Every number in the SRT program's cross-modal and invariance sections has a file here. This repository is the evidence, not the model: encoded states, fitted heads, result JSONs, and run logs.

Papers: arxiv_program/paper.md. Code: github.com/space-bacon/SRT.

Reproduction principle. If a claim is in the paper and you cannot rebuild it from the files here plus the scripts in the GitHub repo, that is a bug and we want to hear about it. Two such bugs have been found by external review and both are fixed: the cross-runtime validation that never loaded the head, and the gemma/gemma v6 head that existed nowhere. See the ledger in §9 of the paper.

Reading mlp_align_full.json correctly. linear_n117000 i2t R@1 0.661 is the no-hard-negative recipe. Heads trained with compositional hard negatives at weight 1.0 land at 0.622 clean i2t by design, trading retrieval for compositional accuracy. Comparing a hard-negative head against 0.661 understates it by about four points.

Heads (all linear, 1024-dim projection, tau 0.05, seed 0, trained on the 118k pairs above with K=4 compositional negatives at weight 1.0):

gemma_v6_head.pt is a rebuild. The original existed neither locally nor here, which external review would have found; it was regenerated from procrustes/train_pairs plus a seed-0 regeneration of the v6 negatives (473,148 at K=4, prep_replace 75,786, adj_transfer 54,894, matching the published counts to the item). It reproduces baseline_centered 0.288 and the shuffled control 0.001 exactly, and linear 0.620 against 0.622. Details in mlp_align_gemma_v6_rebuild.json.

SugarCrepe states, so the benchmark can be scored without a GPU:

Together with the four heads these reproduce every cell of the tower factorial on CPU in seconds via scripts/sugarcrepe_controls.py.

Controls and results:

Which null to trust. Three of the four are blind to the thing they null: the bigram prior, the mean-image ablation, and a random fixed direction never open an image, and on swap_obj two of them sit 17.8 points apart, so the sign of "does image content help" was being set by null choice rather than measurement. The mean-image ablation is reproduced by another split's mean image and defeated by a random direction, so it prices a learned generic-image direction in text space, not image content. Its apparent win over real images was the caption formatting and does not survive normalisation. Candidate position exchange cannot arbitrate either: it is an identity for an independent-scoring cosine.

Use shuffle_pair (real images, trained head, pairing deranged over 20 seeds). It is the only null that varies per cell. Against it, on normalised captions, the heads clear by +16.3 to +18.6 clean-five points.

Normalise the captions first. SugarCrepe pairs a human COCO caption with a model-generated foil and the generator punctuates: the foil is the dotted one in 1,384 of the 1,384 pairs where the two differ, no counter-example in 7,511. States cached under the raw caption carry that signature. Stripping it reverses three conclusions we published: both swap splits go from at or below their deranged control to about ten points above it, the mean-image ablation stops outscoring real images, and the text-tower asymmetry inverts and shrinks to under a point on raw accuracy. The un-normalised numbers are kept because they are what other work quoting SugarCrepe is comparable to, not because they are the better measurement. Normalised state caches are published here as q4/sugarcrepe_txt_norm_*.npz, with q4/sugarcrepe_txt_reenc_*.npz as the same-session re-encode control (floor 0.0042 on any arm).

encoded_L29_n5000.pt, qwen3b_linear_head.pt, mlp_align_qwen3b.json, procrustes_qwen3b.json.

Both models held resident on separate GPUs with the same batch through each, so the comparison carries no re-encode floor. Headline: layers 0 to 14 displaced exactly zero, onset at 15, peak rho 0.074 at layer 48, cosine never below 0.9949. For scale, cross-runtime drift in the same units is rho 1.334 (MLX-Q4) and 0.749 (NF4), so a behavioural weight edit is roughly 10 to 18 times smaller geometrically than a quantizer swap. Capability holds (+0.028 nats on held-out text) while refusal-register text costs +0.479. The effect is register-local, not topic-local: contestedness tiers are flat.

Run logs, anchor sets, and the base-model K-curve.

Then score with scripts/sugarcrepe_controls.py --cell gemma-img_x_gemma-txt.

Conventions two and three are due to Dipankar Sarkar, who reviewed this program in public over fourteen rounds and reproduced our numbers from these files alone at every step.

Lancaster, J. B. (2026). Train Once, Read Everywhere: Substrate Invariance of the Linearly Readable Structure in Frozen Language Models.