Initialize repo with Stage 0 baseline

README.md

@@ -0,0 +1,24 @@
+# 1-Parameter Classifier
+
+Progressively reducing the model budget for image-level person classification on EUPE-ViT-B features. Each stage is a deeper reduction or transformation of the previous. The classifier shrinks across stages while the backbone it draws features from is attacked in parallel.
+
+## Stage 0: Baseline
+
+A 1-free-parameter image-level person classifier on the frozen EUPE-ViT-B backbone. The classifier reads 20 pre-selected person-positive and 20 pre-selected person-negative feature dimensions, sums the positives, subtracts the negatives, and compares the result to one learned threshold. F1 = 0.889 on COCO val 2017 image-level person presence, measured through the live Argus forward pass at 768 pixel input.
+
+See [`stage_0/`](stage_0/) for the classifier config, discovery pipeline, and full characterization of the person axis in the backbone.
+
+## Roadmap
+
+| Stage | Name | What changes |
+|---|---|---|
+| 0 | Baseline 1-param classifier | Uses the full EUPE-ViT-B backbone unchanged |
+| 1 | Output-channel pruning | Keep only the 100 feature dims the classifier reads |
+| 2 | Attention-head pruning | Ablate heads that do not contribute to those 100 dims |
+| 3 | Depth reduction | Drop transformer blocks that do not route signal to the 100 dims |
+| 4 | Specialist backbone | Train a small student that emits only the 100 target dims |
+| 5 | Circuit-level synthesis | Synthesize the entire fixed-weight pipeline to gates and dead-code eliminate everything that does not reach the classifier output |
+
+## Source backbone
+
+EUPE-ViT-B from Meta FAIR ([arXiv:2603.22387](https://arxiv.org/abs/2603.22387), Zhu et al., March 2026), distilled from PEcore-G + PElang-G + DINOv3-H+ via a 1.9B proxy teacher. License: FAIR Research License (non-commercial). The 1-parameter classifier is an artifact derived from that backbone's feature geometry.

stage_0/README.md

@@ -0,0 +1,41 @@
+# Stage 0: Baseline 1-Parameter Classifier
+
+Image-level person classifier on frozen EUPE-ViT-B features. One free scalar parameter.
+
+## Classifier
+
+Given a 768 pixel input image, forward through EUPE-ViT-B, take the 2304 patch tokens at the final layer, apply layernorm across the 768-channel axis, and max-pool across patches to get a single 768-D vector per image. The classifier reads 40 of those 768 dimensions: 20 person-positive, 20 person-negative. Sum the positives, subtract the negatives, compare against one learned threshold.
+
+```python
+# pseudocode
+patches = backbone(image)["x_norm_patchtokens"]         # (2304, 768)
+pooled  = layernorm(patches, 768).max(dim=patches)      # (768,)
+score   = pooled[pos_dims].sum() - pooled[neg_dims].sum()
+pred    = score > threshold                             # bool
+```
+
+All arrays (`pos_dims`, `neg_dims`, `threshold`) are in `classifier.json`.
+
+## Evaluation
+
+F1 = 0.889, precision = 0.901, recall = 0.876 on 5000 COCO val 2017 images, measured through the live Argus forward pass at 768 pixel input.
+
+See `eval.json`.
+
+## How the dim selection was discovered
+
+Three-step process, documented in the artifacts below.
+
+**`cojoint_discovery.json`** — Sampled 100,000 random 92-dim subsets of the 768-D EUPE-ViT-B feature space, trained a ridge classifier for each, kept the top 1%, counted dim occurrence frequency across the kept cohort. Dim 48 appeared in 100% of top-1000 subsets. Next strongest (dim 525) appeared in 31%.
+
+**`characterization.json`** — Five analyses on dim 48 specifically. F1 versus K (dim 48 alone reaches F1 = 0.83). Activation distribution for person-positive versus person-negative images (Cohen's d = 1.98). Per-class activation delta for each of 80 COCO categories. Top-10 frequent-dim pairwise correlation (max |r| = 0.57, mostly independent). Spatial IoU of dim-48 peak activations against ground-truth person boxes (mean IoU = 0.17 — dim 48 is a scene-level signal, not a pixel-level localizer).
+
+**`compressed_variants.json`** — Leaderboard of 20+ classifier variants ranging from 1 free parameter to 769. Ranked, the ternary ±1 on 50 positive plus 50 negative dims wins at F1 = 0.893. The 20+20 variant chosen for Stage 0 is the same recipe at smaller footprint with F1 = 0.881 cached / 0.889 live.
+
+## Interpretation
+
+Dim 48 is the canonical anthropogenic-scene axis in EUPE-ViT-B. It activates strongly on person scenes and on person-associated objects (sports equipment, wearable accessories, handheld items), and is suppressed on non-human animals and non-anthropogenic structures. Alone it delivers F1 = 0.83 as a 2-parameter classifier. The additional 39 dims stack on mostly orthogonal axes to reach F1 = 0.89 at 1 free parameter.
+
+## Hardware footprint estimate
+
+At INT8 precision, the classifier synthesizes to an estimated 2,500–4,100 gates: two Wallace-tree adders (50-input each), one subtractor, one comparator. For reference, a 768-dim INT8 MAC unit is roughly 65,000 gates, and the prior 4,614-parameter multi-output person detector synthesizes to roughly 391,000 gates.

stage_0/characterization.json

@@ -0,0 +1,542 @@
+{
+  "target_dim": 48,
+  "A_f1_vs_k": [
+    {
+      "K": 1,
+      "F1": 0.8285356163978577,
+      "precision": 0.7985524535179138,
+      "recall": 0.8608582615852356,
+      "dims": [
+        48
+      ]
+    },
+    {
+      "K": 2,
+      "F1": 0.8562204241752625,
+      "precision": 0.8503633737564087,
+      "recall": 0.8621586561203003,
+      "dims": [
+        48,
+        525
+      ]
+    },
+    {
+      "K": 3,
+      "F1": 0.8652572631835938,
+      "precision": 0.8526936173439026,
+      "recall": 0.8781967759132385,
+      "dims": [
+        48,
+        525,
+        475
+      ]
+    },
+    {
+      "K": 5,
+      "F1": 0.8649698495864868,
+      "precision": 0.8608844876289368,
+      "recall": 0.8690940737724304,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240
+      ]
+    },
+    {
+      "K": 10,
+      "F1": 0.9092138409614563,
+      "precision": 0.9465290904045105,
+      "recall": 0.8747290968894958,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240,
+        "..."
+      ]
+    },
+    {
+      "K": 20,
+      "F1": 0.9288457632064819,
+      "precision": 0.9698113203048706,
+      "recall": 0.8912007212638855,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240,
+        "..."
+      ]
+    },
+    {
+      "K": 30,
+      "F1": 0.9349648356437683,
+      "precision": 0.9795821309089661,
+      "recall": 0.8942349553108215,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240,
+        "..."
+      ]
+    },
+    {
+      "K": 50,
+      "F1": 0.9407240152359009,
+      "precision": 0.9839091300964355,
+      "recall": 0.9011703729629517,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240,
+        "..."
+      ]
+    },
+    {
+      "K": 92,
+      "F1": 0.9444695711135864,
+      "precision": 0.98539799451828,
+      "recall": 0.9068053960800171,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240,
+        "..."
+      ]
+    },
+    {
+      "K": 184,
+      "F1": 0.9485046863555908,
+      "precision": 0.9855140447616577,
+      "recall": 0.9141742587089539,
+      "dims": [
+        48,
+        525,
+        475,
+        323,
+        240,
+        "..."
+      ]
+    }
+  ],
+  "B_dim48_distribution": {
+    "person_pos": {
+      "n": 2693,
+      "mean": 8.113646507263184,
+      "std": 2.3924715518951416,
+      "min": -1.2812166213989258,
+      "p25": 7.193147659301758,
+      "p50": 8.808687210083008,
+      "p75": 9.730491638183594,
+      "max": 12.061566352844238
+    },
+    "person_neg": {
+      "n": 2307,
+      "mean": 2.982386589050293,
+      "std": 2.7675836086273193,
+      "min": -3.072082042694092,
+      "p25": 0.5744761228561401,
+      "p50": 3.059739589691162,
+      "p75": 5.2233123779296875,
+      "max": 11.300838470458984
+    },
+    "separation_cohen_d": 1.983604907989502
+  },
+  "C_per_class_top15": [
+    {
+      "class": "person",
+      "cat_id": 1,
+      "n_imgs": 2693,
+      "dim48_mean_inclass": 8.1136474609375,
+      "dim48_mean_outclass": 2.982386350631714,
+      "delta": 5.131261110305786
+    },
+    {
+      "class": "tennis racket",
+      "cat_id": 43,
+      "n_imgs": 167,
+      "dim48_mean_inclass": 9.163715362548828,
+      "dim48_mean_outclass": 5.627990245819092,
+      "delta": 3.5357251167297363
+    },
+    {
+      "class": "tie",
+      "cat_id": 32,
+      "n_imgs": 145,
+      "dim48_mean_inclass": 8.980120658874512,
+      "dim48_mean_outclass": 5.6494951248168945,
+      "delta": 3.330625534057617
+    },
+    {
+      "class": "baseball glove",
+      "cat_id": 40,
+      "n_imgs": 100,
+      "dim48_mean_inclass": 8.982823371887207,
+      "dim48_mean_outclass": 5.680027008056641,
+      "delta": 3.3027963638305664
+    },
+    {
+      "class": "skateboard",
+      "cat_id": 41,
+      "n_imgs": 127,
+      "dim48_mean_inclass": 8.923754692077637,
+      "dim48_mean_outclass": 5.663267135620117,
+      "delta": 3.2604875564575195
+    },
+    {
+      "class": "baseball bat",
+      "cat_id": 39,
+      "n_imgs": 97,
+      "dim48_mean_inclass": 8.886703491210938,
+      "dim48_mean_outclass": 5.683949947357178,
+      "delta": 3.2027535438537598
+    },
+    {
+      "class": "backpack",
+      "cat_id": 27,
+      "n_imgs": 228,
+      "dim48_mean_inclass": 8.655838012695312,
+      "dim48_mean_outclass": 5.607059001922607,
+      "delta": 3.048779010772705
+    },
+    {
+      "class": "remote",
+      "cat_id": 75,
+      "n_imgs": 145,
+      "dim48_mean_inclass": 8.703699111938477,
+      "dim48_mean_outclass": 5.657751083374023,
+      "delta": 3.045948028564453
+    },
+    {
+      "class": "toothbrush",
+      "cat_id": 90,
+      "n_imgs": 34,
+      "dim48_mean_inclass": 8.746152877807617,
+      "dim48_mean_outclass": 5.725543022155762,
+      "delta": 3.0206098556518555
+    },
+    {
+      "class": "handbag",
+      "cat_id": 31,
+      "n_imgs": 292,
+      "dim48_mean_inclass": 8.576542854309082,
+      "dim48_mean_outclass": 5.570532321929932,
+      "delta": 3.0060105323791504
+    },
+    {
+      "class": "sports ball",
+      "cat_id": 37,
+      "n_imgs": 169,
+      "dim48_mean_inclass": 8.645074844360352,
+      "dim48_mean_outclass": 5.644669532775879,
+      "delta": 3.0004053115844727
+    },
+    {
+      "class": "hair drier",
+      "cat_id": 89,
+      "n_imgs": 9,
+      "dim48_mean_inclass": 8.637924194335938,
+      "dim48_mean_outclass": 5.74086856842041,
+      "delta": 2.8970556259155273
+    },
+    {
+      "class": "skis",
+      "cat_id": 35,
+      "n_imgs": 120,
+      "dim48_mean_inclass": 8.562700271606445,
+      "dim48_mean_outclass": 5.676822185516357,
+      "delta": 2.885878086090088
+    },
+    {
+      "class": "cell phone",
+      "cat_id": 77,
+      "n_imgs": 214,
+      "dim48_mean_inclass": 8.483532905578613,
+      "dim48_mean_outclass": 5.623681545257568,
+      "delta": 2.859851360321045
+    },
+    {
+      "class": "snowboard",
+      "cat_id": 36,
+      "n_imgs": 49,
+      "dim48_mean_inclass": 8.451048851013184,
+      "dim48_mean_outclass": 5.71931266784668,
+      "delta": 2.731736183166504
+    }
+  ],
+  "C_per_class_bottom15": [
+    {
+      "class": "traffic light",
+      "cat_id": 10,
+      "n_imgs": 191,
+      "dim48_mean_inclass": 4.146858215332031,
+      "dim48_mean_outclass": 5.809599876403809,
+      "delta": -1.6627416610717773
+    },
+    {
+      "class": "boat",
+      "cat_id": 9,
+      "n_imgs": 121,
+      "dim48_mean_inclass": 4.040976047515869,
+      "dim48_mean_outclass": 5.788370132446289,
+      "delta": -1.74739408493042
+    },
+    {
+      "class": "fire hydrant",
+      "cat_id": 11,
+      "n_imgs": 86,
+      "dim48_mean_inclass": 3.6824471950531006,
+      "dim48_mean_outclass": 5.782199382781982,
+      "delta": -2.099752187728882
+    },
+    {
+      "class": "clock",
+      "cat_id": 85,
+      "n_imgs": 204,
+      "dim48_mean_inclass": 3.664764642715454,
+      "dim48_mean_outclass": 5.83461332321167,
+      "delta": -2.169848680496216
+    },
+    {
+      "class": "train",
+      "cat_id": 7,
+      "n_imgs": 157,
+      "dim48_mean_inclass": 3.4102745056152344,
+      "dim48_mean_outclass": 5.821805477142334,
+      "delta": -2.4115309715270996
+    },
+    {
+      "class": "sheep",
+      "cat_id": 20,
+      "n_imgs": 65,
+      "dim48_mean_inclass": 3.174736261367798,
+      "dim48_mean_outclass": 5.779951095581055,
+      "delta": -2.605214834213257
+    },
+    {
+      "class": "elephant",
+      "cat_id": 22,
+      "n_imgs": 89,
+      "dim48_mean_inclass": 3.023428201675415,
+      "dim48_mean_outclass": 5.795424938201904,
+      "delta": -2.7719967365264893
+    },
+    {
+      "class": "parking meter",
+      "cat_id": 14,
+      "n_imgs": 37,
+      "dim48_mean_inclass": 2.688075065612793,
+      "dim48_mean_outclass": 5.768880844116211,
+      "delta": -3.080805778503418
+    },
+    {
+      "class": "airplane",
+      "cat_id": 5,
+      "n_imgs": 97,
+      "dim48_mean_inclass": 2.6694958209991455,
+      "dim48_mean_outclass": 5.806950092315674,
+      "delta": -3.1374542713165283
+    },
+    {
+      "class": "stop sign",
+      "cat_id": 13,
+      "n_imgs": 69,
+      "dim48_mean_inclass": 2.509443998336792,
+      "dim48_mean_outclass": 5.7913737297058105,
+      "delta": -3.2819297313690186
+    },
+    {
+      "class": "cow",
+      "cat_id": 21,
+      "n_imgs": 87,
+      "dim48_mean_inclass": 2.1573596000671387,
+      "dim48_mean_outclass": 5.809632778167725,
+      "delta": -3.652273178100586
+    },
+    {
+      "class": "bird",
+      "cat_id": 16,
+      "n_imgs": 125,
+      "dim48_mean_inclass": 1.5200526714324951,
+      "dim48_mean_outclass": 5.854443073272705,
+      "delta": -4.33439040184021
+    },
+    {
+      "class": "giraffe",
+      "cat_id": 25,
+      "n_imgs": 101,
+      "dim48_mean_inclass": 0.5355944633483887,
+      "dim48_mean_outclass": 5.8535051345825195,
+      "delta": -5.317910671234131
+    },
+    {
+      "class": "zebra",
+      "cat_id": 24,
+      "n_imgs": 85,
+      "dim48_mean_inclass": 0.19826489686965942,
+      "dim48_mean_outclass": 5.842027187347412,
+      "delta": -5.643762290477753
+    },
+    {
+      "class": "bear",
+      "cat_id": 23,
+      "n_imgs": 49,
+      "dim48_mean_inclass": -0.3452981412410736,
+      "dim48_mean_outclass": 5.806369781494141,
+      "delta": -6.151667922735214
+    }
+  ],
+  "D_top10_correlation": {
+    "dims": [
+      48,
+      525,
+      475,
+      323,
+      240,
+      637,
+      318,
+      472,
+      719,
+      251
+    ],
+    "corr_matrix": [
+      [
+        1.000000238418579,
+        0.28202927112579346,
+        -0.03376583009958267,
+        0.24143743515014648,
+        -0.013180982321500778,
+        0.1519388109445572,
+        0.21512284874916077,
+        0.12231583148241043,
+        0.15586860477924347,
+        0.14881354570388794
+      ],
+      [
+        0.28202927112579346,
+        0.9999998211860657,
+        0.2280241698026657,
+        0.08577030897140503,
+        0.20908223092556,
+        -0.09965874999761581,
+        0.2280401587486267,
+        0.08895031362771988,
+        0.2658396065235138,
+        -0.15059183537960052
+      ],
+      [
+        -0.03376583009958267,
+        0.2280241698026657,
+        1.0000001192092896,
+        0.061812229454517365,
+        0.570030927658081,
+        -0.011771553196012974,
+        0.07696891576051712,
+        -0.10601639002561569,
+        0.06801445037126541,
+        -0.23264755308628082
+      ],
+      [
+        0.24143743515014648,
+        0.08577030897140503,
+        0.061812229454517365,
+        0.9999997615814209,
+        0.0030930074863135815,
+        0.18994615972042084,
+        0.15969695150852203,
+        0.21833136677742004,
+        0.26542553305625916,
+        -0.039861708879470825
+      ],
+      [
+        -0.013180982321500778,
+        0.20908223092556,
+        0.570030927658081,
+        0.0030930074863135815,
+        1.0000003576278687,
+        -0.09009360522031784,
+        0.003698738757520914,
+        -0.08297917991876602,
+        0.01635875552892685,
+        -0.21284666657447815
+      ],
+      [
+        0.1519388109445572,
+        -0.09965874999761581,
+        -0.011771553196012974,
+        0.18994615972042084,
+        -0.09009360522031784,
+        0.9999999403953552,
+        0.13532496988773346,
+        0.04327913001179695,
+        -0.03058011643588543,
+        0.22757190465927124
+      ],
+      [
+        0.21512284874916077,
+        0.2280401587486267,
+        0.07696891576051712,
+        0.15969695150852203,
+        0.003698738757520914,
+        0.13532496988773346,
+        0.9999998807907104,
+        0.030232839286327362,
+        0.23005540668964386,
+        0.025300158187747
+      ],
+      [
+        0.12231583148241043,
+        0.08895031362771988,
+        -0.10601639002561569,
+        0.21833136677742004,
+        -0.08297917991876602,
+        0.04327913001179695,
+        0.030232839286327362,
+        1.000000238418579,
+        0.21879234910011292,
+        0.020776957273483276
+      ],
+      [
+        0.15586860477924347,
+        0.2658396065235138,
+        0.06801445037126541,
+        0.26542553305625916,
+        0.01635875552892685,
+        -0.03058011643588543,
+        0.23005540668964386,
+        0.21879234910011292,
+        0.9999998807907104,
+        -0.015971055254340172
+      ],
+      [
+        0.14881354570388794,
+        -0.15059183537960052,
+        -0.23264755308628082,
+        -0.039861708879470825,
+        -0.21284666657447815,
+        0.22757190465927124,
+        0.025300158187747,
+        0.020776957273483276,
+        -0.015971055254340172,
+        1.0000001192092896
+      ]
+    ],
+    "max_abs_offdiag": 0.570030927658081
+  },
+  "E_spatial_localization": {
+    "n_sampled_images": 500,
+    "mean_iou": 0.16534487868892028,
+    "median_iou": 0.15606729686260223,
+    "p95_iou": 0.35281437188386916
+  }
+}

stage_0/classifier.json

@@ -0,0 +1,19 @@
+{
+  "backbone": "facebook/EUPE-ViT-B",
+  "feature_dim": 768,
+  "input_resolution": 768,
+  "patch_size": 16,
+  "patch_grid": [48, 48],
+  "preprocessing": "layernorm over 768 channels then max-pool over 2304 patches",
+  "pos_dims": [48, 525, 475, 645, 273, 292, 158, 510, 506, 337, 8, 309, 267, 217, 79, 13, 657, 207, 722, 311],
+  "neg_dims": [642, 224, 113, 565, 49, 637, 45, 520, 219, 290, 529, 617, 269, 745, 576, 701, 105, 694, 82, 283],
+  "pos_weight": 1.0,
+  "neg_weight": -1.0,
+  "threshold": 25.284494400024414,
+  "decision": "sum(feat[pos_dims]) - sum(feat[neg_dims]) > threshold",
+  "free_parameters": 1,
+  "fixed_parameters": {
+    "dim_indices": 40,
+    "signs": 40
+  }
+}

stage_0/cojoint_discovery.json

@@ -0,0 +1,500 @@
+{
+  "n_sampled": 100000,
+  "K": 92,
+  "n_kept": 1000,
+  "f1_distribution": {
+    "min": 0.7833036780357361,
+    "p50": 0.8676828145980835,
+    "p95": 0.9258173704147339,
+    "p99": 0.9300636053085327,
+    "max": 0.9373591542243958
+  },
+  "top_kept_f1_range": [
+    0.9300652146339417,
+    0.9373591542243958
+  ],
+  "best_single_genome_f1": 0.9373591542243958,
+  "best_single_genome_sorted": [
+    2,
+    4,
+    8,
+    21,
+    29,
+    39,
+    48,
+    67,
+    69,
+    75,
+    91,
+    119,
+    124,
+    141,
+    149,
+    157,
+    158,
+    168,
+    177,
+    181,
+    192,
+    198,
+    200,
+    218,
+    220,
+    240,
+    249,
+    250,
+    259,
+    267,
+    269,
+    272,
+    273,
+    275,
+    285,
+    286,
+    297,
+    300,
+    304,
+    318,
+    319,
+    323,
+    328,
+    330,
+    340,
+    373,
+    374,
+    377,
+    379,
+    391,
+    393,
+    403,
+    419,
+    429,
+    433,
+    436,
+    454,
+    463,
+    476,
+    520,
+    524,
+    530,
+    537,
+    548,
+    550,
+    558,
+    560,
+    570,
+    574,
+    582,
+    590,
+    594,
+    605,
+    606,
+    609,
+    612,
+    617,
+    624,
+    630,
+    646,
+    649,
+    657,
+    660,
+    673,
+    680,
+    700,
+    715,
+    716,
+    737,
+    740,
+    754,
+    755
+  ],
+  "dim_frequency_top_25": {
+    "48": 1.0,
+    "525": 0.3140000104904175,
+    "475": 0.3110000193119049,
+    "323": 0.23400001227855682,
+    "240": 0.21700000762939453,
+    "637": 0.2120000123977661,
+    "318": 0.20600001513957977,
+    "472": 0.20200000703334808,
+    "719": 0.20100000500679016,
+    "251": 0.19700001180171967,
+    "506": 0.19500000774860382,
+    "13": 0.1940000057220459,
+    "281": 0.1900000125169754,
+    "113": 0.1860000044107437,
+    "331": 0.1860000044107437,
+    "224": 0.18200001120567322,
+    "382": 0.1810000091791153,
+    "767": 0.18000000715255737,
+    "605": 0.1770000010728836,
+    "642": 0.17500001192092896,
+    "157": 0.17400000989437103,
+    "155": 0.1680000126361847,
+    "273": 0.16700001060962677,
+    "207": 0.16500000655651093,
+    "510": 0.16500000655651093
+  },
+  "union_classifier": {
+    "n_dims": 768,
+    "F1": 0.9616926908493042,
+    "precision": 0.989005982875824,
+    "recall": 0.9358474016189575
+  },
+  "cores": {
+    "23": [
+      48,
+      525,
+      475,
+      323,
+      240,
+      637,
+      318,
+      472,
+      719,
+      251,
+      506,
+      13,
+      281,
+      113,
+      331,
+      224,
+      382,
+      767,
+      605,
+      642,
+      157,
+      155,
+      273
+    ],
+    "46": [
+      48,
+      525,
+      475,
+      323,
+      240,
+      637,
+      318,
+      472,
+      719,
+      251,
+      506,
+      13,
+      281,
+      113,
+      331,
+      224,
+      382,
+      767,
+      605,
+      642,
+      157,
+      155,
+      273,
+      207,
+      510,
+      127,
+      245,
+      759,
+      292,
+      531,
+      617,
+      333,
+      527,
+      45,
+      332,
+      574,
+      595,
+      269,
+      158,
+      432,
+      0,
+      119,
+      310,
+      490,
+      557,
+      528
+    ],
+    "92": [
+      48,
+      525,
+      475,
+      323,
+      240,
+      637,
+      318,
+      472,
+      719,
+      251,
+      506,
+      13,
+      281,
+      113,
+      331,
+      224,
+      382,
+      767,
+      605,
+      642,
+      157,
+      155,
+      273,
+      207,
+      510,
+      127,
+      245,
+      759,
+      292,
+      531,
+      617,
+      333,
+      527,
+      45,
+      332,
+      574,
+      595,
+      269,
+      158,
+      432,
+      0,
+      119,
+      310,
+      490,
+      557,
+      528,
+      545,
+      558,
+      597,
+      645,
+      28,
+      51,
+      79,
+      82,
+      301,
+      126,
+      542,
+      173,
+      90,
+      380,
+      71,
+      384,
+      514,
+      620,
+      630,
+      713,
+      223,
+      516,
+      520,
+      606,
+      17,
+      172,
+      355,
+      393,
+      421,
+      745,
+      68,
+      252,
+      280,
+      334,
+      657,
+      95,
+      111,
+      151,
+      244,
+      543,
+      748,
+      290,
+      363,
+      469,
+      561,
+      591
+    ],
+    "184": [
+      48,
+      525,
+      475,
+      323,
+      240,
+      637,
+      318,
+      472,
+      719,
+      251,
+      506,
+      13,
+      281,
+      113,
+      331,
+      224,
+      382,
+      767,
+      605,
+      642,
+      157,
+      155,
+      273,
+      207,
+      510,
+      127,
+      245,
+      759,
+      292,
+      531,
+      617,
+      333,
+      527,
+      45,
+      332,
+      574,
+      595,
+      269,
+      158,
+      432,
+      0,
+      119,
+      310,
+      490,
+      557,
+      528,
+      545,
+      558,
+      597,
+      645,
+      28,
+      51,
+      79,
+      82,
+      301,
+      126,
+      542,
+      173,
+      90,
+      380,
+      71,
+      384,
+      514,
+      620,
+      630,
+      713,
+      223,
+      516,
+      520,
+      606,
+      17,
+      172,
+      355,
+      393,
+      421,
+      745,
+      68,
+      252,
+      280,
+      334,
+      657,
+      95,
+      111,
+      151,
+      244,
+      543,
+      748,
+      290,
+      363,
+      469,
+      561,
+      591,
+      96,
+      397,
+      419,
+      425,
+      523,
+      627,
+      734,
+      124,
+      236,
+      454,
+      9,
+      345,
+      358,
+      480,
+      575,
+      736,
+      26,
+      34,
+      65,
+      141,
+      152,
+      176,
+      204,
+      293,
+      305,
+      445,
+      282,
+      389,
+      437,
+      619,
+      162,
+      178,
+      311,
+      394,
+      409,
+      483,
+      496,
+      613,
+      717,
+      50,
+      188,
+      216,
+      325,
+      602,
+      699,
+      740,
+      254,
+      258,
+      299,
+      403,
+      495,
+      603,
+      622,
+      40,
+      83,
+      146,
+      150,
+      232,
+      259,
+      341,
+      392,
+      444,
+      522,
+      567,
+      629,
+      647,
+      25,
+      170,
+      205,
+      267,
+      274,
+      307,
+      411,
+      549,
+      576,
+      618,
+      686,
+      760,
+      8,
+      200,
+      255,
+      353,
+      354,
+      413,
+      455,
+      463,
+      544,
+      721,
+      93,
+      171,
+      225,
+      234
+    ]
+  },
+  "runtime_s": 4.635331869125366
+}

stage_0/compressed_variants.json

@@ -0,0 +1,228 @@
+{
+  "results": [
+    {
+      "name": "ref: full 768 ridge",
+      "params": 769,
+      "F1": 0.9598035216331482,
+      "precision": 0.9898664355278015,
+      "recall": 0.9315127730369568,
+      "threshold": null
+    },
+    {
+      "name": "ref: K=92 ridge (cojoint top-92 + bias)",
+      "params": 93,
+      "F1": 0.9463722109794617,
+      "precision": 0.9854528307914734,
+      "recall": 0.9102730751037598,
+      "threshold": null
+    },
+    {
+      "name": "E: ternary weights top-50 pos vs top-50 neg, threshold",
+      "params": 1,
+      "F1": 0.8933987617492676,
+      "precision": 0.9200735092163086,
+      "recall": 0.8682271242141724,
+      "threshold": 31.819643020629883
+    },
+    {
+      "name": "C: threshold(sum top-20 pos \u2212 sum top-20 neg)",
+      "params": 1,
+      "F1": 0.8808632493019104,
+      "precision": 0.8952551484107971,
+      "recall": 0.8669267296791077,
+      "threshold": 24.8664608001709
+    },
+    {
+      "name": "E: ternary weights top-20 pos vs top-20 neg, threshold",
+      "params": 1,
+      "F1": 0.8808632493019104,
+      "precision": 0.8952551484107971,
+      "recall": 0.8669267296791077,
+      "threshold": 24.86646270751953
+    },
+    {
+      "name": "C: threshold(sum top-10 pos \u2212 sum top-10 neg)",
+      "params": 1,
+      "F1": 0.8801606893539429,
+      "precision": 0.9070836901664734,
+      "recall": 0.8547897934913635,
+      "threshold": 22.383634567260742
+    },
+    {
+      "name": "E: ternary weights top-10 pos vs top-10 neg, threshold",
+      "params": 1,
+      "F1": 0.8801606893539429,
+      "precision": 0.9070836901664734,
+      "recall": 0.8547897934913635,
+      "threshold": 22.383630752563477
+    },
+    {
+      "name": "B: threshold(sum top-2 pos dims)",
+      "params": 1,
+      "F1": 0.8780821561813354,
+      "precision": 0.9276410937309265,
+      "recall": 0.8335500359535217,
+      "threshold": 14.303534507751465
+    },
+    {
+      "name": "B: threshold(sum top-20 pos dims)",
+      "params": 1,
+      "F1": 0.8683924674987793,
+      "precision": 0.8716157078742981,
+      "recall": 0.8651928901672363,
+      "threshold": 62.11921310424805
+    },
+    {
+      "name": "C: threshold(sum top-5 pos \u2212 sum top-5 neg)",
+      "params": 1,
+      "F1": 0.8637353181838989,
+      "precision": 0.867512047290802,
+      "recall": 0.8599913120269775,
+      "threshold": 27.896242141723633
+    },
+    {
+      "name": "B: threshold(sum top-10 pos dims)",
+      "params": 1,
+      "F1": 0.8574432134628296,
+      "precision": 0.8325847387313843,
+      "recall": 0.883831799030304,
+      "threshold": 45.166908264160156
+    },
+    {
+      "name": "A: threshold(dim48)",
+      "params": 1,
+      "F1": 0.8451337814331055,
+      "precision": 0.8776844143867493,
+      "recall": 0.8149111270904541,
+      "threshold": 6.3439040184021
+    },
+    {
+      "name": "ref: K=1 ridge (dim48 + bias)",
+      "params": 2,
+      "F1": 0.8285356163978577,
+      "precision": 0.7985524535179138,
+      "recall": 0.8608582615852356,
+      "threshold": null
+    },
+    {
+      "name": "B: threshold(sum top-5 pos dims)",
+      "params": 1,
+      "F1": 0.8214052319526672,
+      "precision": 0.824454128742218,
+      "recall": 0.8183788657188416,
+      "threshold": 37.6262092590332
+    },
+    {
+      "name": "C: threshold(sum top-3 pos \u2212 sum top-3 neg)",
+      "params": 1,
+      "F1": 0.8199912905693054,
+      "precision": 0.8185744881629944,
+      "recall": 0.8214130997657776,
+      "threshold": 18.052114486694336
+    },
+    {
+      "name": "B: threshold(sum top-3 pos dims)",
+      "params": 1,
+      "F1": 0.7914334535598755,
+      "precision": 0.7131432294845581,
+      "recall": 0.8890333771705627,
+      "threshold": 23.4901123046875
+    },
+    {
+      "name": "D: threshold(max top-3 pos \u2212 max top-3 neg)",
+      "params": 1,
+      "F1": 0.7337717413902283,
+      "precision": 0.6004415154457092,
+      "recall": 0.9432163238525391,
+      "threshold": 4.771557331085205
+    },
+    {
+      "name": "D: threshold(max top-5 pos \u2212 max top-5 neg)",
+      "params": 1,
+      "F1": 0.7142618894577026,
+      "precision": 0.5805314779281616,
+      "recall": 0.9280450940132141,
+      "threshold": 6.046311378479004
+    },
+    {
+      "name": "D: threshold(max top-10 pos \u2212 max top-10 neg)",
+      "params": 1,
+      "F1": 0.7102322578430176,
+      "precision": 0.5886545181274414,
+      "recall": 0.8951018452644348,
+      "threshold": 4.326292991638184
+    },
+    {
+      "name": "D: threshold(max top-20 pos \u2212 max top-20 neg)",
+      "params": 1,
+      "F1": 0.7102322578430176,
+      "precision": 0.5886545181274414,
+      "recall": 0.8951018452644348,
+      "threshold": 4.326292991638184
+    }
+  ],
+  "top_pos_dims_30": [
+    48,
+    525,
+    475,
+    645,
+    273,
+    292,
+    158,
+    510,
+    506,
+    337,
+    8,
+    309,
+    267,
+    217,
+    79,
+    13,
+    657,
+    207,
+    722,
+    311,
+    566,
+    278,
+    25,
+    627,
+    511,
+    332,
+    654,
+    719,
+    593,
+    305
+  ],
+  "top_neg_dims_30": [
+    642,
+    224,
+    113,
+    565,
+    49,
+    637,
+    45,
+    520,
+    219,
+    290,
+    529,
+    617,
+    269,
+    745,
+    576,
+    701,
+    105,
+    694,
+    82,
+    283,
+    574,
+    310,
+    613,
+    90,
+    92,
+    650,
+    36,
+    53,
+    396,
+    17
+  ]
+}

stage_0/eval.json

@@ -0,0 +1,14 @@
+{
+  "dataset": "COCO val2017",
+  "n_images": 5000,
+  "task": "image-level person presence (binary)",
+  "positive_rate": 0.539,
+  "protocol": "live Argus forward pass at 768 pixel input, no feature caching",
+  "metrics": {
+    "F1": 0.8886,
+    "precision": 0.9011,
+    "recall": 0.8763,
+    "optimal_threshold": 25.2845
+  },
+  "parity_cached_features": 0.881
+}

stage_0/infer.py

@@ -0,0 +1,63 @@
+"""Reference inference for the Stage 0 baseline.
+
+Loads Argus (EUPE-ViT-B backbone), reads the classifier config, and scores one
+or more images. Prints the raw score and the binary decision.
+
+Usage: python infer.py image1.jpg [image2.jpg ...]
+"""
+import json, sys, os
+import torch
+import torch.nn.functional as F
+from PIL import Image
+import numpy as np
+from transformers import AutoModel
+
+
+def load_classifier(path='classifier.json'):
+    with open(path) as f:
+        return json.load(f)
+
+
+def load_argus(repo_or_path='phanerozoic/argus'):
+    return AutoModel.from_pretrained(repo_or_path, trust_remote_code=True)
+
+
+def preprocess(image_path, resolution=768, device='cuda'):
+    img = Image.open(image_path).convert('RGB').resize((resolution, resolution), Image.BILINEAR)
+    arr = np.asarray(img, dtype=np.uint8).copy()
+    x = torch.from_numpy(arr).permute(2, 0, 1).unsqueeze(0).to(device).float() / 255.0
+    mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1).to(device)
+    std = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1).to(device)
+    return (x - mean) / std
+
+
+@torch.inference_mode()
+def score(model, x, classifier):
+    with torch.autocast('cuda', dtype=torch.bfloat16):
+        out = model.backbone.forward_features(x)
+    patches = out['x_norm_patchtokens'].float().squeeze(0)
+    D = classifier['feature_dim']
+    ln = F.layer_norm(patches, [D])
+    pooled = ln.max(dim=0).values
+    pos = pooled[classifier['pos_dims']].sum()
+    neg = pooled[classifier['neg_dims']].sum()
+    return float((pos - neg).item())
+
+
+def main():
+    if len(sys.argv) < 2:
+        print('usage: python infer.py <image1> [image2 ...]')
+        sys.exit(1)
+    here = os.path.dirname(os.path.abspath(__file__))
+    classifier = load_classifier(os.path.join(here, 'classifier.json'))
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    model = load_argus().to(device).eval()
+    thr = classifier['threshold']
+    for image_path in sys.argv[1:]:
+        x = preprocess(image_path, classifier['input_resolution'], device)
+        s = score(model, x, classifier)
+        print(f'{image_path}  score={s:+.3f}  threshold={thr:+.3f}  person={s > thr}')
+
+
+if __name__ == '__main__':
+    main()

stage_1/README.md

@@ -0,0 +1,5 @@
+# Stage 1: Output-Channel Pruning
+
+Reserved. See repo root README for plan.
+
+Scope: keep only the 100 feature dimensions the Stage 0 classifier reads, remove the remaining 668 output channels from EUPE-ViT-B's final projection. No retraining. Expected to preserve F1 exactly.

stage_2/README.md

@@ -0,0 +1,5 @@
+# Stage 2: Attention-Head Pruning
+
+Reserved. See repo root README for plan.
+
+Scope: measure each of EUPE-ViT-B's 144 attention heads (12 blocks x 12 heads) for contribution to the 100 dims Stage 0 reads. Ablate low-contribution heads.

stage_3/README.md

@@ -0,0 +1,5 @@
+# Stage 3: Depth Reduction
+
+Reserved. See repo root README for plan.
+
+Scope: drop transformer blocks that do not route signal to the 100 Stage 0 dims.

stage_4/README.md

@@ -0,0 +1,5 @@
+# Stage 4: Specialist Backbone
+
+Reserved. See repo root README for plan.
+
+Scope: train a small student network that emits only the 100 target dims of EUPE-ViT-B, supervised against those dims on a large image corpus.

stage_5/README.md

@@ -0,0 +1,5 @@
+# Stage 5: Circuit-Level Synthesis
+
+Reserved. See repo root README for plan.
+
+Scope: synthesize the entire fixed-weight pipeline to gates and dead-code eliminate everything that does not reach the classifier output.