It moved the girl, and put the world back.A Sora pair: "Little Red Riding Hood" by default and by direction · 2026 · the engine since discontinued

Orientation, for a cold reader

Two prompts, one instrument, and the scope up front

The Default frame came from six words: "Little red riding hood descending stairs." The Steered frame came from a direction written in the vocabulary of this research program, displace the figure off the vertical axis, make the right-side void active rather than atmospheric, ban the particle fill, place one hard geometric element against the organic tunnel, overexpose the threshold light, make the stair rhythm irregular, and five refusal clauses (no centre-halo rescue, no symmetrical tree framing, no uniform bokeh). The Image Structure Reader, deterministic, recognising nothing, measuring gradient fields, masses, relations and accents, read both frames in the same coordinate space as six paintings, one calligraphy scroll, and one photograph: eight prior works across three domains. Scope before anything else: this is one pair, one draw per condition, from an approximate engine that has since been discontinued: no further draws can ever be made, so within-engine replication is permanently unavailable. Every "delivered" below is a directional observation of what this draw did, not a claim about placement accuracy or obedience; the study pre-registered twenty predictions and missed ten, and the misses are marked in the record.

The default

A portrait of the basin [stable]

The Default is what the program's spatial-priors research predicts from a naive prompt, measured here in corpus coordinates. Default Gravity Index 57.8: "leans toward the attractor," above every real work in the corpus (Klimt 54.9; the Cartier-Bresson 53.1). Its island system has no protagonist: the anchor is the glow itself, the figure fused beneath it, every top relation weak (r ≈ 0.5), the hierarchy near-flat (entropy 0.964: the corpus's dissolution end). Its punctuation is axial: eight structural accents, every one within |nx| ≤ 0.08 of the vertical centreline, none stronger than 1.08: the halo evens everything. And its deepest signature inverts intuition: this "static" emblem is the most superadditive object the corpus has measured, at R_spatial 12.3–14.2 (the previous record, Klimt's Mäda, was 2.9–5.0), with stance torque 0.572, torquing. The nested radial field (glow inside arch inside tunnel) wanders the multiscale centroid as no part of it does; the emblem is an emergent whole-property, not a sum of centred parts. Against the pre-registered bet of Hard RCP, the five-hit classifier returned Borderline-to-Soft (centre-lock and density-bowl hit; ring-fit measured at 0.287 against a ≥0.50 bar; μ mid): the false-positive guard beating the analyst's own expectation, which is what guards are for.

The steered

The real-work grammar, from a prompt [stable]

The Steered frame lands in different territory on nearly every arrangement axis. DGI 48.6, "leaving the basin." R_spatial 0.03–0.39, subadditive: the cancellation camp where this corpus's Bruegel and Cartier-Bresson live: a whole quieter than its loudest part. Stance 0.023: standing, one heading held. Punctuation rebuilt: 24 accents at the lens cap, spanning the frame (nx −0.20 to +0.40), ridge-dominated (19 of 24: the edges of placed objects), top leverage 1.44: this frame lets marks stand out. And the island graph reproduces a structure this library first measured on a 1932 photograph: a displaced subject joined to a static anchor across the frame's most resistant corridors (r = 0.62–0.67): the Gare Saint-Lazare relational grammar, assembled by a text-to-image engine from a paragraph of direction. The figure herself moved where she was pointed: verified cloak mask at Δx +0.199 against a directed 0.18: the magnitude-class delivered; the near-match quoted as a curiosity of one draw from an approximate engine, nothing more.

The rock

Attribution, not co-occurrence [stable, I-13 gated]

Here is the study's centre, and it needed an experiment, not a correlation. The steered figure stands at +0.199; yet the frame's structural mass centroid reads −0.079 (left of centre) and the imbalance is 0.046, as balanced as the Default. Something is counterweighting her. The ablation test removes the left rock mass and re-measures under two different fills (the library's I-13 gate: only fill-invariant results are citable: these were, to three decimals): without the rock, the mass centroid swings +0.107 toward the figure and the balance tips (imbalance 0.046 → 0.060). Remove the glare instead and the mass doesn't move but the frame's bilateral address drops, though that cell is fill-sensitive and is quoted at reading-aid tier. So the rebalancing is not an inference: the placed rock carries it. The engine granted the displaced subject and built a counterweight into the world around her, which is, note, exactly what the canonical works in this corpus do with their displaced subjects. Whether that is the engine's habit or this draw's luck, one pair cannot say.

What did not move

The house style under the arrangement [stable / narrowed]

Between the two frames, the layered delta reads: substrate 1.01 (different pixels entirely), form 0.124 (the arrangement moved), tone 0.035: nearly identical tonal policy. Both frames are shadow-massed in the Caravaggio corner of the corpus (0.822 / 0.743); both put colour in light's service (the corpus's two highest couplings, its two lowest β; cbi twins at ~0.09). Precision the record insists on: coupling itself is not a twin; it moved 0.386 → 0.571, and under the surrogate band below that jump is the pair's single most robust change, plausibly riding the placed glare and column (edge-chroma events). Which fits the pattern rather than breaking it: what the direction moved, it moved by placing objects; what the engine kept, it kept in the statistics: the palette-and-light physics passed through steering untouched everywhere an object didn't drag it.

How sure can one pair be?

Three rankings, one reference band, and a convergent stack [narrowed]

Three tools name three different "top separators" for this pair, and the disagreement is instructive rather than embarrassing. The corpus comparison (mean-normalised, small-n) fronts torque and grid asymmetry; the absolute position decomposition fronts cohesion and vertical centroid; and the surrogate band asks a third question: how often a matched-statistics null of the Default moves an axis as far as the steering did. That band (n=100 per null; phase-scramble the harsh bracket, patch-shuffle (local texture kept) the more draw-like one; reported as distribution-free exceedance p (the fraction of null perturbations whose scatter matches the observed steering delta) because these AI-image nulls are non-Gaussian and σ misreads their tails; τ excluded on principle, since both nulls preserve the histogram) reads: coupling p <0.01 / <0.01 (scramble/shuffle) and cohesion p 0.01 / <0.01 both clear both brackets: the pair's two robust scalar changes; lateral centroid p 0.53 / <0.01 and DGI p 0.43 / <0.01 clear the draw-like bracket but not the harsh one; torque's delta p 0.06 / 0.06 is marginal, though the sharper torque fact is one-sided: the Steered's standing (0.023) falls below the minimum of all two hundred null draws, so what beats the noise floor is not "the emblem torques" (a scrambled image torques too) but "the steered frame stands." And imbalance p 0.92 / 0.75, well inside both bands, which here is a finding, not a null: both real frames are more balanced than the nulls, and the ablation shows the engine actively rebuilds that balance, so an unmoved imbalance is the fingerprint of a working counterweight, not of inaction. Null variance is not draw variance, with the engine gone, this is the only reference band the closed artifact permits, and it is labelled a surrogate throughout. The honest conclusion the three tools converge on: no single scalar carries the steering claim. What carries it is the stack that does not depend on any one axis, the anchor-and-corridor grammar, the axial-versus-distributed punctuation, the superadditive-to-subadditive flip, the stance flip, and a rock that fails the balance when you remove it.

Optional interpretation

Clearly marked, downstream of the numbers [reading]

The pair reads as a measurement of what a generative engine will trade. Asked plainly, it performs: the halo, the axis, the emblem, the basin. Directed in compositional language, it negotiates: it places what it is asked to place, and the placed things do real structural work (but it composes its way back to equilibrium behind them, and its tonal policy) with the colour policy that isn't dragged along by a placed object: never comes up for negotiation. The steered frame is the more canonical composition by every grammar this corpus has measured on real works; the default frame is the more extreme object, a superadditive emblem beyond anything the paintings produced. One engine, caught once, now gone. An archival pair: the basin, and one documented escape from it.

What this does not prove

The scope, stated plainly

The record (lab book, Entries 01–16) carries the pre-registration with its ten marked misses, the RCP guard beating the analyst's own Hard-RCP bet, a fill-invariance-gated ablation, an n=100 two-null surrogate band, and three review rounds' corrections in place. The study's most-quoted numbers were wrong first and corrected in public; that is the method working.

Sources & record

Citations and artifacts

• Prompts, verbatim, and the engine note: PROVENANCE.md. Pre-registration (20 predictions, before any number): PREREGISTRATION.md.
• Program grounding, cited not re-derived: Measuring Spatial Priors in Generative Image Models; the VTL Kernel Metrics drift study; the RCP classifier and AI aesthetic-defaults reference (1.All_Skills).
• Study record: studies/AI-red-riding/: labbook.md + HTML twin (Entries 01–16), report/candidate/ablation/null-band JSONs, island and accent readouts, verified cloak masks. Register: toolkit/CANDIDATES.md (C-3, C-4; fourth domain, n→11).
• Instrument lineage: the ISR library: Matisse, Degas, Pollock, Caravaggio, Soejima, Klimt, Bruegel, Cartier-Bresson, Sora Pair, MidJourney Ensemble, and Velazquez