# WorldAlign research log This file records research decisions, failed hypotheses, and the current experimental gate. Exact metrics live in the result reports and machine-readable artifacts. The collaborator-facing statement of the project is maintained in `PROJECT_CONCEPT.md`. ## Research target The target is a real multimodal large-model system, not a geometry result or a toy correspondence benchmark. The working hypothesis is that independently trained, sufficiently informative unimodal representations may recover the same world structure up to a change of coordinates. Multimodal alignment could then be obtained by descending a high-level shared-world energy, without paired contrastive supervision. Important constraints: - Do not treat the word "graph" as a commitment to explicit semantic nodes. It is only an analogy for richly constrained relations in learned states. - Node-wise or sample-wise sufficiency is not enough. The observations must be overcomplete enough to constrain relations among world states. - Generic encyclopedia text is not a closed counterpart of the visible world: it contains many latent factors that images cannot identify. - The target method is not an image-to-language MLP bridge. The bridge baseline is useful only as an interface/control. - Paired examples may be used for held-out diagnosis, never by the unpaired energy, optimizer, initialization, model selection, or stopping rule. ## 2026-07-28: learned-bridge baseline Flickr30k, frozen DINOv2 and frozen Qwen were aligned with static GW/SWD/isometry objectives. - Cross-modal relational similarity is measurably non-random. - Static GW produces sharp but mostly incorrect correspondences. - A paired bridge verifies that the frozen-language-model interface works. - The unpaired learned bridge does not produce usable multimodal ability. See `RESULTS.md`. ## 2026-07-29: overcomplete observations Five independent captions of the same scene were treated as an observation orbit. On 512 held-out examples, the correct-vs-shuffled relation-energy gap increased from 0.316 with one caption to 0.582 with five captions. This supports the narrower claim that relational overcompleteness makes the correct configuration easier to distinguish. It does not establish that the current energy identifies the correct absolute configuration. ## 2026-07-29: energy descent without a cross-modal map Free Qwen-side latent particles were initialized from real text states and optimized using: - standardized relation-field matching; - multiscale conditional-neighborhood matching; - sliced text-population matching; - proximity to text-only prototypes. Weak manifold regularization briefly improved held-out R@10 from 2.34% to 12.11%, but left the decodable language manifold. Strong regularization improved it only to 6.25% at the best intermediate step. Language-prior-only and shuffled-vision controls showed that caption-generation improvements were not reliably image-conditioned. ## 2026-07-29: convergence audit and falsification The original logged total was not a fixed objective because relation and conditional weights were annealed. Therefore that curve could not establish convergence. Starting from the strong-manifold (`m30`) endpoint, the final weights were frozen and optimization was continued at one tenth of the learning rate: - fixed energy decreased from 2.767 to approximately 1.54; - R@10 decreased from 5.86% to approximately 3--4%; - median rank worsened from 103 to approximately 120--123; - random R@10 for 256 candidates is 3.91%; - gradients remained nonzero and Adam continued to oscillate. Most decisively, the true paired language configuration has fixed energy about 4.49, while an incorrect synthetic configuration reaches about 1.54. Therefore the present energy is falsified as an alignment objective: \[ E(\text{correct paired state}) > E(\text{incorrect synthetic state}). \] The early retrieval gain was a trajectory transient, not evidence that further energy descent would recover the correct gauge. More steps or optimizer tuning cannot repair this ordering. ## 2026-07-29: on-manifold assignment gate Decision, confirmed with the project owner: restrict the configuration space to permutations of real frozen text states. On this space the language-only terms of the falsified energy are permutation-invariant constants, and the continuous counterfeit route is closed by construction. Per the standing gate, only ranking and local-ordering diagnostics were run; no new descent. Result: the falsification survives the restriction. - Global ordering always passes: the true assignment ranks below all 1,000 random permutations in every condition (z 2.7--18.7). - Local ordering always fails: 6.8--33% of transpositions of the true assignment lower the energy (exact closed-form enumeration); at N=5,000 the count is 2.70M of 12.50M. Scale does not repair it. - Exact steepest 2-swap descent from the truth keeps 2.7--13.9% of nodes; random-start quenches reach energies 0.9--1.9 below the true configuration at 0.0--0.4% accuracy. On-manifold counterfeits exist in every condition. Mechanism, from three audits: pooled states are extremely anisotropic (99.5% of Flickr caption pairs above cosine 0.8); the cross-modal share of standardized relational structure is Spearman 0.19--0.31; that share is spectrally concentrated (top-16 population directions: 0.367; full space: 0.306; whitening collapses it to 0.03). Five-caption orbits and visible relations denoise within this coarse subspace; distribution-valued edge channels (std/q10/q90 of view-pair cosines) carry no shared signal and dilute the energy; model scale was already known to be inert. A roughly sixteen-dimensional topical code cannot separate 512 instances, so no energy over these node states can identify the assignment. See `MANIFOLD_RESULTS.md`. ## 2026-07-29: view-level shared structure inside VG nodes The per-node view permutation hidden at preparation time is exactly replayable from the seed and the private image IDs; the replay reconstructs all 5,000 released phrase bundles verbatim and is used only for evaluation. - Within-node relation fields align across modalities: matched Spearman 0.128 against -0.002 shuffled, z = 46.6 over 5,000 nodes. - With the true node-level alignment as a coarse frame, population-context profiles match the 16 views at 14.5% accuracy against 6.25% chance (top-3 29.5%); a larger frame helps (200 nodes: 12.9%). This is an upper bound: a blind system must substitute its own recovered coarse frame. - Within-node relations alone stay near chance (7.0%), so the population context carries the signal. The fine-grained shared structure missing at node level exists one level down, in the compositional structure that pooling discards. ## 2026-07-29: Ricci-flow control and static view-level gate Two same-day continuations, both requested before the structural redesign. Ollivier-Ricci flow on each modality's kNN geometry (independent, identical hyperparameters, diffusion-only baseline for attribution) more than doubles the global ranking z (Flickr 21.4 to 47.8; VG 6.9 to 22.5) and changes no verdict: 9--14% improving transpositions, descent retention 1--5%, counterfeits at about half the true energy in all four conditions. Geometric smoothing is a frame ingredient, not a resolver; it cannot be part of the alignment transform itself (information loss), matching the whitening result. The static view-level gate fails at scale: with the true coarse frame, none of 5,000 nodes has the true view assignment as a strict local minimum (32.0% improving for the within-node relational energy, 28.3% for the profile energy, 25.5% joint). Frame-free internal-field matching is null at 512 nodes over three seeds (R@10 at chance, zero margin-decile precision, true-pair z 0.1--1.8 after a capacity null; the negative z seen at 96-node smoke scale did not replicate). Consequence: the view level needs model-internal relations or true re-encoding interventions -- the R3/R8 battery in `STRUCTURE_DESIGN.md` is the critical path. ## 2026-07-29: representation batteries and the first passed basin audit The R6 and R3 batteries of `STRUCTURE_DESIGN.md` were run the same day. R6, within-scene noise whitening: directions maximizing between-scene over within-scene variance, fitted per modality on unpaired orbits (train-only for Flickr, holdout-only for VG). Cross-modal relational Spearman rises 0.298 to 0.358 (Flickr orbit), 0.189 to 0.326 (single caption, now above the raw orbit mean), 0.213 to 0.614 (VG both sides, k=8); improving-swap fractions collapse 7.8% to 0.5% and 23.5% to 0.25%. Population whitening had destroyed the same signal; the only difference is whitening the within-scene rather than the total covariance. R3, model-layer readout: joint-context encoding on all 5,000 VG nodes. Attention weights are modality-specialized (best attention-attention matched rho 0.076); in-context states carry rho 0.216 against the 0.128 isolated baseline (z 69.6 vs 46.6). Relations do live in the model layer, and the shared readout is the mixed states, not the mixing weights. Full gate on the new states: nine of nine VG conditions pass the basin audit -- ten random-restart quenches per run all end 5--19% above the true energy, replicated over subset seeds 0--2, for both isolated projected states and the in-context composite (which roughly doubles the margin at matched k). First energies in the project whose descent finds nothing below the truth. Flickr still fails by under 1%: one visual view means no visual orbit to whiten, so two-sided multi-view observation is a data-protocol requirement. The true assignment is still not a strict local minimum (0.17--0.28% improving swaps; descent keeps 52--59%), so what is licensed is soft-coupling recovery, not hard matching. See `BATTERY_RESULTS.md`. ## 2026-07-29: blind recovery enters the hard phase First Phase-2 experiment on the gate-passing states, hidden-shuffle protocol, three search arms. Parallel tempering (eight replicas, exact closed-form deltas, 60k rounds) equilibrates into the quench band of the gate report -- ending 9--20% above the true energy at 1--3% accuracy -- and never enters the true funnel; entropic Sinkhorn and spectral-band homotopy stall at +62--100% with zero verified seed mass. Discrete annealing beats continuous relaxation everywhere and still loses to the landscape. Diagnosis: the classic hard phase of planted assignment -- the gate shows the configuration is information-theoretically identifiable while local dynamics equilibrate into an exponential shelf of wrong states. The gap between the statistical and algorithmic thresholds is now the central quantity. The N-scaling probe closes the cheapest hypothesis: at matched per-node move budget (N=2,048, 320k rounds), the reachable shelf returns to the same +9--10% band as N=512 and the small accuracy signal vanishes. Density alone does not approach an algorithmic threshold in the tested range. The first C5 anchor wave closes the second-cheapest. Pixel- and lexicon-derived unary anchors carry population signal (combined z = 17.2) with no per-pair reliability (top-margin precision 2%, mutual-NN 0.17%), and dense injection into the tempering energy degrades recovery monotonically in the weight (accuracy 3.1% to 1.0--2.0%, relational energy +9% to +30--115%). Scene-level anchors are coarse-redundant with the shared subspace; usable side information must be per-pair reliable or independent of the relational field. Fine-grained anchors (numerals, spatial predicates, box geometry) live at view level with sixteen candidates, so the second wave merges into the view recursion. Remaining search-side lever: cavity/message-passing dynamics. The second wave (2026-07-30) delivers the first seeds. Per-view color, light, size, and position anchors against crop pixels and box geometry: frame-free view matching 11.5% vs 6.25% chance, 35.6% precision on top-1%-margin views (354 seeds; every prior high-confidence channel was 0--4%). Aggregated per-pair Hungarian values give the best frame-free node affinity to date (R@10 4.4--4.9x chance over three seeds), with node-level seed precision still at 2--6%. Pixel and lexicon statistics are exhausted; the next package is a factor graph over node and view assignments -- anchor unaries, relational pairwise, message passing -- plus the untested numeral-to-repetition anchor. See `RECOVERY_RESULTS.md`. ## 2026-07-30: synthetic closed world v0 -- the objective becomes the bottleneck A procedural world with exact closure (43-word vocabulary, unique referents, ring-placement multiplicity, relation-constrained rendering, ground-truth interventions) and from-scratch unimodal towers on disjoint splits. First measurement inverts the natural-data pattern: linear CKA matches (0.278 vs 0.262) while relational Spearman collapses six-fold (0.030 vs 0.184, z 2.4 vs 12.3). The gate fails accordingly; content projection restores ranking (z = 117) but not the basin (counterfeit at -13%). Attribution is clean because data sufficiency is total by construction: InfoNCE's uniformity pressure turned the vision tower into a near-orthogonal instance codebook, erasing between-scene geometry. Sufficiency of the data does not transfer to sufficiency of the representation; the unimodal objective decides which world geometry survives, and R1 is a property of the data-objective pair. The SSL objective is now a phase-diagram axis. Next: reconstruction-family towers (doubling as the masked predictors of the prediction-transfer energy), then the structure identification battery. See `SYNTH_RESULTS.md` and the 2026-07-30 additions to `STRUCTURE_DESIGN.md` (unified element-relation object; the three-rung energy ladder ending in lifted unimodal world models). ## 2026-07-30/31: objective battery, set machinery, and the first phase coordinates The synthetic world's second day closes three questions and opens one. Objective battery, complete: five standard self-supervised recipes (orbit-InfoNCE, SimMIM, their hybrid, data2vec-lite, slot attention at two resolutions) all fail to produce factor-complete global states, each by a distinct mechanism -- shortcut selection, local-only knowledge, non-merging losses, softened shortcut, and spatial tiling instead of object binding. Slot attention carries the most relational sharing (0.128 against 0.030 for InfoNCE) and the set-versus-pool doctrine is confirmed quantitatively: pooling slot sets without correspondence drops color decodability from 0.99 to 0.55. Set-kernel machinery, taken to a hand-crafted ceiling (connected-component descriptors against phrase bag-of-words): seven measured iterations moved the gate from z = -0.1 and 45% improving swaps to z = 51, 0.15% improving swaps, and 77% descent retention. Priced en route: mass-weighting inside matching destroys grading; multi-view field averaging cancels segmentation noise; a factor present on one side only (shape) is a counterfeit lever; so is size-dependent matching bias. End to end, honestly: blind tempering on the best fields still finds assignments 4--9% below the truth at chance accuracy, and size residualization removes shared signal along with the nuisance. With VG's passing configuration at field validity 0.61 and this ceiling near 0.45, the basin-pass boundary at N = 512 is bracketed between them: the first coordinates of the identifiability phase diagram. See `SYNTH_RESULTS.md`. ## 2026-07-31: the Fock lift and the (validity, N) boundary The out-of-the-box program's first invention went to flight the same morning. A symmetric-tensor moment kernel on object sets (the truncated second-quantized lift: composition as algebra, multiplicity as an observable, no matching step and hence none of the measured matching biases) with factor-mirrored one-hot descriptors gives the synthetic world its first full basin pass at N = 64 -- no counterfeit, 92% retention. At N = 512 counterfeits return 18--22% below the truth and recovery stays at chance; residual segmentation error (30% wrong counts) and shape-cluster impurity suffice at eight times the configuration space. Identifiability is now measured as a joint boundary in field validity and population size: pass at (0.45, 64), fail at (0.45, 512), pass at (0.61, 512). Queued inventions, in order: orbit-derived per-entry precision weighting (channel-matched energy), third-order triangle invariants (holonomy against pairwise counterfeits), moment ideals as gauge-free content coordinates; mixed-curvature products noted as a symmetry-breaking device and deprioritized. See `SYNTH_RESULTS.md` and the tool inventory in `STRUCTURE_DESIGN.md`. ## 2026-07-31: the gate statistic was measuring the searcher A methodological correction that supersedes part of the same day's optimism. Third-order (triangle) invariants were added and, in a harness whose descent proposes 64 sampled pairs per step, the truth became essentially immovable at N=512 (retention 99.6%, no counterfeit, z=50), replicated over three hidden shuffles. Two controls dismantled the reading: the pairwise-only energy passes identically in that harness (retention 100%), and long tempering on the triangle energy at N=128 reaches 10% below the truth at zero accuracy. Conclusion: descent retention measures the strength of the search operator, not the identifiability of the energy. Exact all-pairs enumeration (yesterday's harness) is a strictly stronger searcher than sampled proposals, and long tempering is stronger still. The gate statistic is therefore redefined: \[ E(\text{truth}) \le E(\text{deepest state a strong searcher reaches}), \] measured with long parallel tempering from random starts plus a truth-initialized arm, uniformly across candidate energies. Retention and improving-swap fractions remain as diagnostics, never as verdicts. All candidate energies are being re-measured on this basis (`synth_deep_gate.py`); readings from the weaker harness are marked superseded in `SYNTH_RESULTS.md`. ## 2026-07-31: closed form, spectral solvers, and a clean Tier 0 Three results and one protocol decision, all same day. Closed form: the triangle energy equals a constant minus 2 trace(M^3)/6 with M the elementwise product of permuted text and visual fields, so one batched matrix product scores hundreds of permutations over all C(N,3) triples. Validated to 3e-7 against brute force; opens trace(M^k) for any k. Iteration time per configuration falls from hours to minutes. Spectral solvers (Umeyama, GRAMPA) were added because they are polynomial and dimension-free -- relation fields are N x N whatever the embedding widths -- and therefore bypass the glassy landscape. They return 3--13x chance. Correlated-graph-matching theory explains it: at N=256 the information-theoretic threshold sits near rho = 0.29 and the polynomial threshold near rho = 0.9, and our fields are at 0.51--0.57. The project is squarely in the computationally hard phase, which now has a quantitative account rather than an empirical one. Identifiability is not correlation. The colour-only field reaches rho = 0.863 yet caps at 49% because only 126 of 256 scenes have distinct colour multisets; its eigenvalue degeneracy is that collision structure. Full descriptors are 100% distinct. Channel mixing inside one moment kernel dilutes 0.86 to 0.51. Protocol, decided with the user: anchors that declare "red means this hue" are not labels but are hand-supplied cross-modal prior, so results must be reported by tier. Tier 0 forbids any declared correspondence. Achieved the same day on world v1 (Zipf-skewed factor marginals, since uniform v0 makes the correspondence information-theoretically unrecoverable): text factor families discovered by mutual exclusivity, vision colour classes by k-means on raw RGB, the two paired by marginal frequency rank -- 10/10 entries semantically correct, fields at rho = 0.570 with a 97.3% identifiability ceiling. See `SYNTH_RESULTS.md`. ## 2026-07-31: first recovery, and it was segmentation all along The rho decomposition settled the target: oracle segmentation gives field correlation exactly 1.0 and the Tier 0 derived dictionary is free (oracle and derived fields are identical), while connected-component segmentation gives 0.5696. Three unimodal fixes closed most of it -- Gestalt appearance grouping (exact group count 71.9 to 90.2%), distance-transform watershed for touching ring members (exact object count 74.2 to 100%), and one-dimensional k-means on radial extent for size classes (52.7 to 87.9%, because area confounds size with shape and the classes are gap-separated rather than equally populated). Field correlation went 0.570 to 0.671 to 0.928, and GRAMPA recovery followed it 5.1% to 9.4% to **53.9%** at N=256 (0.39% chance), replicated at 53.9/53.9/54.7% over three hidden shuffles, and 26.3% at N=600 (0.17% chance). First end-to-end world matching in the project: no pairs, no declared dictionary, no learned cross-modal map -- factor families from mutual exclusivity, colour correspondence from marginal frequency rank, object states from pixels, assignment from a polynomial spectral solver on two independently built relation fields. Composing the two solvers closes most of the remaining gap: GRAMPA initialisation 53.5%, plus pairwise-energy steepest descent 86.7%, plus pairwise-and-triangle refinement **95.3%**. Spectral never gets trapped but rounds badly; the energy rounds well but cannot find the basin; the composition does both. The final state sits 0.04% below the truth in energy at 95.3% accuracy -- residual field noise, not a counterfeit. The theoretical account held throughout: nothing worked below rho = 0.9 and recovery appeared as soon as the fields crossed it. Reporting-bias worlds show the derivation degrades gracefully -- text-versus-pixel rank correlation 1.00/0.98/0.84/0.77 gives dictionaries of 8/8/6/6 out of ten -- with errors concentrated in the close-frequency tail, so margin-calibrated entries are the natural seed set. The recovered assignment then transfers: used as pseudo-pairs for a ridge map and applied to 200 held-out scenes that took no part in matching, retrieval is 93.0% exact against 0.5% chance, colour set F1 0.989, count multiset 97.5%. The random-pair control sits exactly at chance and the shuffled-image control drops colour F1 to 0.295, so the output is image-conditioned and extrapolates. Rungs four and five of the evidence ladder are closed inside the synthetic world; matching is a bootstrap for the map, not the deployment mechanism. Unchanged by this: object states are hand-built upper-bound descriptors that no tested self-supervised recipe reproduces, and the world is procedurally generated. See `SYNTH_RESULTS.md`. ## 2026-08-01: reporting bias, and three solver-versus-information calls Worlds v3 and v4 test whether the derived dictionary survives reporting bias. v3 couples colour to shape cyclically and biases mentions toward rare colours, so captions name half the groups and the pixel and text colour rankings diverge in the tail; v4 replaces the cyclic coupling with a distinct Dirichlet shape profile per colour. v3 results: marginal rank 4/10, joint structure 6/10, joint plus marginal under exact enumeration of the 720 shape permutations 7/10, oracle shape columns with measured clusters 8/10, oracle labels 10/10. Vision-side extraction is not the limit (colour cluster purity 1.000, shape 0.998). The 6/10 initially looked like an information bound -- ten colours over six shapes leaves exactly four twin profiles, and exactly four appeared -- but the twins separate by frequency, and the combined evidence has minimum pairwise distance 0.2322 against 0.0008 for shape alone. The bound is 10/10; what is left is one entry of solver gap and two of estimation noise from the thinned joint table. v4 results: exact enumeration recovers 10/10 under the same bias and mention rate, across frequency weights including zero. Non-degenerate joint structure suffices on its own, which is the regime real corpora occupy. Three times now a limit has looked intrinsic and proved to be the solver's: descent retention, the triangle-energy pass, and this ceiling. The rule stands -- compute the bound, do not infer it. See `SYNTH_RESULTS.md`. ## 2026-08-01: factor families on real language First step of the port to natural data, text side only, on 80,000 Visual Genome region descriptions with no lexicon and no labels. The synthetic pipeline found families by mutual exclusivity: template phrases carry exactly one colour word, so colour terms never co-occur. That does not transfer. In real descriptions colour words are statistically independent or better -- observed over expected co-occurrence is 1.29 for black with white, 0.93 for red with blue -- while head nouns are the genuinely exclusive family (0.14 for car with truck). Run as written, the method recovers clean noun classes and no attribute families at all. Mutual exclusivity is a template artifact. Standard distributional word-class induction, of which exclusivity is a degenerate case, does transfer. Positive PMI context vectors, truncated SVD, k-means over the 300 most frequent words yields a colour family at 0.67 purity (red, brown, gray, dark, grey, pink, tan, purple) plus coherent object categories: vehicles, animals, clothing, people, water scenes. Colour terms split into three clusters rather than one, and the split is structural rather than noisy -- white and blue cluster with sky and clouds, because what a colour word modifies shapes its distribution. That dependence is the signal cross-modal value alignment feeds on. Next: the vision side under unsupervised segmentation, then the field correlation as the go/no-go before any recovery run. ## 2026-08-01: first natural-data coordinate, and a correction The Tier 0 recipe was ported to Visual Genome: DINOv2 patch features, spectral segmentation of the affinity graph, segment feature classes on one side; distributional word classes on the other. Field correlation at the hidden pairing is **0.166**, far under the 0.9 that recovery needs, so by our own go/no-go no recovery run is warranted yet. A correction came with it. Aligning vision classes to text classes by frequency rank and by an oracle give bit-identical fields -- 0.1662 both ways -- while agreeing on 0 of 23 classes. Moment fields are built from within-modality inner products, and permuting feature coordinates consistently leaves every inner product unchanged, so relation fields are gauge-invariant in the class labels and the dictionary never enters them. The same was true unnoticed in the synthetic world, where oracle and derived dictionaries produced identical fields. The dictionary is required for encoding a new image into text coordinates, not for matching; matching has always run on the fields alone. The 0.166 therefore reads as representation quality, and localises the deficit at vision-side granularity: six segments per image, background included, against sixteen phrases about specific objects. A sixteen- segment run is measuring whether granularity is the binding constraint. ## 2026-08-01: natural data reaches 0.656, and segmentation was not the wall Full port to Visual Genome, measured by the go/no-go statistic. Three results overturned expectations, each checked rather than inferred. Discretisation was the dominant loss: hard class codes give 0.166, continuous states 0.489. Relation fields need scene similarity only, so quantising states first is pure attrition; the synthetic world hid this because its states are discrete by construction. Segmentation is not the bottleneck on photographs. Oracle region boxes buy 0.061 over unsupervised spectral segmentation of DINOv2 patches (0.550 against 0.489), the inverse of the synthetic world where segmentation was the whole gap. Flat untextured scenes are adversarial for object discovery in a way photographs are not. Content projection is again the largest lever: 0.489 to **0.656**, matching its earlier role at node level (0.21 to 0.61). Protocol-legal and oracle-box configurations land within 0.007, so the residual is not about object localisation. Control: frozen general-purpose encoders underperform corpus-fitted statistics (0.19 for DINOv2 crop CLS against Qwen phrase states), so in-context features and corpus-specific distributional vectors are the better substrate. Augmentation orbits were then tried and give nothing: four random resized crops per image, segmented independently with fields averaged, move the correlation from 0.6559 to 0.6559. Re-renders in the closed world resample layout, which is nuisance by construction; random crops perturb framing, which self-supervised patch features already absorb, so the average has nothing to cancel. Multi-view observation means separate photographs, not transformations of one. Verdict: 0.656 against the 0.9 recovery needs; no recovery run, and the orbit lever is unavailable on this corpus. See `NATURAL_RESULTS.md`. ## 2026-08-01: landscape reshaping does not substitute for correlation At the natural-data field quality of 0.656, every solver returns chance. Full-rank spectral initialisation gives 0.0039 against 0.0039 chance, refinement 0.0078, and a rank-truncation ladder from four to full rank stays at or below chance throughout. Coarse-to-sharp smoothing is the standard way to widen basins and thin decoys, and it recovers nothing here. The reading is categorical rather than about this one method. Below the polynomial threshold the deficit is information the algorithm class cannot exploit, not a basin it fails to locate, so reshaping has nothing to work with. Solver-side candidates should therefore wait on the correlation rather than compete with it, and the only non-correlation lever still open is population size, since the phase boundary is joint in the two. Compute note: NCSA Delta A40 was brought in for the correlation-raising battery -- larger backbones and finer segmentation over the same corpus -- with three array tasks starting immediately against a queue of 810, 9,624 GPU-hours remaining on the account. ## 2026-08-01: shrinking the population is backwards, and the truth is not the optimum Population size was the last non-correlation lever, on the reasoning that the hard phase is joint in correlation and size, so small blocks would need less correlation. The sign is the other way. The information-theoretic threshold is ρ_IT ≈ √(4 log N / N), which **rises** as N falls: 0.29 at 256 scenes, 0.51 at 64, 0.83 at 16. The gate settles it directly rather than by citation. At N = 16, 24, 32, 48, 64 and 96, the deepest state a strong searcher reaches — spectral start plus 200 random restarts, each run to a local optimum under exact steepest descent — is deeper than the truth, in three replicates out of three at every size. Where the truth is not the optimum, no searcher of any cost finds it and the algorithm class stops mattering, which also disposes of "accept exponential cost". Route D closes. ## 2026-08-01: the correlation was never the statistic that governs recovery The project's central gate — correlation at truth must reach about 0.9, price any representation in minutes without running a search — is wrong, and a controlled truncation shows it. Project the synthetic fields that recover to rank *r* and match blind. Correlation stays at 0.902, 0.906, 0.928 across ranks 4, 8, 16 while recovery goes 6.2%, 12.9%, **95.6%**. A field past the supposed threshold recovers 13%. We misread the theory rather than being failed by it. Matching thresholds assume exchangeable full-rank noise, where each of N²/2 entries constrains the pairing independently; a rank-*r* shared component supplies about *rN*, an eighth as many at 256 scenes with *r* near ten. The gate becomes a joint condition on correlation and on the width of the shared spectrum, measured by principal angles between the two fields' leading eigenspaces against a scene-shuffled null. Neither suffices alone: 18 shared directions at correlation 0.508 fails, 11 at 0.902 fails. With the old gate retired the excuse for never searching natural data went with it. Five blind trials on the best natural field recover **0.0000 against 0.0039 chance** — refinement moves away from the truth, which is what descent does when the truth is not the optimum. The old verdict was right; its reasoning was not. Diagnosis: on photographs both modalities are individually rich — vision effective rank 40, text 48 — and agree on 15 directions. Each side is rich about something and they are rich about different things. The intersection barely responds to anything on the vision side: 6 to 16 segments costs one direction, annotated boxes buy one, third-order moments in the set kernel raise the vision field's own rank by seven and the shared count by one. The text side moves it from 10 to 17 and saturates; more scenes do not help (23 to 19 from 256 to 1024). Free gains along the way: eigenvalue-weighted projection over a wide basis plus 128-dimensional text vectors take the correlation from 0.656 to 0.716 and the shared count from 10 to 16. Hubness was refuted as an inflation hypothesis — the additive degree model is 1–5% of field variance and removing it slightly *raises* the correlation. See `RANK_RESULTS.md`. ## 2026-08-01: two operational lessons **Never `scancel -u` on Delta.** Clearing three failed array tasks that way also killed four DAGFormer H200 jobs queued nineteen minutes earlier on the same shared account. Restored from `sacct`'s recorded WorkDir and SubmitLine, losing queue position only. Rule and recovery recipe written into `DELTA_HPC.md`. **The vision extraction was CPU-bound on its own eigendecomposition,** one dense symmetric problem of side grid² per image, costing more than the forward pass that produced the features. Moving it to the accelerator already holding the model took the batch time from 130 s to 1.9 s, a 68× speedup that makes the whole backbone battery cheap rather than overnight. ## Current experimental gate The node-level ordering-and-basin gate is passed on VG by content-projected states (isolated and in-context composite). The standing rules extend rather than retire: 1. Blind recovery next: unpaired soft-coupling optimization (Sinkhorn over real states, preserving the decodable path) must place calibrated coupling mass on the hidden truth from scratch. Its score is coupling mass at truth and verified seed precision, never retrieval means alone. 2. The view-level gate stands unchanged for the coarse-to-fine recursion: static view fields failed it; in-context view states plus projection are the next candidate, with replayed view truth evaluation-only. 3. Any new energy or state family still passes ordering before descent, including single-transposition local ordering. 4. Flickr-style single-view data cannot host the current pass; two-sided multi-view observation is a protocol requirement for now.