A pre-semantic reader finds no structural mechanism for what the tradition says the inscription does, yet independently recovers the two-zone mass and contrasting void the record and scholarship describe, once it learns to distrust its own first lens. It corrects the eye on two counts, once it is pointed at the right channel. The mass: the default edge lens fragments on the brushwork and misreads the picture as a texture field (I-31 a third time; the dome-under-read is construction, a flat wash is edge-quiet at any fidelity, though the degree of foreground fragmentation is entangled with a single aged plate). Only the value squint recovers the real structure, and it is not a single monumental peak but a bipolar dark–void–dark field, the misted foreground carrying nearly the dome's weight. The void: the broad bright expanse is not blank paper but worked. Masking (the aging divided out, the mid-tones pulled) resolves the fog as coherent pale-wash form, 93% in connected bodies, wrapping the rock bases. Working the mist is standard ink technique, not a discovery, and an earlier attempt to prove it by raw luminance was withdrawn (the aged sheet's uneven soiling confounds any blank-vs-wash comparison); what is distinctive, and sourced, is the opposite pole, Jeong Seon's bold, heavy, wet dark against the school's light-and-dry norm. On the samjeol balance claim the legs land in distinct states and none confirms it. The common phrasing, that the ink and seals counterweight the granite "cliffs on the left", is refuted fill-invariantly (removing the cluster makes the picture more balanced, imbalance 0.070 → 0.051), as much a finding about the phrasing: there is no dominant left mass, the composition is right-heavy, and the inscription reinforces that lean. The colour claim (the seals as the whole colour design) is untestable at this fidelity, not refuted, the faded seals lost under aged-paper warmth on a single JPEG. And the fairer claim, that the inscription activates the empty upper-right, was the decisive open test, and it is now run and not supported: the mark is the sole ink in a real blank pocket (consistent with the tradition, the "most isolated mark" overclaim already retracted by a systematic null, 155 cells more isolated), but that pocket is one of several comparable blanks (the upper-left often larger), and the located, singled-out void the reading needs is not there. What the instrument can do here is adjudicate which sense of an "active void" survives measurement: not a located counter-mass (refuted), but the midtone mist acting as the value-separator between the two dark poles, the bipolar structure's own divider. Contrast-separator holds; counter-mass does not. So the tradition's balance, where it is more than that separation, lives in the semantic layer the instrument declines (C-1), and where it made a checkable structural claim the structure does not back it. And where the study opened skeptical, it closes on an earned convergence: the bipolar two-zone mass and what I am calling the contrast-separator void, recovered from pixels, are what the record and scholarship describe from iconography and viewpoint (the official NT-216 record describes the grand granite and reads the white fog in contrast against the dark; the Korean sources give the peak and the foreground house at two opposed viewpoints, 부감법 and 고원법). The pre-semantic reader and the humanists reach the same structure by different roads. The durable products are the bipolar-mass correction and its convergence with the scholars, the void read correctly (worked, shown by masking, though working the mist is standard; the distinctive hand is the bold wet dark), and two method findings from retracted overclaims: a blank-versus-worked luminance test is void on an aged plate, and a method finding from the "most isolated mark" overclaim: an extremum needs a systematic null, not a chosen foil (I-32).
Russell Parrish · Parallax Metrology · 2026. Computational results are deterministic and reproducible from the scripts named in each entry; art-historical context is triangulation, cited in the appendix. More info: www.parallaxmetrology.com, Full critique
The brief: open a regime the corpus had never read, monumental mass against an active void, in ink monochrome, from outside the European and American traditions (the corpus had one non-Western work, Soejima's calligraphy, Study 5). Jeong Seon (1676–1759), founder of Korean true-view landscape (jingyeong sansuhwa), painted Inwangjesaekdo in 1751 in his seventies, of the granite peak Inwangsan above his own Seoul neighbourhood. The digital source is the Wikimedia plate at 4449×2583 (~82 px/cm; aspect 1.722 vs the 1.745 sheet, so undistorted), public domain (artist d. 1759). All numbers below measure this file at a stated working resolution.
The study's opening call inverts the Gauguin default, where the signature was benign and set aside. In the samjeol (three-perfections) tradition of Joseon literati culture, poetry, calligraphy and painting are one act, not three crafts; the yobaek (여백, unfilled paper) is an active compositional element; and the black title-block and red seals in the upper-right void are held to counter-balance the granite. Read at the plate, the upper-right block is Jeong Seon's own and contemporaneous: the title 仁王霽色, his art name Gyeomjae, the date 辛未閏月下浣 (late in the leap month of 1751), and two seals. Decision: include the inscription and seals as composition; treat any later collector seals accreted over 270 years as a separate stratum (the Malevich lesson). One honest flag on that decision: inscription authorship on Jeong Seon's works is genuinely contested, on the companion National Treasure Geumgang jeondo the 60-character inscription has been argued (Gang Gwan-sik) to be a later custodian's hand, not Jeong Seon's, on stroke-signature and iconographic grounds. So folding the 1751 block into "the single act" is an assumption we make and flag, not a fact we assume, and it is exactly why the H4 stroke question matters.
| condition | observation | decision |
|---|---|---|
| surface | clean cream paper, bold high-contrast ink, light foxing and a fine fibre texture | the opposite of aged silk: no craquelure to fake mass (I-11 minor). PM check once, do not assume |
| near-monochrome | ink on paper; the only true chroma is the two red seals | set β / cbi aside as substrate (I-12/I-25); read tonal; make the seals a targeted colour test (H2) |
| border | masses touch top & left edges; sides & bottom are blank yobaek margins | --edge-mode raw (I-1); confirm at the first-run gate |
| the inscription | a dense dark mark-cluster + red seals on bright paper, upper-right | will fire as edges/accents/chroma; per the inclusion decision that is correct, measure its role (H1,H2,H4) |
The standard-metric predictions are in labbook.md §pre-registration (mass
cohesive and upper-left, tau high, tonally bimodal, construction tonal). The spine of the study is four
hypotheses drawn straight from the samjeol reading, each pre-registered as confirmatory and each a
real bet the instrument could lose:
| # | hypothesis (from the tradition) | the test |
|---|---|---|
| H1 | the inscription cluster is a mass counterweight to the granite (an operationalization of the common gloss that it balances the cliffs 'on the left') | read with vs without it; ablating it should INCREASE imbalance (balance_dependency > 0) |
| H2 | the red seals are the entire colour composition | chroma should be near-degenerate everywhere except the seals |
| H3 | the yobaek is an active void, a bright counter-mass, not background | soft-read / void machinery: one large located void, not diffuse |
| H4 | one hand, one gi: calligraphy and landscape strokes share a signature | stroke-stats on the Soejima axes; inscription closer to the landscape than to an unrelated hand |
H1 and H2 are the founding pair, and they do not land in the same place. H1 is refuted as operationalized, and that is partly a finding about the operationalization (there is no left mass to oppose). H2 is not refuted but untestable at this plate's fidelity. The fairer form of the balance claim (negative-space activation) is open, its decisive test un-run. H3 and H4 remain un-run. A reading fitted to numbers has no failed bets; this one carries three distinct epistemic states, kept distinct.
Raw read, no persistence, 2600px working copy (runs/run_read.py, core untouched).
Headline numbers: kernel delta (+0.074, +0.035) near-centre;
mu 0.193; mask_mode TEXTURE_FIELD;
tau 0.379; shadow/mid/high 0.18 / 0.48 / 0.34;
Notan field_void 0.090; omega 0.032 (all families agree on placement);
construction tonal, grayscale-survival 0.994; chroma_support p98 0.018
(near-mono, not degenerate); void_fraction 0.72.
The value tools recover what the eye sees, and then correct it. A horizontal-band analysis of the dark-ink mass (full-res proxy of the Run-1 Notan-field result) shows the structure is not one mass but two, split by the mist:
notan_check.png shows the value squint holding both as
coherent masses where the island panel fragmented them.The founding hypothesis, tested by reading the picture with and without the upper-right cluster (masked to paper, box x 0.85–0.995 × y 0.03–0.33) and ablating it for balance_dependency. The result is the opposite of the prediction on every channel:
| channel | with inscription | without | reading |
|---|---|---|---|
| field imbalance (instrument) | 0.070 | 0.051 | removing it makes the picture MORE balanced: it is on the heavy side |
| dark-ink imbalance (fill-invariant) | 0.039 | 0.034 | same direction on the value channel; the cluster is 1.7% of the dark mass |
| field mass-centroid & saliency | — | both shift LEFT ~0.02–0.03 | the cluster is barely salient; it pulls attention slightly right, not opposing a left mass |
In a near-monochrome sheet the two red seals are the only chroma, so the chromatic-accent channel and the absolute-chroma guard become a targeted colour measurement of the seals alone. Measured:
With mass-counterweight refuted, the fairer form of the tradition's claim is negative-space activation: not that the inscription counterweights a mass, but that it activates the empty space, a different and testable thing. A first pass reported it CONFIRMED, calling the inscription "the single most isolated mark in the painting" against six hand-picked landscape marks. Told to go adversarial and not to Goldilock, the pass re-ran the test as a systematic map, and the headline did not survive.
labbook.md (house rule 3: corrections appended, not erased).What survives is narrower and honest. First a definition, load-bearing across this entry and Entry 07: true yobaek is paper left near its blank white (within ~3% of paper-white), while pale wash is committed midtone ink, such as the mist. (Run 6 later showed this luminance line is itself confounded by the sheet's uneven ageing, so the pocket specifics below are held loosely; the H3 verdict rests on the isolation and rival-void results, not on the blank-paper map.) Two fairer measures, and a real counterexample:
| measure | value | reading |
|---|---|---|
| void-territory (Voronoi of empty paper) | 1.5% of ink → nearest to 7.4% of empty paper | ~5× its ink share, but partly automatic for anything parked in a corner |
| true blank-paper, inscription corner | 0.36 (sole dark ink there) | genuine yobaek + a dark accent, a real regime |
| true blank-paper, upper-left corner | 0.38, bare | equally blank but bare: emptiness alone doesn't force a mark, so placement reads as a choice the instrument can't confirm |
| a false counterexample, ruled out | lower-left 0.87 'empty' | homogeneous pale wash (18% true-white), not yobaek; discarded |
The one leg that would settle the structural half of the balance claim: does the blank paper form a
single located void body with the inscription set within it. Pre-registered before measuring (decision rule:
supported only if the void is located AND the inscription is interior to it AND no comparable rival void
exists). Measured on the true blank paper (near paper-white), connected components, no erosion, across a
threshold sweep because the answer proves threshold-sensitive, and that sensitivity is the result
(runs/run4_yobaek_void.py).
| leg | prediction | result |
|---|---|---|
| H3b: is the fog blank paper or worked? | the tradition says blank (여백) | WORKED, by masking (not luminance): flat-fielding out the aging and pulling the mid-tones resolves the fog as coherent form (93% in connected bodies), wrapping the rocks. The earlier luminance 'proof' (fog darker than a 0.84 blank) is RETRACTED: circular baseline + uneven paper aging (Run 6). Working the mist is standard ink technique, not distinctive |
| H3c: inscription interior to the LARGEST blank body | yes | not robust: true in only 1 of 6 threshold settings (the strictest); flips to upper-left / lower-left when relaxed |
| H3d: is there a comparable rival void? | predict-against-self | yes, always: the inscription's pocket and the upper-left pocket are separate in 6/6 settings, the upper-left often larger |
| H3a: is the blank one located body? | one dominant void | sub-result (feeds H3): the largest pocket holds 24–40% of blank, so not diffuse specks, but no single body dominates and which is largest is threshold-dependent |
g2_void_topology:
128 tiny voids, largest only 0.3% of frame, no dominant body). So the negative-space form of the balance
claim, left OPEN after Run 3-V, now closes as not structurally supported. A methodological note kept
where it fell: a first pass used binary_opening and got a false "blank is shattered into 1093
specks" result; the erosion destroyed the thin, foxed blank. Corrected to a no-erosion sweep. Same family as
I-32, the null's construction, not the foil, manufactured the artifact.The samjeol unity claim's structural form: are the calligraphy strokes and the landscape
strokes one gesture system? "One hand" has a fatal confound (inscription and landscape are the same ink on
the same sheet through the same JPEG), so instead of a chosen foil this uses a within-work null on
the gesture axes that measure the mark, not the material (contour turning-angle variance, inflection
density, orientation entropy; runs/run5_one_hand.py): does the inscription's signature fall
inside the landscape patches' own spread, and, as a reference point only, where does a different hand land?
All four pre-registered hypotheses are now run, and the honesty of the study is in keeping the outcomes distinct across four different epistemic states. Refuted: the mass-counterweight operationalization (Entry 04, robust and fill-invariant), which mostly teaches that the composition is right-heavy with no left mass to oppose. Untestable at this fidelity: the seals-as-colour claim (Entry 05), owed a better plate, not disproven. Not supported: negative-space activation (Entries 06 and 07), where the inscription does sit in real blank paper but not in a located, singled-out void, and the decisive test came back negative. Undecidable: one hand, one gi (Entry 08), where the gesture measure is degenerate and fails its own control. So on no structural mechanism the pre-semantic instrument can reach is the samjeol claim confirmed: refuted, untestable, unsupported, undecidable. The tradition's reading is not thereby false, it lives in the semantic / Gestalt layer the instrument declines (C-1), but where it made a checkable structural claim the structure does not back it. The clean, finished result the eye and the brief both got wrong is the mass: not a single monumental peak but a bipolar dark–void–dark field, the foreground rivalling the dome, and it survives a flat-field correction for the paper's aging. On the void the instrument corrects itself: the fog is worked, not blank (masking, Run 6), though working the mist is standard ink technique, so what it can adjudicate is which sense of an "active void" holds up, the mist as the light value-separator between the two dark poles (survives), a located counter-mass (does not). The one sourced distinctiveness is the other pole: Jeong Seon's bold, heavy, wet dark against the Southern School's light, dry, delicate norm.
| prediction (Step 0) | outcome | verdict |
|---|---|---|
| mu high (one dominant connected mass) | mu 0.193, low; TEXTURE_FIELD (edge fragments on brushwork) | miss (I-31 finding) |
| mass cohesive, upper-left | near-centre, BIPOLAR (dome + foreground split by mist) | miss (correction) |
| tau high, tonally bimodal | tau 0.379 moderate; mid-key (misted washes fill the midtone) | miss (correction) |
| construction tonal (survives grayscale) | tonal, grayscale-survival 0.994 | hit (strong) |
| omega low (families agree on placement) | omega 0.032 | hit |
| H1 inscription is a mass counterweight (as operationalized) | refuted: it is on the heavy side, imbalance 0.070→0.051 | REFUTED, predicted-against (finding) |
| H2 seals are the whole colour composition | substrate out-chromas the faded seals on this plate | UNTESTABLE at fidelity (not disproven) |
| H1b inscription activates the void (fair form) | sole ink in a real blank pocket, but not the extremum and not a located singled-out void; intent is C-1 | NOT SUPPORTED (H3 run, Entry 07) |
| H3 the yobaek is one located active void | not a singled-out void; and the fog is worked, not blank (masking) — but the luminance test of that was retracted, aged-paper confound (Run 6) | miss; void reopened & corrected |
| H4 one hand, one gi (gesture unity) | gesture axes saturated; a different hand (Soejima) sits inside the landscape cloud | UNDECIDABLE (measure degenerate) |
The founding bets did not land in one place, and the scorecard keeps the states distinct: H1 refuted, H2 untestable at this fidelity, H3 run and unsupported, H4 undecidable. The misses carried the study: the bipolar-mass correction, the I-31 recurrence, and the reversal that turned a claimed confirmation into I-32. The void was reopened after the fact (Run 6): the fog is worked, not blank (masking), but the luminance argument for it was retracted as an aged-paper artifact, and working the mist proved standard, not distinctive. All four hypotheses are now run; none confirms the tradition on a structural mechanism.
Lab book complete on the structural legs, honest — 2026-07-30. Nine entries; all four pre-registered hypotheses run and kept in distinct states: H1 (mass counterweight) refuted, H2 (seals as colour) untestable at this plate's fidelity, H3 (negative-space activation) run and not supported (the yobaek is distributed comparable pockets, not one activated void), H4 (one hand, one gi) undecidable (the gesture measure is degenerate and fails its own control). None confirms the samjeol reading on a structural mechanism; it lives in the C-1 layer. Three overclaims were caught (two before shipping, one after): the "most isolated mark" reversal, promoted as I-32; an erosion artifact inside H3; and, reopened by RP's artist read, the "emptiness is painted" luminance claim, retracted as an aged-paper confound (Run 6). Done by masking, the fog IS worked, but that is standard technique; the sourced distinctiveness is the bold wet dark against the school's light-dry norm. The positive yield: the bipolar mass (aging-robust, convergence-confirmed), the sourced dark-handling distinctiveness, and I-32. Everything rides on one aged plate, flagged where it bites. Structural legs run; the critique, written before the void was settled, is to be rebuilt from this record.