diff options
Diffstat (limited to 'SYNTH_RESULTS.md')
| -rw-r--r-- | SYNTH_RESULTS.md | 539 |
1 files changed, 539 insertions, 0 deletions
diff --git a/SYNTH_RESULTS.md b/SYNTH_RESULTS.md new file mode 100644 index 0000000..0e6617f --- /dev/null +++ b/SYNTH_RESULTS.md @@ -0,0 +1,539 @@ +# Synthetic closed world v0: the objective becomes the bottleneck + +Run date: 2026-07-30 + +## Question + +The natural-data experiments could never verify representational +sufficiency: closure and over-completeness were properties of the data we +happened to have. The synthetic world makes them constructions. Scenes +are sets of object groups (count, size, color, shape) with spatial +relation constraints; captions mention exactly the discrete state +(vocabulary 43 words); renders realize it with resampled continuous +nuisance, so visual orbits share exactly the world state and differ +exactly by modality-private variation. 26,000 scenes, 56,000 renders, +3,578 ground-truth intervention variants; vision-only and text-only +splits disjoint; every dial is a generator argument. + +Both towers are trained from scratch on their own modality: a compact ViT +(5.3M parameters) with InfoNCE over natural orbit positives -- no flips, +which would erase left/right, and no color jitter, which is world content +-- and a word-level causal LM (5.6M parameters) to cross-entropy 0.42, +near the floor of the templated language. Closure is verified visually +and referentially (color-shape pairs unique per scene, exact multiplicity +by ring placement). + +## Result: perfect closure, collapsed relational sharing + +On 1,000 held-out pairs, evaluation-only: + +| Quantity | Synth v0 | Natural (Flickr, frozen) | +|---|---:|---:| +| Linear CKA | 0.278 | 0.262 | +| Pairwise relational Spearman | 0.030 | 0.184 | +| Shuffle z | 2.4 | 12.3 | + +The closed world shares global statistics as strongly as natural data and +relational structure six times more weakly. The assignment gate agrees: +the true configuration keeps 6.6% of nodes under descent with a +counterfeit 26% below it, and content projection -- which on VG passed +the basin audit -- rescues ranking power (z = 117, improving swaps 0.35%) +but not the basin (retention 24.4%, counterfeit at -13%). + +## The mechanism is the objective, not the data + +InfoNCE with a uniformity term optimizes scene discrimination: every +scene is pushed apart from every other, and two scenes sharing "three red +circles" gain nothing from sharing it. The vision tower converged to a +near-orthogonal instance codebook whose between-scene geometry is +arbitrary -- exactly the structure the relational energy needs and +exactly what the objective erases. Data sufficiency was total by +construction, so the deficit is attributable: **sufficiency of the data +does not transfer to sufficiency of the representation; the training +objective decides which world geometry survives.** The requirement R1 of +`STRUCTURE_DESIGN.md` is a property of the data-objective pair. + +For the project's central question this is a constructive turn: alignment +without pairs needs unimodal objectives whose representations retain +compositional world geometry, and the synthetic world measures exactly +that retention. The SSL objective joins vocabulary, ontology, scene +count, and orbit multiplicity as a phase-diagram axis. + +## Objective battery: the two canonical objectives fail in opposite ways + +Ground-truth factor probes (linear, fitted per modality on its own split) +localize what each representation retains. The text tower is nearly +lossless and is not the bottleneck. The vision towers: + +| Probe target | InfoNCE (CLS) | SimMIM (patch mean) | SimMIM (foreground-weighted) | Text tower | +|---|---:|---:|---:|---:| +| Color presence | 0.992 | 0.572 | 0.717 | 0.997 | +| Shape presence | 0.540 | 0.535 | 0.645 | 0.992 | +| Group count R2 | 0.61 | 0.12 | 0.11 | 0.94 | +| Object total R2 | 0.44 | 0.33 | 0.35 | 0.90 | +| Relation count R2 | 0.20 | 0.03 | 0.04 | 0.94 | + +InfoNCE performs shortcut feature selection: color multisets are nearly +unique per scene, so discrimination is solvable from color alone and +shape is discarded at chance level. SimMIM inpaints well (reconstruction +MSE 0.0088) and pools to almost nothing: the scene is 80% background, the +information lives distributed in patch tokens, and no simple pooling +recovers it (CKA drops to 0.114, relational z to 0.8; the gate collapses +to 0.2% retention). Discriminative-global objectives select shortcuts; +reconstructive-local objectives never build a global summary. + +The naive hybrid (orbit-InfoNCE on the CLS plus masked pixel +reconstruction on the patches, weight 25) is falsified: color stays at +0.983, shape stays at chance (0.532), relational Spearman turns negative +(-0.063, z = -3.9), and the gate keeps 4.3%. The two losses settle +separately -- the CLS keeps the color shortcut while shape remains +distributed in patch tokens. Joint pressure on different tokens does not +merge into one summary. + +A first prediction-transfer energy was built and gated on the weak +towers: the vision substitution kernel from masked inpainting compared by +foreground-patch Chamfer distance, the text kernel from +relation-sentence continuation likelihood (relation sentences force +referential dependence; whole-caption continuations only reward +repetition, and the LM separates same-scene from different-scene +continuations at two-to-one per-token likelihood but grades partial +overlap poorly). The gated energy reaches z = 2.9 at 512 scenes with 31% +improving swaps: falsified as built. Every energy constructed today +bottlenecks at the same place -- the text side is near-lossless and the +vision side is factor-blind, so the vision objective remains the single +point of failure. + +## The objective battery, complete: five recipes, five distinct failures + +| Vision objective | Color | Shape | Groups R2 | Relational rho (z) | Failure mode | +|---|---:|---:|---:|---:|---| +| InfoNCE (orbit positives) | 0.992 | 0.540 | 0.61 | 0.030 (2.4) | shortcut selection: color only | +| SimMIM (pixel targets) | 0.572 | 0.535 | 0.12 | 0.024 (0.8) | local knowledge, no global summary | +| Hybrid (InfoNCE + SimMIM) | 0.983 | 0.532 | -- | -0.063 (-3.9) | losses settle separately | +| data2vec-lite (EMA feature targets) | 0.743 | 0.545 | 0.18 | 0.059 (3.7) | same, softer | +| Slot attention (5-6 slots) | 0.549 | 0.618 | 0.48 | 0.128 (7.9) | slots tile space, not objects | + +No standard self-supervised recipe at this scale produces +factor-complete global states. Slot attention comes closest structurally +-- best shape decodability, four times the relational sharing of InfoNCE +-- and fails by degenerating into near-uniform spatial tiles (alpha +masses 0.13--0.20 across six slots where object binding would put 0.7+ +on background), a known failure on flat untextured scenes. Its pooled +readout also scrambles per-object factors, which is the set-versus-pool +doctrine restated: color drops from 0.99 (InfoNCE pooled) to 0.55 the +moment states are sets averaged without correspondence. + +## The set-kernel machinery, taken to its ceiling + +With no learned tower delivering object states, model-free +connected-component analysis provides them (the background is flat, so +components are objects; ring merging at distance 30 restores +multiplicity at 70% count accuracy). Text sets are per-group phrase +bag-of-words vectors. Scene-to-scene relations are unweighted Hungarian +matching values; the gate runs on these set-kernel fields. Iterations, +each fixing one measured defect: + +| Iteration | z | Improving swaps | Kept | Counterfeit margin | +|---|---:|---:|---:|---:| +| Slot-vector sets | -0.1 | 45% | 3% | deep | +| Decoded sprite sets | 2.7 | 32% | 6% | deep | +| CC pixels x LM phrases | 1.8 | 32% | 5% | deep | +| CC descriptors x BOW | 4.5 | 13% | 9% | deep | +| + unweighted matching | 52.0 | 0.73% | 37% | -11% | +| + multi-view field average | 59.9 | 0.13% | 69% | -6.8% | +| + shape descriptors | 51.2 | 0.15% | 77% | -4.1% | + +Three lessons priced along the way: mass-weighting inside the matching +destroys similarity grading (0.16 to 0.45 against soft truth); averaging +per-view fields cancels segmentation errors (the orbit dividend again); +and a missing factor on one side (shape, absent from the first visual +descriptor while present in every text set) is exactly the kind of +asymmetry counterfeits exploit. + +## End to end: close, and honestly not there + +The best configuration holds 77% of the truth under exact descent, and +blind tempering still finds assignments 4--6% below the true energy at +chance accuracy. One further asymmetry was identified -- matching values +correlate with set sizes differently per modality (0.29 text, 0.13 +vision) -- and residualizing both fields against size features does not +close the gap either: tempering ends 9.2% below the truth at 0.78% +accuracy, because the correction removes shared signal along with the +nuisance. Each counterfeit so far has pointed to a specific asymmetry +between the two fields; the remaining ones (segmentation error, +descriptor-versus-word semantics) resisted this window's budget. + +The comparison across datasets brackets the identifiability threshold. +On VG, both-sides content projection reached cross-modal field validity +near 0.61 and the basin audit passed with margins; here the upper-bound +fields sit near 0.45 and every recovery run counterfeits. At N = 512, +the pass boundary for this energy family lies somewhere between those +two numbers -- the first coordinates on the phase diagram this world was +built to draw. + +## The Fock lift: matching-free set kernels pass at 64 and counterfeit at 512 + +Replacing the Hungarian matching value with a symmetric-tensor moment +kernel -- phi(S) concatenates the degree-1 and degree-2 moments of the +object set, scenes compare by normalized inner product -- removes the +matching step and with it every measured matching bias, while making +composition algebraic and multiplicity an observable. Paired with +factor-mirrored one-hot descriptors (color, small/large with medium +unmarked as in text, count, k-means shape cluster; size binned after +residualizing the renderer's count-dependent radius shrink): + +- N = 64: the first full basin pass of the synthetic world -- no + counterfeit in any quench, best random descent 11% above the truth, + 92% retention. +- N = 512: retention 75--76% and z = 38--45, but counterfeits return at + 18--22% below the truth, and blind recovery stays at chance. The + remaining asymmetries -- 30% segmentation count errors and shape-cluster + impurity -- are enough at eight times the configuration space. + +The bracket sharpens: this energy family passes at N = 64 and fails at +N = 512 on fields of validity near 0.45, while VG passed at N = 512 near +0.61. Identifiability is a joint (validity, N) boundary, and the world +now measures both axes. + +## Third-order invariants, and the correction they forced + +Triangle invariants (for a node triple, the product of the three edge +values, compared across modalities) were added to the moment-kernel +fields. Under a descent harness that proposes 64 sampled pairs per step, +the result looked decisive: at N=512 the truth kept 99.6--100% of its +assignment with no counterfeit below it and z = 48--50, replicated over +three hidden shuffles. + +Two controls dismantled that reading. The pairwise-only energy passes +identically in the same harness (retention 100%, z = 40), so the third +order was not what changed; and long tempering on the triangle energy at +N=128 reaches 10.3% below the truth at zero accuracy. What differed from +the earlier failing runs was the searcher: exact all-pairs enumeration +takes the globally best swap each step, sampled proposals do not, and +tempering with a long budget beats both. + +**Descent retention measures the search operator, not the energy.** The +gate statistic is corrected to compare the energy of the truth against +the deepest state a strong searcher reaches, and every candidate energy +is being re-measured on that basis. The ablation does carry one clean +finding that survives: triangle-only has 11.3% improving transpositions +against 0.63% for the sum, so the two orders are complementary -- +third-order terms deepen the basin while second-order terms smooth the +local landscape. + +## Precision weighting, tried and instructive + +The channel-matching idea was tested and fails with a clean mechanism: +weighting relation entries by orbit-derived precision (inverse variance +across re-rendered views) makes N=64 worse than the unweighted kernel +(counterfeit 46% below the truth; descent from the truth keeps 56%). +Orbit variance measures reproducibility, not correctness -- the channel +splits into variance and bias, view resampling exposes only the variance +half, and systematically biased entries are the most reproducible ones, +so precision weighting amplifies exactly the entries counterfeits live +in. A full Nishimori match needs the cross-modal residual model, which +requires pairs. The legal half-channel is now measured to be the wrong +half. + +## Closed form, and what it revealed + +The triangle energy has a closed form. Writing M for the elementwise +product of the permuted text field with the visual field, the pairwise +term is a constant minus 2 sum(M) and the all-triple term is a constant +minus 2 trace(M^3)/6, because traces of powers are invariant under +simultaneous row-column permutation. One batched matrix product scores +hundreds of candidate permutations over every C(N,3) triple, replacing a +200,000-triple sample and a 5.7 ms per-proposal Python loop. Validated +against brute force to 3e-7. The same identity opens the whole family +trace(M^k): fourth and fifth order closed loops now cost one more matrix +product each. + +Three diagnostics followed, each changing what we thought we knew. + +**Descent retention measures the searcher.** Sampled-proposal descent +keeps the truth for every candidate energy at N=512; exact all-pairs +enumeration is stronger; long tempering is stronger still and reaches +10% below the truth. The gate statistic is corrected to compare the +truth against the deepest state a strong searcher reaches. + +**Spectral solvers place us in the hard phase.** Umeyama and GRAMPA run +in polynomial time on N x N fields and need no common embedding +dimension, so they sidestep the glassy landscape entirely -- and they +return 3--13x chance, never recovery. Correlated-matching theory says +so: at N=256 the information-theoretic threshold is near rho = 0.29 and +the polynomial-algorithm threshold near rho = 0.9, and our fields sit at +rho = 0.51--0.57. Identifiable, not efficiently recoverable, exactly as +observed. (Our fields are structured rather than Wigner, so the theory +guides rather than proves.) + +**Correlation and identifiability are different axes.** The colour +channel alone reaches rho = 0.863, far above the mixed descriptor, but +its scenes collide: only 126 distinct colour multisets among 256 scenes, +capping any matcher at 49% and collapsing to 25% at N=1000. The +eigenvalue degeneracy that breaks spectral methods is that collision +structure in spectral form. Full descriptors are 100% distinct, so the +information is present; mixing strong and weak channels into one moment +kernel is what dilutes rho from 0.86 to 0.51. + +## Tier 0: the cross-modal dictionary, derived rather than declared + +Anchors that map "red" to a hue band are not labels -- no pair is ever +consulted -- but they are hand-supplied cross-modal prior. Tier 0 +forbids them, and the derivation splits by factor type: ordered factors +(count, size) match by their intrinsic order with the residual ambiguity +enumerated; unordered factors (colour, shape) can only be matched by +distributional statistics. + +That exposed a defect in world v0: factor values were sampled uniformly, +so colour marginals and co-occurrences are symmetric and the +correspondence is information-theoretically unrecoverable. Real corpora +are the opposite, and this is precisely the structure unsupervised word +translation runs on. World v1 samples colour and shape with a Zipf +exponent of 1 (colour marginals 0.317 down to 0.035), which is both more +realistic and sufficient. + +On v1 the derivation runs entirely on unimodal statistics of disjoint +splits: + +1. text families emerge from mutual exclusivity -- tokens of one factor + never co-occur inside a group phrase, so a greedy pass recovers + {a, two, three, four}, the ten colour words, and {small, large} with + no declared list; +2. vision colour classes come from k-means on raw component RGB, with no + hue rule; +3. the two sides pair by marginal frequency rank. + +**The derived dictionary is 10/10 semantically correct.** Tier 0 fields +then reach rho = 0.570 -- above the v0 declared-dictionary Tier 1 value +of 0.508 -- with a 97.3% identifiability ceiling, and GRAMPA rises to +13x chance at eta = 1. Zero declared cross-modal information. + +## First recovery: 54% at Tier 0 + +The rho decomposition made the target unambiguous. With oracle +segmentation the field correlation is exactly 1.0 and the derived +dictionary costs nothing (oracle and derived give identical fields); with +connected-component segmentation it is 0.5696. **The entire gap was +segmentation.** + +Three unimodal fixes, each measured: + +| Fix | Measure | Before | After | +|---|---|---:|---:| +| Gestalt appearance grouping | exact group count | 71.9% | 90.2% | +| Distance-transform watershed | exact object count | 74.2% | 100% | +| One-dimensional k-means on radial extent | size class | 52.7% | 87.9% | + +The watershed matters because connected components merge touching ring +members: the renderer leaves a gap of 0.37 radii between neighbours, +under two pixels for small objects, so a quarter of all scenes lost +objects before any grouping logic ran. The size fix matters because area +confounds size with shape -- a cross of a given radius has half the area +of a circle -- so extent replaces area, and natural breaks replace +terciles because the size classes are separated by gaps rather than +equally populated in extent. + +Field correlation rises 0.5696 to 0.6708 to **0.9285**, crossing the +polynomial-recovery threshold, and GRAMPA follows it exactly: + +| Fields | rho | GRAMPA best | +|---|---:|---:| +| distance-merge | 0.570 | 5.1% | +| Gestalt grouping | 0.671 | 9.4% | +| watershed + extent sizes | **0.928** | **53.9%** | + +**Recovery is 53.9--54.7% exact scene identification across three hidden +shuffles, against 0.39% chance -- 138x.** No pairs, no declared +dictionary, no learned cross-modal map: the colour correspondence is +derived from marginal frequency rank, the factor families from mutual +exclusivity, the object states from pixels, and the assignment from a +polynomial spectral solver on two independently built relation fields. + +This is the first end-to-end world matching result in the project, and +it lands exactly where correlated-matching theory said it would: nothing +worked below rho = 0.9, and recovery appeared as soon as the fields +crossed it. + +The phase diagram now carries measured recovery rather than pass/fail: + +| rho | N | Recovery | Chance | +|---:|---:|---:|---:| +| 0.570 | 256 | 5.1% | 0.39% | +| 0.671 | 256 | 9.4% | 0.39% | +| 0.928 | 256 | **53.9%** | 0.39% | +| 0.926 | 600 | **26.3%** | 0.17% | + +Recovery falls with population size at fixed correlation, which is the +expected shape: the correlation a polynomial solver needs grows with N. +Both large-N and small-N points sit far above chance (158x and 138x), so +the boundary is a gradient rather than a cliff in this range. + +## Composing the two solvers: 95% + +The spectral solver and the energy are complementary in exactly the way +their failure modes suggest. Spectral reads the coarse correspondence out +of the eigenstructure with no search, so it never gets trapped, but its +rounding is noisy. Exact steepest descent on the closed-form energy fixes +rounding but cannot find the basin from a random start. Composed at +N=256: + +| Stage | Accuracy | +|---|---:| +| GRAMPA initialisation | 53.5% | +| plus pairwise-energy refinement | 86.7% | +| plus pairwise-and-triangle refinement | **95.3%** | + +Scaling is gentler than the spectral numbers alone suggest. At N=600 the +spectral initialisation halves to 26.2% but the composed solver only +falls to 84.7%, and there the refined energy sits *above* the truth +(0.7056 against 0.6623), meaning the truth remains the better +configuration and the solver simply stopped short -- a healthier state +than the marginal undercut seen at N=256. + +| N | Spectral | plus pair | plus pair and triangle | Refined energy vs truth | +|---:|---:|---:|---:|---| +| 256 | 53.9% | 87.5% | **96.1%** | 0.04% below | +| 600 | 26.2% | 80.0% | **84.7%** | 6.5% above | + +The final configuration at N=256 has energy 0.64738 against the truth's 0.64981 -- +below it by 0.04% at 95.3% accuracy. That is a different creature from +the counterfeits of the previous weeks: not a wrong configuration +exploiting a blind spot, but residual field noise making a nearly correct +assignment marginally preferable. The remaining 5% is measurement error +in the fields, not a failure of identifiability. + +**Blind recovery of 95.3% of a hidden image-caption correspondence, with +no pairs, no declared dictionary, and no learned cross-modal map.** + +## The recovered correspondence transfers to unseen scenes + +Matching is transductive, so the result only matters if the recovered +assignment can be spent. Treating it as pseudo-pair supervision for a +ridge map from vision states to text states, and applying that map to 200 +held-out scenes that took no part in matching: + +| Measure | Recovered pairs | Random-pair control | Shuffled-image control | +|---|---:|---:|---:| +| Held-out retrieval, exact | **93.0%** | 0.5% | -- | +| Colour set F1 | **0.989** | 0.399 | 0.295 | +| Count multiset exact | **97.5%** | 2.5% | 8.5% | +| Group count exact | **97.5%** | 17.0% | 58.0% | + +Chance retrieval is 0.5%. The random-pair control lands exactly on +chance, so nothing here comes from the map family or the state geometry; +the shuffled-image control collapses colour F1 from 0.989 to 0.295, so +the output is conditioned on the image rather than on a description +prior. + +This closes rungs four and five of the evidence ladder inside the +synthetic world: the output is image-conditioned, and the correspondence +extrapolates to scenes the matcher never saw. Matching is a bootstrap, +not the deployment mechanism -- once enough correct pairs exist, the map +carries the rest forward. + +## What the demonstration actually shows: matching verifies, statistics align + +The framing matters more than the numbers. Finding a permutation +presupposes a bijection between the two corpora, and real unpaired data +has none: image and text collections overlap only partially in the world +states they cover, one image answers to many descriptions or none, and +granularity differs. If the method needed that bijection it would not +transfer. + +It does not. The step that establishes cross-modal correspondence -- the +derived dictionary -- reads the text families from the 6,000 text-only +scenes and the colour classes from the 6,000 vision-only scenes, and +those two sets are disjoint. No instance appears on both sides. The +pairing comes from marginal frequency, a statistic of each corpus alone. + +The 95.3% matching and the 93.0% transfer are therefore verification, not +mechanism: they confirm at instance level that the factor coordinates +recovered from disjoint corpora are the right ones. On real data, where +no hidden truth exists to score, held-out generation against a +shuffled-image control replaces matching accuracy as the measurement. + +The conditions this leaves are weaker than pairing but not vacuous. The +two corpora must reflect a similar world distribution, since frequency +ranks are only comparable when both sides sample the same world; +reporting bias shifts the text marginal away from the visual one, which +the bias worlds show degrades the dictionary gracefully (rank correlation +1.00/0.98/0.84/0.77 gives 8/8/6/6 correct entries) rather than breaking +it; and co-occurrence structure, which real corpora have in abundance and +this world lacks by construction, is the bias-robust upgrade over bare +frequency. + +What remains open is unchanged and should not be blurred by this result: +the object states here are hand-built upper-bound descriptors, and no +self-supervised recipe tested has produced their equivalent; and the +world is procedurally generated. The demonstration establishes that the +matching machinery works end to end when the representations are good +enough, not that learning produces such representations. + +## Reporting bias, and the discipline of separating solver from information + +The frequency-rank dictionary depends on both corpora ranking values the +same way, which real corpora do not: text mentions the salient, pixels +count the common. Two worlds test it. World v3 adds colour-shape coupling +(each colour prefers one shape) and strong reporting bias, so captions +name only 50% of the groups present and the pixel and text colour +rankings diverge in the tail (brown moves from ninth to fourth, pink from +sixth to tenth). World v4 replaces the cyclic coupling with a distinct +Dirichlet shape profile per colour. + +On v3 the derivation degrades and the recovery of the cause matters more +than the number: + +| Evidence and solver | Correct | +|---|---:| +| Marginal frequency rank | 4/10 | +| Joint shape structure, alternating solver | 6/10 | +| Joint plus marginal, exact enumeration of shape permutations | 7/10 | +| Oracle shape columns, measured clusters | 8/10 | +| Oracle labels, all signals | 10/10 | + +The first reading was that 6/10 is an information-theoretic ceiling, +since four of ten vision colours have near-identical shape profiles -- +ten colours cycling over six shapes leaves 10 mod 6 = 4 duplicates, and +the twin count came out at exactly four. That reading was wrong, and +checking it rather than asserting it is the point. The twins differ +sharply in frequency; combining both signals separates all ten (minimum +pairwise evidence distance rises from 0.0008 to 0.2322), so the ceiling +is 10/10. What remains is a one-entry solver gap and a two-entry +estimation gap, the latter because reporting bias leaves common colours +with a quarter of their mentions and thins the joint table. + +On v4, where the coupling gives every colour a distinct shape profile, +exact enumeration recovers **10/10 under the same strong reporting bias +and 50% mention rate**, and does so across a range of frequency weights +including zero. Joint structure alone suffices when it is +non-degenerate, which is the situation real corpora present: colour and +object co-occurrence there is far richer than a cyclic preference. + +The pattern to carry forward is procedural. Three times in two days a +result looked like a property of the problem and turned out to be a +property of the solver: descent retention, the triangle-energy pass, and +this ceiling. Each was settled by computing the bound rather than +inferring it. + +## Next + +1. Push field validity: segmentation is the dominant residual noise + (70% exact group counts); texture-world slot binding (running) or a + supervised-free segmentation refinement would move validity directly. +2. Third-order triangle/holonomy invariants -- the remaining queued + invention: counterfeits to date satisfy pairwise statistics only, and + triangle tensors multiply the constraint count by N per node. +3. The objective-battery finding stands: no standard SSL recipe meets + the ceiling spec; binding fails on flat scenes at both resolutions + (sorted alpha masses 0.23--0.18 against uniform 0.20), and the + texture world tests whether within-object structure fixes it. + +## Main artifacts + +- `artifacts/synth_v0/manifest.json`, `scenes.private.json`, + `captions.json`, `interventions.private.json` +- `artifacts/synth_v0/{vision,text}_tower.pt` +- `artifacts/synth_v0/{vision,text,text_orbits}.pt` +- `artifacts/synth_v0/diagnostics.json` +- `artifacts/synth_v0/gate_test512_orbit.json` +- `artifacts/synth_v0/content_gate_k32.json` |