 mission to Beijing, resignation in the Seikanron split, later privy councillor and Home Minister; art name Sōkai 蒼海.
• Nakata Yūjirō, The Art of Japanese Calligraphy (Weatherhill/Heibonsha, 1973) — script registers (kaisho / gyōsho / sōsho, tensho) and the Meiji karayō revival.
• Fu Shen et al., Traces of the Brush: Studies in Chinese Calligraphy (Yale University Art Gallery, 1977) — flying-white (kasure / feibai) and ink-method connoisseurship.
• Lothar Ledderose, Ten Thousand Things: Module and Mass Production in Chinese Art (Princeton, 2000) — the modular expectations cursive writing plays against.
• Rudolf Arnheim, Art and Visual Perception (California, 1954/1974) — perceptual weight and dynamic balance.
• Instrument lineage and prior studies: the ISR library lab books for Matisse, Degas, Pollock, and Caravaggio; the program's published papers at parallaxmetrology.com.

Computational results in this lab book are deterministic and reproducible from the scripts named in each entry (environment: instrument/.venv, Python 3.14); the close-out audit (Entry 20) reproduces 8/8 logged headline numbers and holds both live verdicts at all 7 parameter configurations. Art-historical context above draws on the published scholarship listed here.