Coverage: all plotted cells have seeds 0, 1 and 2.
Figures
Downloads
per_seed_results.csv and aggregate_seed_results.csv contain the plotted values. Existing per-task bootstrap confidence intervals are carried through in the per-seed CSV when present; the bar error bars are seed variability.
Aggregate Table
| Direction | Dataset | Metric | Mechanism | n | Direct | Grounded mean +/- SD | Delta mean +/- SD | Directionality |
|---|---|---|---|---|---|---|---|---|
| BGE -> RepTok | COCO-1000 | CLIP similarity | Activated dynamic + sigmoid keep | 3 | 0.28647 | 0.291675 +/- 0.000239 | +0.00520567 +/- 0.000239 | higher is better |
| BGE -> RepTok | COCO-1000 | CLIP similarity | Activated dynamic + softmax keep | 3 | 0.28647 | 0.290629 +/- 0.000851 | +0.00415929 +/- 0.000851 | higher is better |
| BGE -> RepTok | COCO-1000 | CLIP similarity | Activated pinned + sigmoid keep | 3 | 0.28647 | 0.293101 +/- 0.000112 | +0.00663126 +/- 0.000112 | higher is better |
| BGE -> RepTok | COCO-1000 | CLIP similarity | Activated pinned + softmax keep | 3 | 0.28647 | 0.293127 +/- 0.000527 | +0.00665744 +/- 0.000527 | higher is better |
| BGE -> RepTok | COCO-1000 | CLIP similarity | Raw pinned + sigmoid keep | 3 | 0.28647 | 0.292705 +/- 5.71e-05 | +0.00623528 +/- 5.71e-05 | higher is better |
| BGE -> RepTok | COCO-1000 | CLIP similarity | Raw pinned + softmax keep | 3 | 0.28647 | 0.292329 +/- 0.000295 | +0.00585931 +/- 0.000295 | higher is better |
| BGE -> RepTok | COCO-1000 | FID | Activated dynamic + sigmoid keep | 3 | 75.6146 | 69.2706 +/- 0.324 | -6.34405 +/- 0.324 | lower is better |
| BGE -> RepTok | COCO-1000 | FID | Activated dynamic + softmax keep | 3 | 75.6146 | 69.6307 +/- 0.802 | -5.98393 +/- 0.802 | lower is better |
| BGE -> RepTok | COCO-1000 | FID | Activated pinned + sigmoid keep | 3 | 75.6146 | 68.697 +/- 0.121 | -6.91765 +/- 0.121 | lower is better |
| BGE -> RepTok | COCO-1000 | FID | Activated pinned + softmax keep | 3 | 75.6146 | 68.0537 +/- 0.259 | -7.5609 +/- 0.259 | lower is better |
| BGE -> RepTok | COCO-1000 | FID | Raw pinned + sigmoid keep | 3 | 75.6146 | 68.7221 +/- 0.121 | -6.89254 +/- 0.121 | lower is better |
| BGE -> RepTok | COCO-1000 | FID | Raw pinned + softmax keep | 3 | 75.6146 | 69.0655 +/- 0.0842 | -6.54913 +/- 0.0842 | lower is better |
| BGE -> RepTok | COCO-1000 | LPIPS | Activated dynamic + sigmoid keep | 3 | 0.697183 | 0.69602 +/- 0.000304 | -0.00116261 +/- 0.000304 | lower is better |
| BGE -> RepTok | COCO-1000 | LPIPS | Activated dynamic + softmax keep | 3 | 0.697183 | 0.695985 +/- 0.000241 | -0.00119813 +/- 0.000241 | lower is better |
| BGE -> RepTok | COCO-1000 | LPIPS | Activated pinned + sigmoid keep | 3 | 0.697183 | 0.697211 +/- 6.63e-05 | +2.78354e-05 +/- 6.63e-05 | lower is better |
| BGE -> RepTok | COCO-1000 | LPIPS | Activated pinned + softmax keep | 3 | 0.697183 | 0.69724 +/- 0.000408 | +5.72602e-05 +/- 0.000408 | lower is better |
| BGE -> RepTok | COCO-1000 | LPIPS | Raw pinned + sigmoid keep | 3 | 0.697183 | 0.696936 +/- 9.32e-05 | -0.000246982 +/- 9.32e-05 | lower is better |
| BGE -> RepTok | COCO-1000 | LPIPS | Raw pinned + softmax keep | 3 | 0.697183 | 0.69697 +/- 0.000115 | -0.000213166 +/- 0.000115 | lower is better |
| BGE -> RepTok | ImageNet-1000 | CLIP similarity | Activated dynamic + sigmoid keep | 3 | 0.324203 | 0.326046 +/- 0.00014 | +0.00184281 +/- 0.00014 | higher is better |
| BGE -> RepTok | ImageNet-1000 | CLIP similarity | Activated dynamic + softmax keep | 3 | 0.324203 | 0.325806 +/- 0.000292 | +0.00160271 +/- 0.000292 | higher is better |
| BGE -> RepTok | ImageNet-1000 | CLIP similarity | Activated pinned + sigmoid keep | 3 | 0.324203 | 0.326694 +/- 0.000168 | +0.00249083 +/- 0.000168 | higher is better |
| BGE -> RepTok | ImageNet-1000 | CLIP similarity | Activated pinned + softmax keep | 3 | 0.324203 | 0.326585 +/- 0.000177 | +0.00238241 +/- 0.000177 | higher is better |
| BGE -> RepTok | ImageNet-1000 | CLIP similarity | Raw pinned + sigmoid keep | 3 | 0.324203 | 0.326339 +/- 0.000298 | +0.00213664 +/- 0.000298 | higher is better |
| BGE -> RepTok | ImageNet-1000 | CLIP similarity | Raw pinned + softmax keep | 3 | 0.324203 | 0.326415 +/- 9.6e-05 | +0.00221173 +/- 9.6e-05 | higher is better |
| BGE -> RepTok | ImageNet-1000 | FID | Activated dynamic + sigmoid keep | 3 | 51.64 | 50.6868 +/- 0.125 | -0.953152 +/- 0.125 | lower is better |
| BGE -> RepTok | ImageNet-1000 | FID | Activated dynamic + softmax keep | 3 | 51.64 | 50.7774 +/- 0.16 | -0.862517 +/- 0.16 | lower is better |
| BGE -> RepTok | ImageNet-1000 | FID | Activated pinned + sigmoid keep | 3 | 51.64 | 50.4729 +/- 0.00393 | -1.16708 +/- 0.00393 | lower is better |
| BGE -> RepTok | ImageNet-1000 | FID | Activated pinned + softmax keep | 3 | 51.64 | 50.4728 +/- 0.17 | -1.16719 +/- 0.17 | lower is better |
| BGE -> RepTok | ImageNet-1000 | FID | Raw pinned + sigmoid keep | 3 | 51.64 | 50.5645 +/- 0.156 | -1.07546 +/- 0.156 | lower is better |
| BGE -> RepTok | ImageNet-1000 | FID | Raw pinned + softmax keep | 3 | 51.64 | 50.6411 +/- 0.13 | -0.998834 +/- 0.13 | lower is better |
| BGE -> RepTok | ImageNet-1000 | LPIPS | Activated dynamic + sigmoid keep | 3 | 0.647536 | 0.646692 +/- 0.000103 | -0.000843028 +/- 0.000103 | lower is better |
| BGE -> RepTok | ImageNet-1000 | LPIPS | Activated dynamic + softmax keep | 3 | 0.647536 | 0.647548 +/- 0.000159 | +1.2219e-05 +/- 0.000159 | lower is better |
| BGE -> RepTok | ImageNet-1000 | LPIPS | Activated pinned + sigmoid keep | 3 | 0.647536 | 0.646077 +/- 0.000292 | -0.00145831 +/- 0.000292 | lower is better |
| BGE -> RepTok | ImageNet-1000 | LPIPS | Activated pinned + softmax keep | 3 | 0.647536 | 0.64568 +/- 0.00013 | -0.00185593 +/- 0.00013 | lower is better |
| BGE -> RepTok | ImageNet-1000 | LPIPS | Raw pinned + sigmoid keep | 3 | 0.647536 | 0.645413 +/- 0.000147 | -0.00212284 +/- 0.000147 | lower is better |
| BGE -> RepTok | ImageNet-1000 | LPIPS | Raw pinned + softmax keep | 3 | 0.647536 | 0.645638 +/- 0.000314 | -0.00189779 +/- 0.000314 | lower is better |
| DINO -> symbolic | ImageNet-50k | Top-1 accuracy | Activated dynamic + sigmoid keep | 3 | 0.836 | 0.836027 +/- 6.43e-05 | +2.66433e-05 +/- 6.43e-05 | higher is better |
| DINO -> symbolic | ImageNet-50k | Top-1 accuracy | Activated dynamic + softmax keep | 3 | 0.836 | 0.799627 +/- 0.0413 | -0.0363734 +/- 0.0413 | higher is better |
| DINO -> symbolic | ImageNet-50k | Top-1 accuracy | Activated pinned + sigmoid keep | 3 | 0.836 | 0.83598 +/- 0.00014 | -2.00272e-05 +/- 0.00014 | higher is better |
| DINO -> symbolic | ImageNet-50k | Top-1 accuracy | Activated pinned + softmax keep | 3 | 0.836 | 0.81284 +/- 0.0203 | -0.02316 +/- 0.0203 | higher is better |
| DINO -> symbolic | ImageNet-50k | Top-1 accuracy | Raw pinned + sigmoid keep | 3 | 0.836 | 0.830207 +/- 0.00108 | -0.00579335 +/- 0.00108 | higher is better |
| DINO -> symbolic | ImageNet-50k | Top-1 accuracy | Raw pinned + softmax keep | 3 | 0.836 | 0.79168 +/- 2e-05 | -0.04432 +/- 2e-05 | higher is better |
| RepTok -> SONAR | ImageNet-50k | Corpus BLEU-4 | Activated dynamic + sigmoid keep | 3 | 11.8117 | 12.3648 +/- 0.423 | +0.553028 +/- 0.423 | higher is better |
| RepTok -> SONAR | ImageNet-50k | Corpus BLEU-4 | Activated dynamic + softmax keep | 3 | 11.8117 | 11.5165 +/- 1.31 | -0.295276 +/- 1.31 | higher is better |
| RepTok -> SONAR | ImageNet-50k | Corpus BLEU-4 | Activated pinned + sigmoid keep | 3 | 11.8117 | 11.2083 +/- 0.941 | -0.603406 +/- 0.941 | higher is better |
| RepTok -> SONAR | ImageNet-50k | Corpus BLEU-4 | Activated pinned + softmax keep | 3 | 11.8117 | 11.3173 +/- 1.57 | -0.494454 +/- 1.57 | higher is better |
| RepTok -> SONAR | ImageNet-50k | Corpus BLEU-4 | Raw pinned + sigmoid keep | 3 | 11.8117 | 12.3779 +/- 0.453 | +0.566166 +/- 0.453 | higher is better |
| RepTok -> SONAR | ImageNet-50k | Corpus BLEU-4 | Raw pinned + softmax keep | 3 | 11.8117 | 11.7102 +/- 1.19 | -0.101519 +/- 1.19 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | ResNet target-class accuracy | Activated dynamic + sigmoid keep | 3 | 0.9525 | 0.957567 +/- 0.00177 | +0.00506669 +/- 0.00177 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | ResNet target-class accuracy | Activated dynamic + softmax keep | 3 | 0.9499 | 0.9546 +/- 0.00329 | +0.00469995 +/- 0.00329 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | ResNet target-class accuracy | Activated pinned + sigmoid keep | 3 | 0.9505 | 0.951633 +/- 0.00142 | +0.00113332 +/- 0.00142 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | ResNet target-class accuracy | Activated pinned + softmax keep | 3 | 0.9468 | 0.956033 +/- 0.00311 | +0.00923334 +/- 0.00311 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | ResNet target-class accuracy | Raw pinned + sigmoid keep | 3 | 0.9509 | 0.402733 +/- 0.115 | -0.548167 +/- 0.115 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | ResNet target-class accuracy | Raw pinned + softmax keep | 3 | 0.9509 | 0.3971 +/- 0.035 | -0.5538 +/- 0.035 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO IoU | Activated dynamic + sigmoid keep | 3 | 0.460004 | 0.458865 +/- 0.00115 | -0.00113982 +/- 0.00115 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO IoU | Activated dynamic + softmax keep | 3 | 0.459704 | 0.448074 +/- 0.00171 | -0.0116296 +/- 0.00171 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO IoU | Activated pinned + sigmoid keep | 3 | 0.46051 | 0.476182 +/- 0.00497 | +0.0156717 +/- 0.00497 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO IoU | Activated pinned + softmax keep | 3 | 0.458074 | 0.476833 +/- 0.00159 | +0.018759 +/- 0.00159 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO IoU | Raw pinned + sigmoid keep | 3 | 0.459411 | 0.463827 +/- 0.0234 | +0.00441679 +/- 0.0234 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO IoU | Raw pinned + softmax keep | 3 | 0.458388 | 0.492687 +/- 0.00288 | +0.0342994 +/- 0.00288 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO detection rate | Activated dynamic + sigmoid keep | 3 | 0.9847 | 0.9822 +/- 0.000889 | -0.00249996 +/- 0.000889 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO detection rate | Activated dynamic + softmax keep | 3 | 0.986 | 0.978367 +/- 0.000379 | -0.00763335 +/- 0.000379 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO detection rate | Activated pinned + sigmoid keep | 3 | 0.9823 | 0.978333 +/- 0.000586 | -0.00396659 +/- 0.000586 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO detection rate | Activated pinned + softmax keep | 3 | 0.9833 | 0.972867 +/- 0.00159 | -0.0104334 +/- 0.00159 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO detection rate | Raw pinned + sigmoid keep | 3 | 0.9822 | 0.991833 +/- 0.00196 | +0.00963328 +/- 0.00196 | higher is better |
| Symbolic -> RepTok | ImageNet 1000 x 10 | YOLO detection rate | Raw pinned + softmax keep | 3 | 0.9848 | 0.986967 +/- 0.0037 | +0.00216673 +/- 0.0037 | higher is better |
Exploratory BBox Class-CE Addendum
This bottom addendum is not one of the six factorial cells above. It reports only the raw-pinned sigmoid bbox grounder retrained with the July-style class-CE term, because the raw bbox variants improved spatial IoU while losing decoded class identity.
Source root: gw_setup/outputs/ch4_bbox_raw_pinned_ce_20260830/evals/raw_pinned_sigmoid_ce/bbox_symbolic_reptok
ce_bbox_raw_pinned_sigmoid_values.csv and ce_bbox_raw_pinned_sigmoid_deltas.csv
| Condition | Seed | Metric | Delta | CI low | CI high |
|---|---|---|---|---|---|
| grounded_seed0 | 0 | YOLO IoU | 0.0289349 | 0.0232689 | 0.0348198 |
| grounded_seed0 | 0 | Detection rate | -0.005 | -0.0086 | -0.00150001 |
| grounded_seed0 | 0 | ResNet target-class accuracy | -0.011 | -0.0177 | -0.00440001 |
| grounded_seed1 | 1 | YOLO IoU | 0.0240556 | 0.0178918 | 0.0301982 |
| grounded_seed1 | 1 | Detection rate | -0.0111 | -0.0152 | -0.0072974 |
| grounded_seed1 | 1 | ResNet target-class accuracy | -0.00479996 | -0.0116 | 0.00209999 |
| grounded_seed2 | 2 | YOLO IoU | 0.0109504 | 0.00380066 | 0.0178778 |
| grounded_seed2 | 2 | Detection rate | -0.00380003 | -0.00760262 | -0.000100017 |
| grounded_seed2 | 2 | ResNet target-class accuracy | -0.0444 | -0.0535 | -0.0355 |
Thesis Asset Package
Download the Page-36 final Chapter 4 thesis handoff bundle: ch4_page36_final_thesis_assets_20260831.zip.
This bundle contains copied final-result CSV/JSON/bootstrap files, selected qualitative assets, provenance tables, prose method notes, and lightweight controller-step diagnostics. It does not rerun completed external evaluations.