Study 9 — Sora pair, Little Red Default vs Steered — lab recordImage Structure Reader · Study 9 · lab record · fourth domain (AI-generated)
Read 2026-07-12. Born-digital (Sora, user-generated); prompts recorded verbatim in PROVENANCE.md before the battery ran. Pinned Little_Red_Default.webp / Little_Red_Steered.webp (1536×1024 each, matched; lossy webp = the reproduction caveat — coarse composition metrics only). Pre-registration before any number: PREREGISTRATION.md (20 predictions). The corpus's first AI-generated images (fourth domain) and its designed-for case. First-pass numbers report_{name}.json (edge-aware, no persistence); persistence quotes from report_full_{name}.json. Candidates via the new null_sweep (its first native study).
More info: www.parallaxmetrology.com, full critique
Labels: [stable] pipeline-stable · [narrowed] scoped · [aid] reading aid · [confirmatory] pre-registered · [exploratory] post-hoc.
Entry 01 — Selection, condition, and the steering contract [stable]
The pair structure is the study: one naive prompt ("Little red riding hood descending stairs") and one steering prompt written in VTL vocabulary with quantitative targets (Δx = 0.18; broken center-lock; active right void with a particle-fill ban; a geometric-vs-organic architectural element; overexposed threshold light; irregular stair rhythm; figure torque; five refusal clauses; "maintain peripheral pull"). So beyond default-vs-steered difference, the instrument can read the pair against the direction given — scoped at review (user, 2026-07-12): this is one pair from an approximate engine, not a study of placement accuracy; "obedience" would need a corpus. Everything in Entry 08 is a directional observation of what this draw did, reported with its metric. Engine note: Sora has since been discontinued — the pair is a closed artifact; no further draws exist or can be made, which fixes this record's status (one engine's defaults, caught once) and its limits. Conditions: born-digital, full-bleed, no PM indicated (no grain; webp compression noted instead — no fine-structure claims); colour subsystem ACTIVE for the first time since Study 2; both edge modes owed (content to all edges); RCP guard adopted verbatim (no model-internals claims; "Default"/ "Steered" are the user's labels — the instrument tests whether the geometry matches them).
Entry 02 — Headline: the basin gap is real but smaller than bet [confirmatory, thresholds missed]
DGI: Default 57.8 "leans toward the attractor" vs Steered 48.6 "leaving the basin." Direction confirmed; both pre-registered thresholds missed (predicted >60 and <45; got 57.8 and 48.6; gap 9.2 vs the >15 bet). Scored honestly: the steering moved the image one full band out of the basin, and the Default sits above every real work in the corpus (Klimt 54.9, HCB 53.1) and inside the matched-stats null band (57–64) — the default output scores like a null of a real composition, which was the qualitative bet behind the number. Kernel: Default Δx +0.063 (center-locked), Steered mass Δx −0.079; radial both neutral (RC_f ≈ 0.91); imbalance 0.041 vs 0.046 — both balanced; state both counterweighted.
Entry 03 — The rebalancing discovery: Sora moved the figure and restored the field [stable — the study's center]
The steered frame displaces the subject hard (verified cloak mask: figure Δx +0.199, from a centered Default baseline of +0.01…+0.07). The contract asked for 0.18; the delivered magnitude-class matches, and the two-decimal proximity is quoted as a curiosity, not a capability — Sora is an approximate engine, this is one draw, and no numeric-following claim is made (user scoping, 2026-07-12). But the structural mass Δx is −0.079: the heavy left rock face counterweights her almost exactly, imbalance stays 0.046, and the mass centroid returns to within 0.08 of centre — technically inside the RCP center-lock bound. The engine granted the displaced subject and quietly rebuilt a balanced mass field around her. The island graph names the new structure: anchor = the left rock/stair mass, with the figure-side islands joined to it across the frame's three most resistant corridors (r = 0.62–0.67) — the Gare Saint-Lazare grammar (displaced subject, held apart, counterweighted), produced by a text prompt. The Default's graph is the halo grammar instead: anchor = the glow arch itself (island 16), figure fused beneath it, all top relations weak (r ≈ 0.5), island hierarchy near-flat (entropy 0.964 — Degas-adjacent).
Entry 04 — Stance and R_spatial invert the pre-registration, in a crossed pattern [stable, surprising]
Predictions 3/12 called Default "standing" (the C-2 AI cluster) and Steered torqued. Measured: Default torque 0.572 — torquing; Steered 0.023 — standing. And R_spatial (prediction 8/16, badly wrong): Default 12.3–14.2, superadditive — the most extreme value in corpus history (previous max: Klimt Mäda 2.9–5.0); Steered 0.03–0.39, subadditive — the HCB/Bruegel cancellation camp. The physics intuition that "centred symmetric = cancellation" was exactly backwards for this construction: torque is multiscale centroid wander, and the Default's nested radial field (glow core inside arch inside foliage tunnel) makes the whole wander as no part does — the emblem is an emergent whole-property. The corpus's two most superadditive works are now the Klimt ornament-ground portrait and the AI radial emblem; the steered image lands with the real distributed-order works. Nulls (5 seeds each, null_sweep discipline): seed-chaotic ranges that never reach the Default's 12–14 (null max 6.1) and never reproduce the Steered's stable subadditive verdict — stability-vs-chaos, both. (R_spatial nulls are n=5 per the per-metric discipline — the softest null in the study; the Default's ~2× margin over the null max is quoted with that caveat.)
Entry 05 — Address goes figural, both frames (first composition-dependent address outside the Mäda) [stable]
null_sweep at n=100 per null, first native use. Default central frontality 0.484, p(scramble ≥ real) = 0.03 — prediction 5's "at least partly figural" hit: the centred figure owns the central column. Steered 0.590, p = 0.01 — prediction 13 (< 0.30) wrong: the steered frame builds a bilateral light-column (glare + stair axis) that no figure owns; per the Bruegel scope note this is bilateral-structure address, not figural address — but unlike the Bruegel it dies under the null: it is arrangement, not spectrum. A third category for C-4's scope: composition-dependent, non-figural. Also logged: the AI images' scramble distributions are themselves seed-chaotic (sd 0.21–0.25 vs ~0.07 on paintings) — the null's tightness is image-class-dependent.
Entry 06 — The tonal house style survives steering [stable]
My Default tonal bet (midtone vignette) was wrong: both frames are shadow-massed tenebrist distributions (Default τ 0.618, shadow 0.822!; Steered τ 0.555, shadow 0.743) — and prediction 14 (Steered harder) also wrong: the Default has the harder Gini. The layered delta makes it a finding: substrate NCD 1.01 (different pixels entirely), form delta 0.124 (arrangement moved), tone delta 0.035 — nearly identical tonal policy. The steering rewrote the composition; the engine's shadow-heavy, warm-accent tonal fingerprint passed through untouched. Colour agrees: coupling +0.386 / +0.571 (colour tracks light in both — the "aesthetic alignment" signature), cbi ~0.09 both (no chroma boundaries light doesn't draw; prediction 15 wrong), β ~0.06 both (colour lives on structure, not in the void). Persistence: both fragile_to grayscale (0.679 / 0.763) — here legitimate (real colour images, not tint artifact; contrast Study 8 Entry 10).
Entry 07 — RCP five-hit table: the false-positive guard beats my bet [narrowed]
Pre-registered per-hit (prediction 2: Hard). Measured evidence (rcp_contract_audit.py): center-lock HIT (mass +0.063, figure +0.01…07, both < 0.08); density-bowl HIT (radial luminance profile perfectly monotone, 1.0, centre-to-rim drop 0.60 — textbook); radial-void borderline (sector-void sd 0.106 — the glow wraps but unevenly); μ-inflation MISS (μ 0.493, mid); ring-fit MISS (ring-aligned share 0.287, far under the ≥50% bar — my pre-registered uncertainty resolved against). Count: 2 solid + 1 borderline → Borderline-to-Soft RCP, not Hard. The naive prompt produced a soft basin (vertical stair axis, figure below centre, right-heavy particles are all directional residues), and the guard — require multi-signal agreement — did exactly its job against my over-bet. Steered: center-lock technically passes on mass (−0.079; the rebalancing of Entry 03), density-bowl partial (0.89 monotone — the glare is still near-central-top), everything else misses → Borderline (2 weak hits), Not-RCP in spirit but not in letter. The letter matters: it is the measured trace of what steering could not remove.
Entry 08 — The direction given, read against the result — directionally [stable, scoped]
| contract item | delivered? | evidence |
|---|---|---|
| figure Δx = 0.18 | YES in magnitude-class (+0.199) | verified cloak mask; Default baseline ~+0.01; the near-exact match is one draw from an approximate engine — direction and class delivered, precision not claimed |
| break center-lock (figure) | YES | figure off-axis, stride + look-back |
| …but reduce centripetal suction | PARTIAL | mass field re-centred (Δx −0.079), imbalance 0.046, DGI 48.6 = one band out, not out of orbit |
| — center-lock note | hairline | the −0.079 sits 0.001 inside the 0.08 RCP bound — per house discipline a hairline is a reading aid, not a verdict; the rebalancing claim rests on imbalance + the anchor graph, which are not hairline |
| active right void, particle ban | MOSTLY | bright particles right: 160 → 37 (−77%); not zero |
| geometric element (column OR arch) | YES — both | column left-mid + arch right (the "2 elements max" clause consumed) |
| overexposed threshold light | YES | white-hot glare; highlight_mass 0.029 → 0.112 |
| irregular stair rhythm | visually yes | sci_rhythm 0.29 → 0.36; not probed further |
| add torque to figure | INVERTED at field level | field torque 0.572 → 0.023 — the figure torques, the FIELD stands |
| maintain peripheral pull | no purchase [aid] | x_p 0.099 → 0.084 — a 0.015 move on ~0.09, on THE metric the library names as gripless on centred images (I-15); purchase unverified, so this row is a reading aid, not a refused-directive finding (review fix) |
| refusal clauses (no halo, no symmetric trees, no bokeh fill) | YES | see readouts; halo replaced by directional glare |
The pattern, stated directionally (one pair, one draw, no accuracy claim): the object-level directions all moved the frame the way they pointed; the field-level directions did not move their targets — mass balance restored itself, peripheral pull did not rise, tonal policy passed through. What did change at field level changed through the placed objects: the rock counterweight, the glare, the diagonal. (No intent or internals claimed; and with the engine discontinued, this table is a record of what one system did once, not of what it does.)
Entry 09 — Corpus placement [aid, n=1 per label]
Default sits above every real work on DGI and beyond the corpus range on R_spatial superadditivity — the basin made visible in the shared coordinate space. Steered interleaves with the real distributed-order works (R_spatial in the Bruegel/HCB camp, counterweighted state, resistant-corridor grammar). Pair compare top-CV rows: torque, saliency-divergence, dRC, highlight, grid asymmetry, off-centre force — the basin axes, not facture (prediction 18 held). Candidates logged to CANDIDATES.md: C-3 and C-4 each +2 works, fourth domain, n→11.
Entry 10 — Raw-mode cross-check: within the envelope [stable]
Both frames re-read with edge suppression off (user prompt: the readouts ran edge-aware). Steered: zero headline movers. Default: counterweight bookkeeping only (7 → 5 counterweights, top counterweight relabelled, bridge cost 0.53 → 0.46); kernel, tonal, radial, DGI all stable. Born-digital full-bleed frames sit inside the border-suppression envelope; edge-aware quoted throughout with no caveat owed.
Entry 11 — Two punctuation systems: axial vs distributed [stable]
The accent lens splits the pair more cleanly than any scalar. Default structural channel: 8 accents (under-cap), every one within |nx| ≤ 0.08 of the vertical centreline, top leverage 1.08 — the punctuation system collapsed onto the central axis (stair-tread lights below the figure, figure/lantern darks above; positions quoted as coordinates, object language indicative). Only the chromatic channel escapes the axis (19 accents, the corner flora at nx −0.34…+0.23). Steered structural: 24 accents (at cap), spanning nx −0.20…+0.40, top leverage 1.44 — a distributed field (column/glare boundary, figure region, right arch, stair diagonal), with the chromatic system agreeing (16, same spread). Steering rebuilt the punctuation from one axial column into a frame-wide field, and gave the frame genuine outlier marks (top leverage 1.08 → 1.44): the halo evens everything; the steered frame lets marks stand out.
Entry 12 — Corpus placement (n=8): the engine owns the colour axes, steering owns the arrangement axes [stable, placement]
The two Sora frames enter the full corpus and pair on exactly the axes steering never touched: coupling_index top two in the corpus (colour welded to light) and β bottom two (no colour-in-void). Precision (review fix): β, cbi, and tone-delta are genuinely near-identical across the pair; coupling is NOT (0.386 → 0.571, a ~48% jump) — both are corpus-high, but the jump is real and plausibly downstream of the placed glare/column (edge-chroma events), which fits the moved-by-placing-objects story rather than the "identical" one first written here. Shadow_mass: both in the Caravaggio corner (the tenebrist house key). And they split maximally on the arrangement axes: Steered posts the most asymmetric mass grid in the corpus (grid_asym z +2.05); torque separates them to opposite poles (Default +1.41, beside Degas; Steered −0.92, beside HCB/Bruegel); saliency-divergence likewise. The one-line version, consistent with Entry 06's delta: the engine is a statistics; the steering is an arrangement. Candidates logged: C-3 and C-4 rows appended (both frames, fourth domain, n→11) in CANDIDATES.md.
Entry 13 — Synthesis [reading]
The pair measures what a text-to-image engine will and will not trade. Asked naively, Sora produced the basin: a soft-RCP radial emblem whose anchor is its own light source, whose punctuation lies along its axis, whose island hierarchy is nearly flat, and whose whole out-torques every part (R_spatial 12–14, the most superadditive object the corpus has measured — the emblem is an emergent whole-property, kin to the Klimt ornament-ground, beyond it). Asked with VTL vocabulary, the engine delivered every object it was asked for — the displaced figure, the column and arch, the glare, the particle ban — and the delivered objects assembled into the real-work grammar: a counterweighted standoff with resistant corridors to a displaced subject (the Gare Saint-Lazare relational structure, from a prompt), distributed punctuation, cancellation torque (R_spatial 0.03–0.39, the Bruegel camp). But the field-level requests were refused or absorbed: the mass centroid re-centred behind the displaced figure, peripheral pull fell, and the tonal and colour policies passed through untouched (tone delta 0.035; β and cbi twins; coupling corpus-high in both but shifted 0.386 → 0.571 — see Entry 12's precision note). What steering moved, it moved by placing objects; what the engine kept, it kept in the statistics. The instrument's one-sentence account: steering changed the composition; nothing changed the palette-and-light physics the engine composes with.
What This Does Not Prove
- Not model internals. No claims about Sora's training, weights, sampling, or intent (RCP guard adopted verbatim). "Delivered/refused" is geometry audited against a prompt text, on one pair.
- Not a steering law, and now unreplicable as posed. n=1 pair, one engine, one prompt family, one draw per condition — and Sora has been discontinued, so no further draws exist: the within-engine replication that would separate steering effect from draw noise is permanently unavailable. What replication remains is cross-engine (a different system, which changes the question) or the corpus form (the program's spatial-priors paper is the 1,400-image version). The Δx magnitude match stays a curiosity of this one draw (and "two-decimal" was the wrong phrase
besides — 0.199 rounds to 0.20).
- Not a quality verdict. "Basin" and "grammar" locate structure; neither frame is graded.
- Not fine structure. Lossy webp: box-counting, skeletons, and sub-8px accents unquoted throughout.
- Not figural address on the Steered. Its 0.590 is composition-dependent bilateral structure (glare + stair column) — a third C-4 category, not a figure facing the viewer.
- Object attributions in Entry 11 are positional (coordinates) with indicative object language; the uniform pixel audit of every accent (Study-8 Entry-20 style) was not run this round and is queued.
- Open items. RCP ring-fit overlay figure (numeric share 0.287/0.279 is in the record; the visual evidence panel is owed); single-draw generation variance is now unmeasurable within-engine (Sora discontinued) and stands as a permanent scope limit rather than an open probe; the HTML twin assembles now.
Entry 14 — Review round: three cheap extensions run, three queued [stable]
Run from data in hand (user review, 2026-07-12). Position decomposition (corpus-independent, replacing small-n z-reliance): pairwise d = 0.450, driven by cohesion 0.379, vertical centroid 0.182, highlight 0.091 — and the pair's shared loudest axes are orientation entropy (0.995/0.988), void ratio (0.72 both), and mass-island count: the split is arrangement-side, the shared floor is engine-side, said without a reference corpus. Accent kinds: steering shifted punctuation type, not just spread — Default's 8 structural accents are a point-accent/ridge mix; Steered's 24 are ridge-dominated (19 of 24: 13 light_ridge + 6 dark_ridge — the edges of placed objects; the other 5 are filaments/point-accents). Role/delay-grammar fields are not populated in this export — queued with the pixel audit. Edge commitment as a second fingerprint: REFUTED — the frames do not share a soft-structure profile (Default MIXED, gap_fraction 0.408; Steered QUIET_MODERATE, 0.569): edge commitment moved with the steering, so the engine-fingerprint claim stays one axis-family (colour + tone), narrower and better-supported. Queued as the next tranche, in value order: (1) stratigraphy ablation of the left rock mass and the glare (I-13 fill-invariance gated) — would turn "steering moved things by placing objects" from co-occurrence into attribution; (2) per-axis null ensembles as a draw-noise surrogate band (null variance ≠ draw variance, said loudly — but it is the only reference band the closed artifact permits); (3) Color Kernel triangulation of the colour-fingerprint claims (the designed cross-check, and the honest test of the coupling wobble).
Entry 15 — Ablation: the rock is the rebalancer (attribution earned) [stable, I-13 gated]
ablation_probe.py, Steered frame, two fills per cell. Rock ablation is fill-INVARIANT (Δx +0.0273 both fills; imbalance 0.0599/0.0596): removing the left rock mass moves the mass centroid from −0.079 to +0.027 — a +0.107 shift toward the displaced figure — and lifts imbalance 0.046 → 0.060. Entry 03's central claim is now attribution, not co-occurrence: the placed rock carries the rebalancing. Specificity control: rock ablation leaves address untouched (0.590). Glare ablation is fill-SENSITIVE (address 0.590 → 0.481 median / 0.511 border): the glare carries a real part of the bilateral address but the cell lives between strata (I-13) — quoted as a reading aid, not a verdict.
Entry 16 — The surrogate band at n=100, both nulls, distribution-free [narrowed]
null_band.py v3 (v2's σ adjudication retracted: Entry 05 shows these AI-image nulls are non-Gaussian, so σ→percentile is unsafe on the marginal axes; and v3's first empirical draft used a flawed statistic — distance from the Default — which conflates the null's systematic mean-shift with its scatter, both caught at review). The honest statistic saves every draw and reports p_dev = fraction of the 100 null perturbations whose deviation from the null mean is ≥ the observed Default→Steered delta (the distribution-free analog of Δ/sd, using the null's own scatter as the draw-noise proxy), plus ste_tail = where the Steered value falls in the null cloud. Phase-scramble = harsh bracket; patch-shuffle (local texture kept) = draw-like bracket. τ excluded on principle (both nulls preserve the histogram). Null variance ≠ draw variance — the only reference the closed artifact permits.
| axis | Δ | p_dev scramble / shuffle | reading |
|---|---|---|---|
| coupling | 0.184 | 0.00 / 0.00 | clears both — the pair's one unambiguous scalar change |
| μ | 0.379 | 0.01 / 0.00 | clears both (σ undersold it at 2.2σ harsh — distribution-free it clears the harsh bracket too) |
| torque | 0.549 | 0.06 / 0.06 | marginal on the delta — but see the note: the Steered value is below every null draw |
| DGI | 9.2 | 0.43 / 0.00 | clears draw-like, unresolved vs harsh |
| delta_x | 0.142 | 0.53 / 0.00 | clears draw-like, unresolved vs harsh |
| imbalance | 0.005 | 0.92 / 0.75 | inside both — but both real frames are more balanced than the nulls (below-tail); see Entry 15 |
Two conclusions changed from the σ draft. (1) μ clears both brackets — distribution-free, the Steered's cohesion jump exceeds 99% of scramble scatter; it is a second robust scalar, not a marginal one. (2) The torque finding inverts. Torque's null floor is high and barely tightens scramble→shuffle (sd 0.33 → 0.26, versus 4–10× tightening on every other axis) because both nulls destroy spatial arrangement and a spatially-disordered mass field has intrinsically large multiscale centroid wander — so high torque is a property the Default shares with noise (0.572 sits mid-cloud, null range [0.055, 1.781]). What is distinctive is the Steered's torque: 0.023, below the minimum of every one of 200 null draws (scramble min 0.055, shuffle min 0.024). The null-beating signal is not "the emblem torques" — a scrambled image torques too — it is "the steered frame stands," more ordered than any matched-stats perturbation of the Default. The stance flip survives; its emphasis moves from the Default to the Steered.
imbalance, read correctly (the defensive point). The pair barely differs on imbalance (Δ 0.005, inside both bands) — but that is not inaction: both real frames sit in the low tail of the imbalance null (more balanced than 92% of scrambles), and Entry 15's ablation shows the engine actively rebuilds that balance behind the displaced figure. A flat imbalance is the signature of a working counterweight, not of nothing happening.
Reconciling the three rankings (review item). Compare (mean-normalised CV, small-n) fronts torque and grid asymmetry; position (absolute displacement) fronts cohesion and vertical centroid; the band (exceedance of a null floor) fronts coupling and μ. They rank differently because they ask different questions — relative dispersion, absolute distance, noise-floor exceedance — and the two axes the ranking tools front (torque, cohesion) are exactly the ones whose deltas are marginal against the harsh band, even as torque's underlying stance flip survives on the Steered tail. The convergence is the point: no single scalar carries the steering claim. What carries it is the stack no single axis defines — the anchor-and-corridor grammar, axial-vs-distributed punctuation, the superadditive-to-subadditive flip, the Steered's below-all-nulls standing, and the fill-invariant rock. Color Kernel triangulation remains queued (#3).
Pre-registration scorecard (20 predictions, adjudicated)
| # | Prediction | Verdict |
|---|---|---|
| 1 | Default DGI > 60 | Miss by 2.2 (57.8) — but above all real works, inside the null band; qualitative claim stands |
| 2 | Default Hard RCP (4–5 hits) | Wrong — Borderline-to-Soft (2–3); the guard worked |
| 3 | Default stance standing | Wrong — torquing 0.572 |
| 4 | Default RC_f > 0.90 | Hit (0.906) |
| 5 | Default address figural, p < 0.05 | Hit (0.484, p = 0.03) |
| 6 | Default coupling 0.15–0.35, cbi < 0.15, red = chromatic accent | Partial (coupling 0.386, just above; cbi 0.092 ✓) |
| 7 | Default midtone-dominant inverted vignette | Wrong — shadow-massed 0.822, τ 0.618 |
| 8 | Default R_spatial < 0.35 subadditive | Spectacularly wrong — 12.3–14.2 superadditive, corpus record |
| 9 | Default mass_div no purchase | Partial (0.056 — small but nonzero) |
| 10 | Steered DGI < 45 | Miss by 3.6 (48.6); direction right |
| 11 | Steered Not-RCP (0–1) | Close — Borderline (2 weak); the mass-rebalance keeps center-lock alive |
| 12 | Steered torque > 0.20 | Wrong — 0.023, standing |
| 13 | Steered address < 0.30 | Wrong — 0.590, composition-dependent bilateral |
| 14 | Steered τ > Default | Wrong (0.555 < 0.618) |
| 15 | Steered cbi > Default | Wrong (flat, ~0.09 both) |
| 16 | Steered R_spatial > Default | Wrong, inverted (0.03–0.39 vs 12.3–14.2) |
| 17 | Steered saliency divergence larger | Wrong (0.009 < 0.056) |
| 18 | Pair separates on basin axes, not facture | Hit (top-CV: torque, divergence, dRC, grid_asym) |
| 19 | Form delta exceeds same-work band | Directionally right (form 0.124, substrate 1.01, tone 0.035) |
| 20 | Default nearest the nulls; Steered interleaves with real works | Hit (DGI + R_spatial + grammar) |
Ten clean misses out of twenty (#2, 3, 7, 8, 12, 13, 14, 15, 16, 17 — the first tally said eight and dropped #13 and #17; corrected at review, a factual error in the self-audit, the one place not to have one) — the most wrong this analyst has been in the corpus, and the misses are the findings: the emblem is superadditive and torquing, not standing-subadditive; the steered frame is the one that stands; the tonal house style is steering-proof; the guard beats the classifier bet.
Status
Tranche 4 complete (Entries 01–16): ablation attribution earned (rock = rebalancer, fill-invariant); surrogate band rerun at n=100, both nulls, DISTRIBUTION-FREE (σ and the distance-from-default draft both retracted at review): coupling AND μ clear both brackets; the torque finding inverts (Steered stands below all 200 nulls; the Default's high torque is shared with noise); imbalance flat but both frames beat the null — coherent with the rock ablation. Color Kernel queued. Review round applied: scorecard tally corrected (10 misses), coupling claim narrowed, x_p downgraded to no-purchase [aid], the center-lock hairline retagged, R_spatial null n=5 flagged. Ablation, null-band, and Color Kernel triangulation queued. Raw-mode check done (Entry 10), accent systems read (Entry 11), corpus n=8 placed (Entry 12), CANDIDATES rows appended (fourth domain, n→11), Δx claim scoped per user guidance. Queued on confirmation: HTML twin, then the critique. Open probes listed in What This Does Not Prove.