Diego Velázquez — Las Meninasoil on canvas, 1656 · Museo del Prado P01174 · 320.3 × 279.1 cm · public domain

Entry 01 — Condition, and the source decision

Most of the noise was the reproduction, not the painting

Extreme dark key, measured on the first file grabbed (a Prado colour-corrected "FXD" derivative): median L 46, 72% of pixels below L70, 5% above L150. On a surface this dark the I-11 speck check is mandatory, and it fired at the worst tier: PM verdict "meaningful" (0.0178). The cause was the file, not the canvas. The FXD's shadow-lift and sharpening had raised the black floor (a hazy band the eye read as glare) and amplified craquelure. Re-deriving the working image from the raw Prado gigapixel (26065×30000, 10× box-average to 2600px) cut dark-void high-frequency noise from 6.86 to 2.92, deepened the void from L45 to L22, and dropped the verdict to "light" (0.0096).

Entry 02 — Four named edge artifacts

A corpus first: the repoussoir object

Las Meninas brought a condition the corpus hadn't met: a large foreground repoussoir object — the canvas-back Velázquez stands behind (x ≈ 0.02–0.16) — whose hard right edge is the strongest edge in the picture. Measured artifacts, each with handling: the canvas-back edge + left dark margin (x < 0.17); two physical canvas seams at x ≈ 0.259 and 0.636 (3.9× and 3.6× the dark-void median vertical-edge energy — the canvas is stitched from three widths); the lit doorjamb at the right edge (x > 0.975); a floor contour-trap along the bottom (y > 0.93). Handling: masked for the accent read (and any future armature read); thin and minor for mass, tone and saliency. Seam 0.259 passes through the Velázquez and Sarmiento probe boxes (noted where those are quoted); the mirror and doorway boxes are clear of both seams — checked, since the mirror's local substance carries the headline.

Entry 03 — The accent read, cleaned

88% of the default channel was one object's edge

The default structural accent channel returned 24 accents of which 21 were artifact — a column of fragments down the canvas-back's edge plus the doorjamb (the Degas I-7 edge-bias in its most extreme observed form). The chromatic channel (σ6) was 33% artifact; after removing the four named zones, 12 real accents remain and they land on the figure frieze: Velázquez's face, the búcaro-menina, the Infanta, Maribárbola, Pertusato, the chaperones, the dog. Top leverage 0.83 falling smoothly to 0.49 — distributed punctuation, no climax.

Is the Santiago cross one of the twelve? Checked (review 6). The red cross of Santiago on Velázquez's breast is post-1659 paint — added ~3 years after the picture, after the knighthood (per Laing not by the King, which kills the legend). If it were one of the twelve accents, a punctuation mark of the study would not belong to the 1656 composition. Rendered and checked (scout_velazquez_cross.png): the two accents inside his figure box sit on his eye and the red paint on his palette — both original; the cross itself carries no accent. The read is clean of the later addition.

Entry 04 — Headline: the buried sovereigns

Full local substance, one two-hundredth of the frame

subject_probes.py (I-4 mass-locked reads on the PM-denoised fields; nine verified boxes). The crux — the mirror against the Infanta and the doorway:

The marginalization is by extent, not intensity: the reflection is painted with real substance (mass-mean 0.54, edge 0.36 — not a smudge) and allotted 0.61% of the frame. Same signature as the Icarus, and the comparison now carries its numbers (review: it had been asserted without them):

Entry 05 — The two apertures split

A chamberlain in the bright hole, the sovereigns in the dim one

The dark back wall holds two lit rectangles. They are opposites: the open doorway (José Nieto on the lit stair) is the densest local saliency peak in the back wall — local mean 0.556 — while the mirror is a saliency void at 0.055. The doorway out-pulls the royal mirror 10.1× on saliency, 1.53× on mass [confirmatory, prediction 8; v3 boxes].

And the pull is contrast, not brightness (review 4's sharpening). The Ortega probe puts the two lit planes within 18% of each other in luminance (foreground frieze 0.232, through-door light 0.274) — yet the doorway is the denser local saliency peak by a wide margin. Its optical pull is a local-contrast effect against the dark back wall, not a brightness effect. Which sharpens the study's headline convergence: Snyder & Cohen's vanishing point coincides not with the brightest thing in the picture but with a local-contrast peak — a more specific agreement between the perspectival and optical organizations than brightness alone would give [exploratory sharpening, from measured values].

Entry 06 — State, mass, tone

The shadow pole is a tie, not a summit

State mixed_field / offcenter_mass (not dissolution); the island anchor is the Infanta, robust at object identity across both edge modes (not at location: her body edge-aware, her skirt-base raw); entropy 0.79–0.81. Mass pulls hard down (δy +0.214; bottom third 50% of mass vs top 18%) and sits dead-centre horizontally (δx +0.039 — the pre-registered "slightly left" was wrong; scored half-right). μ 0.561: the bright figure-frieze is one connected structure. Colour is a servant of the drawing: β 0.049, coupling +0.361 (corpus high). DGI 18.8, "authored against gravity" — the most extreme off-basin composition measured.

Tone: shadow_mass 0.897 · τ 0.750 · highlight 0.001 — the Infanta's dress, the brightest object in the picture, holds one part in a thousand of the tonal mass. "Darkest in the corpus" is REFUTED by the same-pipeline compare (Entry 11): the Caravaggio reads 0.903 against the Meninas's 0.896 — a 0.007 gap, a statistical tie at the corpus's shadow pole (and the first-pass "Caravaggio ~0.35" was a misquote: that is the Bruegel's number). Against the measured ~0.13 chain delta, the ordering between the two Spanish tenebrists is undecidable at current chain control.

Entry 07 — R_spatial (C-3)

The cancellation camp

0.15–0.26, subadditive, cut-stable — every half and quadrant carries more torque than the whole; the frieze's local leans cancel into a still composition. Joins Caravaggio, Pollock, Bruegel and the HCB in the cancellation camp [confirmatory, prediction 7 — weak band: 0.1–0.6 spans most of the subadditive range]. Appended to the register.

Entry 08 — Address (C-4): beats the spectrum, fails the matched null

The magnitude miss was the first finding; the withdrawn verdict is the deeper one

Central frontality 0.425, axis locked dead-centre (offset 0.004, lock 0.991) — but mid-pack: below the Klimt (0.663) and the Bruegel (0.537). The pre-registered thesis that Las Meninas might be the corpus's figural-address maximum is refuted [confirmatory-refuted, prediction 6; the 0.4–0.65 band was also weak — it covers two-thirds of the corpus spread].

Figural vs spectral. Address is scale-robust here (0.425 @2600 vs 0.431 @1300, a 1.4 percent shift — the stated justification for sweeping at 1300px). The informative null is the phase-scramble (it preserves the power spectrum, so it tests whether the room's banding could produce the signal; patch-shuffle destroys the spectrum and only asks "any structure at all"). At n=100, 0 of 100 draws reached the real value — honest interim p < 0.01 with the point estimate tail-model-dependent (a normal tail gives ≈0.003, but sd/mean = 0.97 is the exponential signature and an exponential tail gives ≈0.027, a 10× spread the n=100 summary cannot discriminate; "p = 0" and "p ≈ 0.003" are both retired — the first as a resolution floor, the second as an extrapolation). Settled empirically at n=1000 (phase-only, quantiles): null median 0.109, q99 0.398, q99.9 0.430, max 0.481 — 1 of 1000 draws ≥ the real 0.431: p ≈ 0.002 by the conventional (r+1)/(n+1) estimator (r/n = 0.001 is the biased form — the p-=-0 logic at its own floor; review 3), 95 percent CI on the exceedance ≤ ~0.006. Note the real value is no longer above the observed null max (0.481 exceeds it): the verdict rests on the empirical percentile, never on clearing a maximum. FIGURAL stands — under a decision rule recorded honestly as adopted AFTER this sweep landed (binding forward, no retroactive authority): p < 0.01 verdict as written; 0.01–0.05 marginal, no "settled"; ≥ 0.05 withdrawn.

What the percentile decomposes into (shuffle_variance.py). Of the address variance the shuffle produces, only 36% (η²) is explained by which figure is central; 64% is the other six positions. Within that 36% the Infanta is genuinely the best central figure — mean 0.494 when she lands central, vs 0.270–0.365 for every other figure. And the conditional read is the sharpest number in this entry: given the Infanta central — the real configuration — random placements of the other six average 0.494 (n=40 of the N=300 shuffles land her central, ≈ the 300/7 = 43 expected; sd 0.054), while Velázquez's actual placement scores 0.419: −1.38σ within its own group (run B), the 7.5th percentile. Only 3 of 40 random arrangements sharing his central figure score lower. So the +0.7σ decomposes exactly: all of it comes from the Infanta being central (the strongest bilateral body, in the middle), and the other six placements give roughly half of it back — they score below chance, not merely "not maximizing." This is the address-channel version of Entry 03's accent finding (distributed punctuation, 0.83→0.49, no climax), arriving independently through a different lens: two channels, same read — the frieze is arranged to dissipate central address, not concentrate it [exploratory: post-hoc, n=40 conditional group; the Entry-03 convergence is interpretation, the −1.38σ is measured].

The two nulls bracket; neither answers the deepest question. The matched null holds central occupancy fixed — someone is always central — so it tests which figure is central, not whether central occupancy is itself remarkable. Read together the nulls are bounds: phase-scramble too permissive (destroys the figures), figure-shuffle too conservative in exactly this way (it can never vacate the centre). Between them the residue — central bilateral figural mass exists and is real content — survives because neither null can test it. Robustness-checked (null_figure_shuffle_robust.py): not a seed or fill artifact — across 3 seeds × 3 background fills (per-row median, heavy-blur bleed, flat dark-void) the address p stays 0.21–0.40, even where the dark-fill drops the synth-real reference to 0.384.

Entry 09 — Null battery, and its own corrections

Two numbers that looked like one, reconciled by measurement

The review caught the study quoting mass_divergence as 0.080 in one row and 0.0746 in another; a second review caught the first fix mis-attributing the gap, so the missing cell of the 2×2 was run (review_checks.json):

Scale effect −14.9% within the PM pipeline (−24.7% within no-PM — the first fix quoted the no-PM column as if it explained the spine-to-null gap). The null run is the no-PM @1300 cell: it differs from the spine in two ways whose effects partially offset (net −6.9%), and its label now says so. The conclusions survive at the honest magnitudes: mass_divergence is scale-sensitive (−15% under the spine's own pipeline) where address is not (1.4%); the null test remains internally valid (real and nulls share one pipeline and scale: no-PM @1300); the two numbers are never interchangeable, and the study quotes 0.080 (@2600, PM) as the finding. That null test: real 0.0746 vs phase-scramble 0.0086 and patch-shuffle 0.029, no draw reaching the real value (p < 0.01 at n=100) — the eye-vs-weight displacement is compositional. DGI's null was then run (Entry 11), on exactly the review's rationale that extremes are where metrics break — and its own 2×2 was completed to bridge the 18.0-vs-18.8 pair the first draft left unexplained: DGI reads 28.2 (no-PM) / 18.8 (PM, the spine) @2600, and 18.0 (no-PM, the null run's cell) / 14.9 (PM) @1300 — the null-run value matches the spine only by two offsetting pipeline effects, the same accident as mass_divergence, and the two are never interchangeable.

Entry 10 — The literature, crossed with the measurements

Convergences claimed — and condition-split; two questions run [sources Grade C unless pulled; see meninas_sources.md]

A deep source pass (meninas_sources.md — graded A–D per claim, most locators Grade C and unverified; the transmission layer itself audited and found lossy) lands on the study in both directions.

The headline convergence. The established geometric result contra Foucault (Snyder & Cohen 1980; Steinberg 1981) is that the vanishing point is not central but at the doorway, near Nieto's elbow. Entry 05 found the doorway is the densest local saliency peak in the back wall (0.556; it out-pulls the mirror 10.35× on clean jitter draws, Entry 12 — the single global argmax is the Infanta's dress, so the convergence rests on the doorway being a dominant local peak coincident with the vanishing point, not the global maximum) — a completely independent channel, luminance-contrast attention rather than ruler-work on orthogonals, landing on the same object. The picture's optical organization agrees with its perspectival organization, and the mirror-centred reading loses on a second, unrelated axis. Three more: Steinberg's shifting focal centre and López-Rey's three foci ↔ the distributed punctuation (0.83→0.49, no climax) and moderate-magnitude address; Clark 1960's tone thesis ("the tonal relations are true... and it holds") ↔ β 0.049 / coupling +0.361; White 1969's lower-half observation ↔ δy +0.214. And Garrido 1992's reflectography (no underdrawing; painted direct) is the one technical datum behind DGI 18.8: there was no plan — he found it in paint.

The convergences, condition-split (review 4 — the study's own WTDNP applied to its allies). Clark wrote in 1960 and Ortega in 1953; Brealey's 1984 cleaning removed a yellow veil accumulating since the 19th century — both men described a picture wearing ~150 years of dirt, and this study measures the post-cleaning surface. So the split is: Snyder & Cohen and Steinberg are varnish-invariant (perspective geometry survives any cleaning — the headline convergence is safe from the condition record entirely, a point in its favour); Clark and Ortega are cross-condition-state comparisons — both probably survive (a roughly uniform veil scales tone rather than restructuring it, and lit-dark-lit is robust to scaling), but "probably survives" is now said rather than assumed. And Clark carries a reflexive sting: his thesis is that the picture holds though the colours are drab — he saw drab partly through a yellow veil, this study measures drab partly through faded azurite. Two condition states, agreeing. Might be the painting; might be two states meeting in the middle [condition-split: exploratory].

Two of the literature's decidable questions, run today — with their scope stated (review 4). Miller 1998 claimed the mirror betrays itself as a mirror by borrowed light. Measured (miller_probe.json): the mirror interior reads 2.3–2.7× the two Mazo canvases on the same wall and 2.0× the bare wall, while still 0.66× the doorway. This confirms Miller's premise (the mirror is anomalously bright for that wall) and leaves his inference untouched: the Mazo ratio is content-confounded (their subjects are dark — any painted rectangle depicting a bright scene would beat them), and 2.0× bare wall is a bar ordinary paint clears easily (the Infanta clears it by far more). Miller's actual argument — that the brightness can only be borrowed, a consistency claim about the room's light — needs an illumination model, not a ratio. What is genuinely this study's: locally anomalous AND globally weak, both measured [premise confirmed; inference not tested]. Attribution caveat (review 6): the two back-wall canvases are the Miller-probe controls, so their identity is load-bearing — and it is contested. Kahr reads them as Rubens-derived Minerva punishing Arachne + Apollo and Marsyas; the National Trust catalogue (Laing) has Minerva and Arachne + The Judgement of Midas after Jordaens (Mazo's copies in the Prado, inv. 1551/1712). The probe needs only two dark painted rectangles on that wall, which both readings grant — but "Mazo canvases" is one identification of two, flagged. Ortega y Gasset 1953 described a three-part light structure: lit foreground, darkened intermediate zone of silhouettes, lit background. Measured (ortega_probe.json): foreground frieze 0.232 / intermediate 0.074 / through-door light 0.274 — lit-dark-lit holds, the middle at a third of either lit plane. Prediction 9's hypothesis, pre-registered 73 years early, confirmed. The mirror-vs-door attention conflict is scoped, not attacked — the received claim could not be attributed to a named scholar and no eye-tracking study of this painting exists, so the doorway-over-mirror result (10.35× on clean jitter draws) stands as the bottom-up model against an unattributed sentence, with the deciding experiment named.

What the instrument cannot touch, said plainly: Foucault's absent centre is structural, not photometric — 0.61% of frame neither confirms nor refutes it; the mirror-reflects-canvas-or-couple debate is geometry and optics outside these metrics; gaze stays beyond the C-4 ceiling (the literature cannot even agree whether three or five figures look out), and the column sweep's lobes-on-brightness result is evidence that address ≠ gaze — a methodological finding about C-4 itself. Alpers 1983 (meaning in the representation, not the plot) is the nearest thing the field has offered to this instrument's manifesto; the study lands where she pointed.

Entry 11 — The owed items, run

Persistence, the DGI null, Ortega, and the compare

Persistence (full read, PM pipeline): 0.558, fragile_to = low_contrast — compress the tonal range and the structure dies [stable]. That completes a three-way corpus split: Bruegel fragile to blur (fine edges), Klimt to grayscale (colour), Meninas to contrast (tone) — a perfect 1:1 across three works and three perturbations, which is 1-in-6 by chance and a post-hoc pattern [exploratory].

The DGI null (n=100 per family, fields-only, no-PM @1300 for both real and nulls — the same cell throughout; the spine's 18.8 is PM @2600, bridged by the 2×2 in Entry 09): real 18.0 vs phase-scramble [30.2–62.7, median 46] and patch-shuffle [47.4–65.8, median 58]. Direction, stated: lower DGI = further from the centred default-gravity basin, so the real value sitting below every null draw is the confirming side — each re-arrangement scores MORE default-gravity than the painting [stable within its cell]. And a scope the cleanliness demands (review 4): a null with zero overlap in 200 draws may be answering "does this image have any composition?" rather than "is this composition non-default?" — phase-scrambling destroys all arrangement, so any structured image may beat any scramble. The address null was informative precisely because it was close; this one may be free. The matched test is the figure-position shuffle (next).

Compare vs the poles (n=5: + Caravaggio, Bruegel, Klimt, Degas), read as raw values and ranks — z's saturate at this n and carry no magnitude (review 4): the decisive row is tonal — Caravaggio 0.903 vs Meninas 0.896 shadow_mass, a tie at the corpus's shadow pole that refutes prediction 5's superlative (Entry 06). Elsewhere: coupling_index 0.301, rank 1/5 (Caravaggio 0.195, Bruegel 0.192, the moderns ≈0 — colour-describes-form beyond even the other tenebrist); β 0.060, lowest with the Caravaggio's 0.061 (the moderns 0.23–0.38): the two Spanish works pair on colour-tone behaviour across the board; Klimt holds the highlight pole (0.372 vs everyone else ≤0.076), Degas the torque pole. Placement only, n=5 [placement]. The mass_divergence null's cheap-cell caveat stands: it proves non-nullness of the no-PM @1300 cell, transferred to the claim cell across a measured 6.9% gap with an 8.7× margin.

Entry 12 — Jitter: box placement retired as a degree of freedom

200 perturbed box sets; the direction never flips

The review-5 worry, stated fairly: four placement errors were found across v1→v3 and every correction strengthened the headline — unanimity that either means the errors were all conservative, or the corrector was steering. Hand-placement cannot settle that; a jitter analysis can (jitter_probes.py): all nine boxes perturbed independently in position and extent, uniform ±5% of each box's own dimensions (covering the requested ±3–5% band), n=200 draws, the exact I-4 selector rerun per draw on once-computed PM fields. And — the load-bearing addition after review 5 — each draw is tagged clean or failed: the selector reports whether the component it chose is the box's largest qualifying one, so the distribution can be split by whether the I-4 lock held. You can eyeball nine masks; you cannot eyeball eighteen hundred.

Of 200 valid draws, 155 are clean. This run fixes a harness bug AND adopts the largest-component convention for the doorway; the convention × jitter 2×2 (finding 4) shows the bug was worth ~2% and the convention ~167%, so the doorway value here is a convention result, not a bug fix.

Four findings.

(1) The direction never flips — in any of 200 draws, on any crux channel. The Infanta carries at least 13.6× the mirror's mass-share and at least 43× its saliency-share under any box set within ±5% of the pinned one; the clean-draw floors hold at 13.6× / 43.2×. And the doorway out-densities the mirror in every single draw (P = 1.0), floor 5.11× all-draws, 9.53× clean. Box placement is retired as a threat to the finding.

(2) The fat 'all-draws' mass-share tail is a mirror component-flip, not placement — review 5's mechanism, confirmed. ±5% on a box moves an area-based share ~10%, so the ratio should move 15.2 → ~16.8; a q95 of 95.7× requires a box to lose ~84% of its content, which geometry at ±5% cannot do. What can is the selector taking a small off-object blob. Split by lock status the tail collapses: mass-share q95 falls 95.7× → 20.6×. Because a mirror flip makes the mirror catch less, these inflate the ratio and live in the upper tail — so the floor (13.6×) is clean and the headline is safe.

(3) The real diagnostic is the decision margin, not a jitter failure-rate (lock_margins.py). A jitter flip-rate is a noisy proxy; the exposure is a static property of each box — how close the runner-up component sits to the winner. Measured for all nine: doorway margin 0.0015 (a one-pixel shift decides the read), mirror 0.105, Infanta a single qualifying component — no decision exists at all, the rest 0.067–0.33. Two things follow that a flip-rate hid: the Infanta headline never had an exposure (nothing to flip), and a mirror flip would read the runner-up at 0.107 saliency vs the winner's 0.055 — it would weaken the study's ratio, not flatter it. The only knife-edge pointing the study's way is the doorway, and it is handled (finding 4).

(4) The doorway: a harness bug AND a convention change, conflated in one pass — the 2×2 prices them apart (convention_jitter.py). Two things moved between last round's 3.75× and this round's 10.2×, and the honest accounting separates them. (a) A harness bug: the jitter loop computed the box window with int(round(x·w)) while zone_read truncates — a one-pixel mismatch, decisive at the doorway's 0.0015 margin. (b) A convention change: the doorway's selection rule was switched from nearest-to-centre to largest-component, on the annular-target ground (the lit aperture surrounds the box centre — Nieto's dark body sits in the middle — so nearest aims into the hole where a 0.6%-of-box highlight sits; verified in scout_doorway_components.png). Priced by the 2×2: fixed-harness + nearest = 3.83× (bug worth ~2%); fixed-harness + largest = 10.23× (convention worth ~167%). So — correcting last round, which credited the bug — the 3.75× was not a poisoned statistic; it was the correct doorway number under nearest-to-centre. The 10.2× is a convention result on an annular target, justified on structural grounds stated before the number, its worth measured. This is the study's recurring error class (PM × scale; chain × pipeline; Caravaggio × Bruegel: two things move, effect credited to the wrong one) — caught here on myself. The convention removes the coin-flip (0 flips by construction under largest); the value is stable under jitter (10.21× all-draws, 10.35× clean), and every fat-tail mass-share outlier is a mirror lock-flip (15.4× locked, 92.7× flipped), never Infanta or doorway.

The maximum claim, settled box-free. Entry 05 had called the doorway "the saliency maximum, the strongest pull of any object measured" — silently switching between a box-free per-pixel argmax and a ranking over nine hand-boxes. Measured with no box: the global saliency argmax is at (0.537, 0.772), inside the Infanta's dress — not the doorway. The picture's two attention poles are the living Infanta (the argmax) and the doorway (the largest single concentration of top-decile saliency mass, ~54%, though its box overlaps the Infanta's by ~10%); the royals are neither. "The saliency maximum is at the doorway" is withdrawn; the doorway is the densest local peak, which is what the Snyder & Cohen convergence needs.

Asymmetric robustness, measured — and it splits by error scale. The review's gift (mirror saliency-share invariant across the v1→v3 re-pin, 0.336‰ → 0.336‰, ratio 0.998, while the Infanta's shares doubled) is real. Fine jitter shows the large lit figures barely move (frame-share cv ≈ 0.03: Infanta, Sarmiento, Isabel). Both are true: the mirror is invariant under gross error (a body-width miss) because nothing bright sits near it to catch by mistake — the buried thing is stably buried — while the small apertures are the ones the selector abandons under fine jitter (finding 3). The crux survives at both scales (gross: the re-pin moved the ratio 10.4→15.2×, same direction; fine, clean: floor 13.8×).

The uniform-rule question, closed. Isabel's mask-lock fix had changed a rule for one object; the review asked whether it was applied to the other eight. All nine locks were then verified, not asserted (mask_verification.json; masks rendered in scout_all_masks.png): 8/9 chose their box's dominant mass component (7/9 natively — Isabel passes only because her box was treated). The one remaining failure is Velázquez (chosen component 9.1% of box vs largest 18.0%). Applying Isabel's treatment uniformly — re-centring his box on the dominant component's centroid — made him worse, not better (saliency mean 0.137 → 0.038): his box's dominant mass is wall structure, not the painter (dark-on-dark, I-16, with the x≈0.259 seam through the box). So the rule is stated exactly: mass-centroid pinning applies where the dominant component is the object (Isabel, zoom-verified); where no placement yields an object-locked dominant component, the object is flagged unmeasurable-reliable and both readings reported (Velázquez: v3 box 0.387 mass-mean / 0.47% share / 0.137 sal-mean vs centroid-pinned 0.297 / 0.68% / 0.038). He appears in no crux ratio; his rank rows carry the flag [confirmatory hardening, review 5].

Two findings, cleanly split. (a) The headline is convention-invariant on AND off the pin: Infanta ÷ mirror floors 13.56/13.59/13.56× across all three conventions under full jitter (saliency-share floor dips to 38.5× under centroid — still ≥38×). No hand-made selection choice moves the burial result, and that claim now rests on the jittered cells, not the static identity. (b) The doorway is the one convention-sensitive number, and only off the pin: under all-nearest it coin-flips exactly as the 0.0015 margin predicted (52% flip, median 3.83×); under largest it is tight (10.23×, floor 9.53×). The honest statement of the doorway choice: the annular-target argument justifies largest-component on structural grounds; the convention × jitter cells measure what the choice is worth (3.83× → 10.23× median); the static sweep never could and is no longer cited for it. Convention is retired for the headline by measurement, and pinned open-eyed for the doorway by stated principle plus its measured effect.

Pre-registration scorecard

Ten predictions, locked before the first number

What This Does Not Prove

• Not intention. The allocation of frame and light is measured; its purpose is a human reading.
• Cross-study tonal rankings inherit reproduction chains AND the object's condition chain. The reproduction delta is ~0.13 on this painting (raw vs FXD, both pipelines); the condition record (1734 fire, undocumented trim, repainted cheek, stuccoed seams, faded azurite, the 1984 cleaning) is a larger, unmeasured confound — "darkest in corpus" was refuted outright by the same-pipeline compare (Caravaggio 0.903 vs 0.896 — a tie at the pole, undecidable against the 0.13 chain delta); any future tonal ranking remains partly a claim about conservation history.
• Every horizontal-centre claim is provisional on the undocumented trim (prediction 2's ✗; C-4's 0.004 axis lock is the centre of a cut canvas). Two lines of trim evidence, both right-heavy, neither significance-grade: the seam-panel widths (0.259/0.377/0.364, left ~30% narrower) are an inference conditional on an untested assumption (uniform loom widths — Velázquez may have used a narrower left strip), carrying no p-value; Clark's grid corroborates the direction only (fit-significance withdrawn under a threshold sweep, p 0.02–0.91, but the right-heavy direction knob-robust; Entry 11). The study claims a right-heavy trim direction, not a measurement of its size.
• The source list has its own provenance grades (meninas_sources.md): most literature claims used here are Grade C (locators transmitted, unverified); the four priority pulls (Garrido 1992, López-Rey II:306, Snyder & Cohen 1980, Clark 1960) are named before any of those claims becomes load-bearing.
• The fields-only metrics are not scale-robust — an instrument fact, not a study fact (review 5). Across one halving (2600→1300, within-PM): DGI −20.7% (18.8→14.9), mass_divergence −14.9%, address 1.4%. Two of the three nulled metrics drift 15–21% per octave; any future 1300px null sweep owes its 2×2 (scale × PM) up front, not as a retrofit. mass_divergence's null remains a same-pipeline, same-scale test (no-PM @1300) and its value is never the spine value (Entry 09).
• The per-box selector exposure is the I-4 decision margin, not a jitter flip-rate (Entry 12; lock_margins.py): the Infanta headline has no exposure (single component), a mirror flip would weaken the ratio not flatter it (runner-up 0.107 vs winner 0.055), and the doorway's knife-edge (margin 0.0015, an annular target) is handled by a call-site largest-component rule that changes no v3 number. A one-pixel harness bug was fixed, but the convention × jitter 2×2 prices it at only ~2%: the doorway's 3.75×→10.2× lift is a convention change (nearest→largest, annular-target ground), not a bug fix (finding 4). Clean-draw floors 13.6× / 43.2× / 9.53×. Jitter covers fine placement error, not gross mispinning (that is the zoom-verification protocol's job; the v1→v3 episode is what skipping it looks like).
• The address channel is a measurement, not an authored-address verdict. It beats the phase null (p ≈ 0.002 — real content, not spectrum). Against the matched figure-position shuffle it is a 70.4th-percentile arrangement of its own cast, CI [66.4, 74.4] (pooled 148/500 over two independent seeds), of which who-is-central explains only 36% (η²) — and conditional on the Infanta central, the other six placements sit at −1.32σ below random (pooled over two seeds, exploratory): the frieze dissipates central address (Entry 08). The figural verdict does not stand; the finding is the percentile. The two nulls bracket (phase too permissive, figure-shuffle too conservative — it can never vacate the centre), so central occupancy per se is untestable by either. DGI is NOT tested on this null (37% synthesis artifact swamps the 11% effect — a false-transfer refused, Entry 11), not withdrawn. mass_divergence is the one placement metric that survives its null. The "distributed gaze-field" is exploratory prose, refuted by the sweep; gaze itself is beyond the instrument.
• The mirror box includes some frame — a limitation on the 0.61% headline share. The v1→v3 box audit (Entry 04) is the larger lesson: the published overlay was a second hand-copy of the coordinates and hid four mis-aimed boxes; overlays are now generated from the probe's own list, and the Study-7 metrics-layer pattern is owed at critique time. The build now trips if the markdown twin lacks the current pass's markers.
• The prediction-5 prior was calibrated against a corrupted anchor ("Caravaggio ~0.35" was Bruegel's row): one bad input cell produced both a wrong band and a wrong superlative. Pattern-shelf: validate anchor values against the source study's own records before chain-scoping them — the reviews audited the pipeline for two rounds while the bug sat in the data.
• Velázquez is flagged unmeasurable-reliable: dark-on-dark (I-16) with canvas seam 0.259 through his box (Sarmiento's box also crosses a seam), and his I-4 lock is ambiguous under any placement — both readings reported in Entry 12. The mirror and doorway boxes are clear of both seams (checked); all nine locks verified, 8/9 clean.
• Island counts are segmenter ceilings (I-17) — the read rests on anchor identity (robust across edge modes at identity, not location), entropy, and mass-shift.
• Accent claims are lens-capped and cleaned of four named artifacts; the mirror null was cap-checked at 100 (18 found, zero on the mirror).
• R_spatial is a candidate (C-3); predictions 6 and 7 were weak-band confirmations, annotated as such in the scorecard.
• Nothing now owed before the critique. Run across six review rounds: the compare vs the poles, DGI null, Ortega/depth/Miller probes, jitter + decision-margin + all-nine masks + the selector-convention sweep (crux invariant, convention retired, Entry 12), the figure-position shuffle with seed/fill robustness and the η² variance decomposition (address measured at +0.7σ / 74th percentile; DGI not tested, Entries 08/11), and Clark's grid (fit-significance withdrawn under a threshold sweep, direction retained, Entry 11). Only the critique remains, after the user's read (Step 0).
• Single reproduction chain per study; crisp decimals carry an unmeasured variance bar.

Sources & record

• meninas_sources.md — the study's graded source list (A–D per claim; the transmission layer audited and found lossy — wrong DOIs, a corrupted title, a phantom page number, and the dimension error this header first inherited). Priority pulls: Garrido 1992 pp. 579–591; López-Rey 1999 II:306; Snyder & Cohen 1980; Clark 1960 pp. 32–40.
• Museo del Prado, P01174 (Las Meninas, Diego Velázquez, 1656; 320.3 × 279.1 cm): object record; source image the Prado/Google-Earth gigapixel via Wikimedia Commons (raw scan pinned; the colour-corrected FXD derivative measured and rejected in Entry 01).
• Snyder & Cohen 1980; Steinberg 1981; Alpers 1983; Clark 1960; Ortega y Gasset 1953; Miller 1998; López-Rey 1996/1999; Garrido 1992; Mena Marqués 1984 — per the graded list; the two run probes (Miller, Ortega) cite miller_probe.json / ortega_probe.json.
• The mirror-reading tradition (Foucault, Les Mots et les choses, ch. 1) — the absent-centre argument this study puts numbers under, without adjudicating it.
• Prior ISR studies in the same lineage: the Degas abonné (Study 2), the Bruegel Icarus (Study 7 — the comparison table's source), the HCB doubling network (Study 8), the Klimt address pole (Study 6).
• Study record: PREREGISTRATION.md (10 predictions, locked pre-run) · labbook.md (markdown twin) · report_pm_nopersist.json / report_nopersist.json · subject_probes.py/.json · accent_clean.py/.json · null_address.json · null_address_1000.json · null_massdiv.json · review_checks.json (the review's measurement batch: massdiv decomposition, shadow chain, column sweep, mirror cap-check, seam overlap) · jitter_probes.py/.json (n=200 box jitter, harness-bug-fixed) · lock_margins.py/.json (I-4 decision margins) · doorway_repin.py/.json + scout_doorway_components.png (the annular-target diagnosis) · null_figure_shuffle.py/.json + _robust.py (matched null + seed/fill robustness — address measured at +0.7σ, DGI not tested) + shuffle_variance.py (η² decomposition) · selector_sweep.py + convention_jitter.py (convention at pin + convention × jitter) · clark_grid.py + clark_grid_sweep.py (the trim estimator and the threshold sweep that withdrew its fit-significance) · mask_verification.json / scout_all_masks.png (all nine I-4 locks) · make_charts.py · instrument findings I-1…I-19, candidates C-3/C-4 (toolkit/CANDIDATES.md, OBSERVATIONS.md — appended).