consolidation: RESEARCH_OVERVIEW; figures w/ multi-seed error bars + cross-objective + closed-form law; polished 3 drafts (selector numbers, refs, figure callouts)

RESEARCH_OVERVIEW.md

@@ -0,0 +1,56 @@
+# covtoken — Research Overview
+
+Label-free rare-structure geometry in frozen self-supervised vision transformers, for medical
+imaging. One principle, three papers, four gated programs — all compute run as Hugging Face Jobs,
+all experiments reproducible, labels eval-only throughout.
+
+## The principle (one sentence)
+**Rare/critical structure lives in the mid-layer *concentration subspace* of self-distilled
+foundation models; depth (view-invariance/globalization) erodes it; rank/spanning objectives are
+anti-aligned with it; and it is readable, certifiable, and bounded label-free.**
+
+## The three papers
+1. **Method** — *Where Lesions Live: Label-Free Mid-Layer Lesion Subspaces for Token-Economical
+   Medical Imaging* (`paper/working_draft.md`). The label-free lesion subspace + membership pruning
+   + conformal retention certificate + lesion-routed depth, across CT (MedDINOv3) and ultrasound
+   (DINOv2). Includes the coverage-floor negative result.
+2. **Theory / negative-results** — *Rank-Based Representation Objectives Fail for Rare-Signal
+   Retention: A Mechanism and a Predictive Law* (`paper/paper2_rank_objectives_draft.md`). The
+   exact closed-form law (gap=(m−r)/m, r\*=m), the `A(rank,SNR)` alignment surface, the
+   selection-vs-scaling sharpening. Transferable beyond medicine.
+3. **Mechanism / SPIE** — *Where Lesions Live: Mid-Layer Localization in Frozen ViTs, and Why*
+   (`paper/paper3_midlayer_draft.md`). Mid-layer localization + label-free tail-gap selector +
+   the invariance mechanism + the cross-objective causal-by-comparison result.
+
+## The four programs (gated, HF-Jobs reproducible)
+- **v1 (method)** — `gate_reports/` Gates 0–6. Subspace validity, cross-modality localization,
+  membership > saliency pruning, coverage-floor NEGATIVE, conformal cert (0.978 ≥ 0.90), routed
+  depth (1.6× FLOPs @ 98% sensitivity).
+- **v2 (S1–S5)** — `research_v2/`. S1 the closed-form law (✅), S2 mid-layer + label-free selector
+  (✅), S4 detection viable (◑), S3 precondition-prediction hard (✗), S5 conformal validity transfers (✅).
+- **v3 (F1–F4)** — `research_v3/`. F2a alignment surface (✅ synthetic), F1a selection≠scaling (◑),
+  F2b law needs rank-relative-to-background (◑), F3a invariance mechanism (◑), **F3 cross-objective
+  decisive** (✅ — erosion across DINOv2/supervised/MedDINOv3; MAE not separable).
+- **v4 (G1–G3)** — `research_v4/`. **G1a training-free concentration steering = NEGATIVE** (✅
+  honest): globalization is an *entangling* transformation, not a removable nuisance — reinforces
+  S2 and sharpens F3. G2 (cross-domain universality) and G3 (deployable tool + benchmark) specced,
+  not built.
+
+## Headline numbers (multi-seed where applicable)
+- Mid-layer lesion AUROC peak **0.866 ± 0.010** (block 3), final-layer 0.565; label-free tail-gap
+  selector regret **0.006**.
+- Cross-objective depth-erosion ρ(invariance, AUROC): DINOv2 −0.93, supervised −0.73, MedDINOv3 −0.94.
+- The law is **exact** (gap=(m−r)/m, r\*=m, std 0.0 over 40 seeds).
+- Coverage-floor ablation: 0.22 (floor) vs 0.82 (membership) small-lesion retention.
+
+## What's honest about this body of work
+The clean wins (the law, the mid-layer mechanism + selector, the cross-modality method, the
+conformal certificate) are real and reproducible. The deep extensions (steering, concentration-
+preserving pretraining via reconstruction) returned **negatives that sharpened the science** rather
+than producing new methods — reported as first-class results. The remaining novel lever is G2
+(cross-domain universality), a deliberate new project, not a quick eval.
+
+## Reproduce
+All jobs in `jobs/` (PEP-723 uv scripts) run via `hf jobs uv run --flavor <t4-medium|t4-small|cpu>
+--secrets HF_TOKEN -v hf://buckets/Chucks90/eryon-datasets:/mnt <script>`. Specs in
+`research_specs/`. Figures: `paper/figures/` via `paper/make_figures.py`.

paper/make_figures.py

@@ -13,18 +13,21 @@ plt.rcParams.update({"figure.dpi": 150, "font.size": 11, "axes.grid": True,
 
 
 def fig1_layer():
-    layers = ["final\n(12)", "block 6", "block 4", "block 3"]
-    auroc = [0.565, 0.769, 0.865, 0.871]
-    fig, ax = plt.subplots(figsize=(5, 3.4))
-    ax.plot(range(len(layers)), auroc, "o-", color="#1f77b4", lw=2, ms=8)
+    # full 12-block depth curve with multi-seed (n=3) error bars [research_v3/rigor_results.json]
+    blocks = np.arange(1, 13)
+    mean = [0.8612,0.8641,0.8657,0.8375,0.8127,0.7818,0.7366,0.7059,0.6816,0.6802,0.6607,0.6365]
+    std  = [0.0098,0.0089,0.0098,0.0169,0.0182,0.0163,0.0253,0.0268,0.0302,0.0379,0.0373,0.0264]
+    fig, ax = plt.subplots(figsize=(5.4, 3.4))
+    ax.errorbar(blocks, mean, yerr=std, fmt="o-", color="#1f77b4", lw=2, ms=6,
+                capsize=3, label="density-A (mean ± std, 3 seeds)")
     ax.axhline(0.767, ls="--", color="#d62728", label="attention saliency (0.767)")
-    ax.axhline(0.70, ls=":", color="gray", label="AUROC floor")
-    for i, a in enumerate(auroc):
-        ax.annotate(f"{a:.3f}", (i, a), textcoords="offset points", xytext=(0, 8), ha="center")
-    ax.set_xticks(range(len(layers))); ax.set_xticklabels(layers)
-    ax.set_ylabel("token-level lesion AUROC (LIDC)"); ax.set_ylim(0.5, 0.95)
-    ax.set_title("Finding 1: lesion signal lives mid-layer")
-    ax.legend(loc="lower right", fontsize=9)
+    pk = int(np.argmax(mean))
+    ax.annotate(f"peak block {blocks[pk]}\n{mean[pk]:.3f}±{std[pk]:.3f}", (blocks[pk], mean[pk]),
+                textcoords="offset points", xytext=(14, 6), ha="left", fontsize=8.5, color="#1f77b4")
+    ax.set_xticks(blocks); ax.set_xlabel("transformer block (MedDINOv3)")
+    ax.set_ylabel("token-level lesion AUROC (LIDC)"); ax.set_ylim(0.55, 0.92)
+    ax.set_title("Finding 1: lesion signal peaks mid-layer, erodes with depth")
+    ax.legend(loc="lower left", fontsize=8.5)
     fig.tight_layout(); fig.savefig(OUT / "fig1_layer_ablation.png"); plt.close(fig)
 
 
@@ -102,6 +105,46 @@ def fig5_conformal():
     fig.tight_layout(); fig.savefig(OUT / "fig5_conformal.png"); plt.close(fig)
 
 
-for f in (fig1_layer, fig2_crossmodality, fig3_pruning_gain, fig4_ablation, fig5_conformal):
+def fig6_cross_objective():
+    # F3 decisive: depth-erosion across objectives [research_v3/f3_cross_objective.json]
+    blocks = np.arange(1, 13)
+    dino = [0.862,0.880,0.868,0.807,0.729,0.682,0.701,0.676,0.682,0.653,0.621,0.617]
+    sup  = [0.842,0.840,0.831,0.825,0.816,0.797,0.785,0.791,0.785,0.732,0.681,0.658]
+    mae  = [0.611,0.600,0.596,0.600,0.589,0.590,0.587,0.582,0.577,0.570,0.577,0.568]
+    fig, ax = plt.subplots(figsize=(5.8, 3.6))
+    ax.plot(blocks, dino, "o-", color="#1f77b4", lw=2, ms=5, label="DINOv2 (self-distill)  ρ=−0.93")
+    ax.plot(blocks, sup,  "s-", color="#2ca02c", lw=2, ms=5, label="ViT (supervised)  ρ=−0.73")
+    ax.plot(blocks, mae,  "^--", color="#d62728", lw=2, ms=5, label="MAE (reconstruction)  ρ=+0.06")
+    ax.axhline(0.50, ls=":", color="k", alpha=0.4, label="chance")
+    ax.set_xticks(blocks); ax.set_xlabel("transformer block (natural-trained; eval on CT)")
+    ax.set_ylabel("token-level lesion AUROC"); ax.set_ylim(0.45, 0.92)
+    ax.set_title("Mechanism: localizers erode with depth; MAE never localizes")
+    ax.legend(loc="upper right", fontsize=8)
+    ax.annotate("MAE flat & low\n(not density-separable)", (9, 0.578),
+                textcoords="offset points", xytext=(0, -28), ha="center", fontsize=7.5, color="#d62728")
+    fig.tight_layout(); fig.savefig(OUT / "fig6_cross_objective.png"); plt.close(fig)
+
+
+def fig7_law():
+    # S1 closed-form law: spanning vs concentration retention vs signal rank [research_v2/s1_crossover.json]
+    m = 8; ranks = np.array([1, 2, 3, 4, 5, 6, 7, 8, 10, 12])
+    spanning = np.minimum(ranks, m) / m         # = min(r,m)/m
+    concentration = np.ones_like(spanning, float)
+    fig, ax = plt.subplots(figsize=(5.4, 3.6))
+    ax.plot(ranks, concentration, "o-", color="#1f77b4", lw=2, ms=6, label="concentration (energy/membership)")
+    ax.plot(ranks, spanning, "s-", color="#d62728", lw=2, ms=6, label="spanning (effective-rank/RankMe)")
+    ax.fill_between(ranks, spanning, concentration, color="#d62728", alpha=0.12)
+    ax.axvline(m, ls=":", color="gray"); ax.annotate("crossover r*=m", (m, 0.55),
+               textcoords="offset points", xytext=(6, 0), ha="left", fontsize=8.5, color="gray")
+    ax.annotate("retention gap\n(m−r)/m", (2.0, 0.62), fontsize=8.5, color="#d62728", ha="center")
+    ax.set_xlabel("signal effective rank  r  (lesion concentration → diversity)")
+    ax.set_ylabel("fraction of rare signal retained"); ax.set_ylim(0, 1.08)
+    ax.set_title("The law: rank objectives lose rare/low-rank signal")
+    ax.legend(loc="lower right", fontsize=8.5)
+    fig.tight_layout(); fig.savefig(OUT / "fig7_rank_law.png"); plt.close(fig)
+
+
+for f in (fig1_layer, fig2_crossmodality, fig3_pruning_gain, fig4_ablation, fig5_conformal,
+          fig6_cross_objective, fig7_law):
     f()
 print("figures written:", sorted(p.name for p in OUT.glob("*.png")))

paper/paper2_rank_objectives_draft.md

@@ -39,9 +39,9 @@ tokens (r=1 fully aligned, r=m fully diverse), each high energy in a lesion/anom
 a budget we select tokens by (i) concentration = top-energy `||P_L z||^2`, or (ii) spanning =
 farthest-point / effective-rank maximization in L.
 
-**Result.** spanning retention `= min(r,m)/m`; concentration retention `= 1`; **gap `= (m-r)/m`;
-crossover `r* = m`.** A spanning objective retains a signal only in proportion to its rank; it ties
-concentration only when the signal is fully diverse. (Multi-seed CIs: §5.)
+**Result (Fig. 1).** spanning retention `= min(r,m)/m`; concentration retention `= 1`; **gap
+`= (m-r)/m`; crossover `r* = m`.** A spanning objective retains a signal only in proportion to its
+rank; it ties concentration only when the signal is fully diverse. (Multi-seed: §5.)
 
 **Alignment surface.** Adding an SNR axis yields `A(r, SNR)`: concentration dominates iff the
 signal is concentrated (r<m) AND distinct (high SNR); at low SNR both lose the signal. This
@@ -91,5 +91,6 @@ label-free concentration subspace); this paper is the transferable negative + la
 
 ### Appendix — artifacts
 `research_v2/s1_crossover.json`, `research_v3/f1a_f2a_results.json`, `research_v3/f2b_f3a_results.json`,
-`gate_reports/ablation_floor.json`, `gate_reports/NEGATIVE_RESULT.md`, multi-seed CIs in
-`research_v3/rigor_s1_multiseed.json`. All experiments reproducible as HF Jobs.
+`gate_reports/ablation_floor.json`, `gate_reports/NEGATIVE_RESULT.md`, multi-seed in
+`research_v3/rigor_results.json`. Figure: `paper/figures/fig7_rank_law.png` (the law). All
+experiments reproducible as HF Jobs.

paper/paper3_midlayer_draft.md

@@ -14,8 +14,9 @@ and that the decline with depth is **caused by representation globalization (vie
 loss of spatial information** — a mechanism that holds across self-distillation and supervised
 objectives but is absent for masked reconstruction (which never localizes). A label-free
 density/membership probe over patch tokens localizes lesions; its AUROC on chest CT rises from
-0.565 at the final block to 0.871 at block 3, and a label-free statistic of the membership
-distribution (tail-gap / bimodality) selects a layer within 0.004 AUROC of the masked oracle.
+0.565 at the final block to 0.871 at block 3 (peak 0.866 ± 0.010 over 3 seeds), and the **tail-gap**
+of the membership-score distribution selects a layer within **0.006 AUROC** of the masked oracle
+(multi-seed) — entirely label-free.
 Across three objectives — DINOv2 (self-distillation, peak 0.88), supervised ViT (peak 0.84),
 MedDINOv3 (CT self-distillation) — localizability peaks early/mid and erodes with depth, strongly
 anti-correlated with rising flip-invariance (ρ = −0.73 to −0.94). Masked-reconstruction (MAE)
@@ -26,15 +27,18 @@ label-free.
 ## 1. The finding (depth)
 
 Token-level lesion-membership AUROC by block (LIDC, MedDINOv3): final 0.565 → block 6 0.769 →
-block 4 0.865 → **block 3 0.871**. Final-layer features serve the global self-distillation
-objective; the dense local lesion signal is mid/early. Multi-seed error bars: §5.
+block 4 0.865 → **block 3 0.871** (Fig. 1; multi-seed peak 0.866 ± 0.010, monotone decline to
+0.637 ± 0.026). Final-layer features serve the global self-distillation objective; the dense local
+lesion signal is mid/early.
 
 ## 2. Label-free layer selection
 
-The tail-gap `(q99−q50)/(q50−q01)` and bimodality of the membership-score distribution — computed
-with NO masks — select a layer within **0.004 AUROC** of the mask-derived oracle (bimodality
-ρ=0.69 with the AUROC curve across depth). Excess kurtosis is a poor proxy (picks the worst layer).
-So *where to read* is discoverable without annotation.
+We select the operating layer with NO masks from the shape of the membership-score distribution.
+The **tail-gap** `(q99−q50)/(q50−q01)` is the robust selector: multi-seed regret **0.006 AUROC**
+(max 0.011) versus the mask-derived oracle. A bimodality statistic correlates with the depth curve
+(ρ=0.69) but is less stable across seeds (regret 0.062), and excess kurtosis is a poor proxy (picks
+the worst layer). So *where to read* is discoverable from the score distribution alone, without
+annotation.
 
 ## 3. The mechanism (why mid-layer)
 
@@ -56,7 +60,7 @@ Holding training domain constant (natural-image backbones, evaluated on CT):
 
 Depth-erosion + invariance-coupling hold for both objectives that produce a localizer
 (self-distillation, supervised), and for CT-native MedDINOv3 (ρ=−0.94) — three objectives, same
-law. **MAE is the clarifier:** it is flat *and low* (0.59) — masked reconstruction features are not
+law (Fig. 6). **MAE is the clarifier:** it is flat *and low* (0.59) — masked reconstruction features are not
 density-separable for lesions at any depth, so "no collapse" is trivial (nothing to lose). The
 method needs self-distillation/supervised features; reconstruction is the wrong pretext.
 
@@ -76,4 +80,5 @@ headline selector. Cross-objective curves span all 12 blocks per backbone.
 
 ### Appendix — artifacts
 `research_v2/s2_depth_localizability.json`, `research_v3/f2b_f3a_results.json`,
-`research_v3/f3_cross_objective.json`, `research_v3/rigor_s2_multiseed.json`. HF-Job reproducible.
+`research_v3/f3_cross_objective.json`, `research_v3/rigor_results.json`. Figures: `paper/figures/
+fig1_layer_ablation.png` (depth, error bars), `fig6_cross_objective.png` (mechanism). HF-Job reproducible.

paper/working_draft.md

@@ -99,7 +99,7 @@ Datasets: LIDC-IDRI (lung CT), KiTS23 (kidney CT), MSD Task03 Liver, MSD Task07
 ## 4. Results: the label-free localizer
 
 ### 4.1 Lesion signal lives mid-layer (Finding 1)
-Token-level lesion AUROC by depth (LIDC, density-A):
+Token-level lesion AUROC by depth (LIDC, density-A; Fig. 1):
 
 | layer | final (12) | block 6 | block 4 | block 3 |
 |---|---|---|---|---|
@@ -107,7 +107,9 @@ Token-level lesion AUROC by depth (LIDC, density-A):
 
 Final-layer features are tuned for the global self-distillation objective; the dense local lesion
 signal sits mid/early. We fix block 3 (MedDINOv3) as the operating layer; for DINOv2 the optimum
-is block 8 — backbone-dependent, but always mid/late, never final.
+is block 8 — backbone-dependent, but always mid/late, never final. The curve is multi-seed stable
+(peak 0.866 ± 0.010, n=3), the operating layer is selectable without labels (tail-gap selector
+regret 0.006), and the depth-erosion holds across objectives — see the companion mechanism study.
 
 ### 4.2 Cross-anatomy, cross-modality, cross-backbone localization (Finding 2)
 density-A token-level lesion AUROC, with attention-saliency as the label-free comparator: