path: root/SYNTH_RESULTS.md

# 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.)

*Added 2026-08-01: that parenthesis was the whole story and we left it
as a parenthesis. Structured means low-rank, and the threshold formula
assumes full-rank noise, which is why the 0.9 figure later turned out to
predict nothing — see `RANK_RESULTS.md`. The caveat was correctly
identified and not followed up for months.*

**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`