File size: 7,318 Bytes
9497609
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
932f60c
9497609
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
932f60c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9497609
 
 
 
 
 
 
 
 
 
 
 
 
 
12aaf69
9497609
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
---
license: apache-2.0
language:
  - en
tags:
  - interpretability
  - probing
  - cross-modal
  - retrieval
  - negative-results
pretty_name: SRT depth-probe artifacts (which layer, and according to which probe)
---

# SRT depth-probe artifacts

**Where a frozen multimodal backbone carries cross-modal meaning depends on
what you use to look. A raw-cosine probe and a fitted linear probe rank the
layers of the same model differently, on the same states, in the same run.**

Every number here is measured on frozen `google/gemma-4-31B-it`. Nothing in
this repository is a trained product. These are measurement artifacts and the
scripts that produced them, published so the ranking can be checked instead
of taken on trust.

## The result

Nine layers, 5,000 COCO val2017 images, 4,000 fitted and 1,000 held out.
Chance is R@1 0.001 and median rank 500. "Cosine" is the head-free read-out.
"Fitted" is the shipped head recipe, two 1024-d towers under InfoNCE, fitted
per layer on the training split only.

| layer | cosine R@1 | cosine median | fitted R@1 | fitted median |
|---|---|---|---|---|
| L4  | 0.013 | 269 | 0.080 | 21 |
| L12 | 0.047 | 116 | 0.175 | 8 |
| **L20** | 0.100 | 76 | **0.293** | **4** |
| L28 | 0.015 | 329 | 0.172 | 9 |
| L36 | 0.068 | 127 | 0.240 | 5 |
| L44 | 0.137 | 24 | 0.254 | 5 |
| **L47** | **0.167** | **20** | 0.230 | 5 |
| L54 | 0.160 | 8 | 0.193 | 7 |
| L60 | 0.003 | 393 | 0.108 | 16 |

Cosine ranks L47 first. Fitted ranks it fourth, behind L20, L44 and L36.
The two probes disagree about the shape of the model.

Two things follow. A depth profile is a joint property of the backbone and
the read-out, so "layer *k* is the alignment peak" is only meaningful with
the probe named. And L60, which looks like total collapse under cosine
(R@1 0.003, median 393), still supports 108x chance under a fitted head. The
information survives, and it is the cosine read-out that loses reach.

## Scaling, and where it stops replicating

The L20 advantage grows with fitting data. On val2017, five seeds, the same
1,000 images held out:

| train images | 250 | 500 | 1,000 | 2,000 | 4,000 |
|---|---|---|---|---|---|
| L20 − L47 R@1 | +0.009 | +0.013 | +0.030 | +0.032 | +0.061 |

Replicated on a different image distribution, COCO train2017, three seeds,
2,000 held out:

| train images | L20 | L47 | gap |
|---|---|---|---|
| 2,000 | 0.145 ±0.004 | 0.130 ±0.005 | +0.015 |
| 4,000 | 0.192 ±0.009 | 0.180 ±0.011 | +0.012 |
| 8,000 | 0.245 ±0.005 | 0.219 ±0.003 | +0.026 |
| 16,000 | 0.290 ±0.004 | 0.260 ±0.009 | +0.030 |
| 28,000 | 0.332 ±0.004 | 0.289 ±0.001 | +0.043 |

**Scope note.** The direction replicates across both image distributions. The
magnitude does not. At 4,000 training images the gap is +0.061 on val2017 and
+0.012 on train2017, a five-fold difference. Treat the sign as the finding and
the size as distribution-dependent.

Further scope: one backbone, one architecture family, COCO images, English
captions, retrieval as the task. L47 remains the shipped tap for the
SRT-Sunstone artifacts, which were fitted and validated at that layer. Nothing
here has been run at deployment scale, and none of it has been tested on a
second backbone.

## SugarCrepe under a fitted probe

Fitted per layer on the 3,440 val2017 images SugarCrepe does not use, then
applied to its 7,511 triples. The `replace_obj` split, n=1652:

| layer | cosine | fitted |
|---|---|---|
| L4 | 0.514 | 0.557 |
| L12 | 0.528 | 0.611 |
| **L20** | 0.571 | **0.696** |
| L28 | 0.533 | 0.574 |
| L36 | 0.561 | 0.626 |
| L44 | 0.617 | 0.669 |
| L47 | 0.614 | 0.661 |
| L54 | 0.666 | 0.676 |
| L60 | 0.541 | 0.562 |

Object identity reaches its ceiling at L20 and stays flat, so identity is not
a late-stack property under this probe. Arrangement peaks mid-stack instead:
fitted `swap_obj` is 0.596 at L28. The macro ordering also inverts between
probes, with cosine best at L54 (0.569) and fitted best at L36 (0.598).

### Read the swap splits with care

`sugarcrepe_symmetry_audit.json` counts, for every split, how many pairs have
a positive and negative that are the same multiset of words:

| split | n | bag-of-words identical |
|---|---|---|
| swap_obj | 245 | 164 (66.9%) |
| swap_att | 666 | 408 (61.3%) |
| replace_att | 788 | 0 |
| replace_obj | 1652 | 0 |
| replace_rel | 1406 | 0 |
| add_att | 692 | 0 |
| add_obj | 2062 | 0 |

For a bag-identical pair, any model whose text representation ignores word
order scores both captions the same and lands at chance by construction. Both
swap splits are majority such pairs, so a swap score cannot separate content
quality across that whole class of models. This is a property of the data and
not a defect, since the swap splits exist to probe word order. It does change
how the number reads. Reproduce with `scripts/sugarcrepe_symmetry_audit.py`,
which runs on the benchmark JSONs alone and evaluates no model.

## Banked null

`l60_fitted_readout.json` is a failed experiment, published because it failed
its own control. A ridge/PCA read-out on 936 images scored 0.5222 at L47
against a head-free 0.5655, so the instrument was weaker than no instrument.
The regularisation was also chosen on the test split. The result is void and
the later InfoNCE refit (`l60_head_refit.json`, `depth_fitted_profile.json`)
supersedes it. It is here so the path to the working version is visible.

## Other artifacts

- `crosslingual_convergence.json`. Opus-100 English pivot, frozen L47, 400
  pairs per language, per-language centring. English–Mandarin 0.335 (134x
  chance), English–Arabic 0.297, English–Finnish 0.247. Confound stated in
  the file: one model was trained on all three languages, so this is not evidence about
  independently evolved systems.
- `space_capacity.json`. Retrieval against growing pools. Pool 1,000 gives
  median 1; 123,287 gives median 70. The pool grows 123x while rank grows
  70x. Participation ratio 84.5 of 1024 dimensions, 99% of variance in 210.
- `invariant_core.json`. 13 multi-backbone claims across 8 hosts, 11 hold.
  Worth reading for what it concedes: 7 of the 11 describe how to measure,
  and only 4 describe what a model does.
- `sugarcrepe_layer_sweep.json`. The cosine-only sweep that the fitted
  re-test above supersedes.

## Reproducing

`scripts/` contains the exact code. `l60_head_refit.py` encodes and fits the
depth profile, `tap_layer_scaling.py` runs the seeded scaling curves,
`tap_layer_full.py` is the full-scale version, `sugarcrepe_fitted_depth.py`
does the compositional re-test. Each writes the JSON it is named after.

Every fit uses a held-out split, and `l60_head_refit.py` enforces a control:
a fitted head at L47 must beat head-free, or the instrument is rejected.

## Anisotropy

These states are strongly anisotropic and raw cosine on them is not directly
interpretable. Comparisons here are centred by the pool mean. Any reuse of
these artifacts should do the same and should report a random floor.

## Citation

```bibtex
@software{srt_depth_probe_2026,
  title  = {SRT depth-probe artifacts: probe-dependent layer ranking in a
            frozen multimodal backbone},
  author = {Lancaster, Burton},
  year   = {2026},
  url    = {https://github.com/space-bacon/SRT}
}
```

Apache-2.0. Source: [github.com/space-bacon/SRT](https://github.com/space-bacon/SRT).