Update paper/results to completed heldout80 and publish rolling latest release

Author: Marcel

.github/workflows/ci.yml

@@ -120,7 +120,7 @@ jobs:
             Rolling paper build from `main`.
 
             Source commit: `${{ github.sha }}`
-          prerelease: true
-          make_latest: false
+          prerelease: false
+          make_latest: true
           files: release/main.pdf
           overwrite_files: true

README.md

@@ -19,9 +19,10 @@ Completed evidence so far:
 - `layer16,last1` was a false lead: useful in early pilots, not robust on held-out data.
 - Broader training washed out the site-specific signal.
 - The oracle rows are reference patches, not ceilings.
+- Completed heldout80: `targeted_mlp(layer34,last1)` reached `h8=0.1156`, `h16=0.0641` vs no-patch `h8=0.0172`, `h16=0.0133`.
 
-Active validation:
-- `helmas3n/artifacts/reports/targeted_site_study_v5_holdout80/` is the current held-out80 rerun.
+Latest validation artifact:
+- `helmas3n/artifacts/reports/targeted_site_study_v5_holdout80/` is the completed heldout80 report.
 
 ## Where to start reading
 

helmas3n/README.md

@@ -102,6 +102,14 @@ Each sample stores:
 3. KV uplift baseline + direct comparison support.
 4. Flow refinement placeholder module (not default).
 
+## Latest Results
+
+- Completed heldout80 targeted-site rerun: `artifacts/reports/targeted_site_study_v5_holdout80/`.
+- `targeted_mlp(layer34,last1)` is the strongest learned method:
+  - `h1=0.7750`, `h4=0.2031`, `h8=0.1156`, `h16=0.0641`
+  - no-patch baseline: `h1=0.0375`, `h4=0.0250`, `h8=0.0172`, `h16=0.0133`
+- `targeted_mlp(layer16,last1)` does not generalize on heldout80.
+
 ## Notes
 
 - `configs/extract.yaml` now defines non-identical low/full regime controls via `text_overrides`.

paper/README.md

@@ -34,4 +34,4 @@ The repository tracks paper sources and generated analysis assets. LaTeX tempora
 
 ## Scope of the draft
 
-The draft is written as a living research document. It uses the completed pilot, layer-sweep, suffix-span, and targeted-site evidence as the current result base, while calling out the larger held-out80 validation as the active next gate.
+The draft is written as a living research document. It now includes the completed heldout80 targeted-site validation as part of the core evidence base, alongside the completed pilot, layer-sweep, and suffix-span studies.

paper/figures/layer_sweep_sensitivity.pdf

@@ -1,14 +1,14 @@
-\begin{tabular}{lcccccccc}
+\begin{tabular}{lcccccc}
 \toprule
-Method & Pilot h1 & Pilot h4 & Pilot h8 & Pilot h16 & Held-out h1 & Held-out h4 & Held-out h8 & Held-out h16 \\
+Method & h1 & h4 & h8 & h16 & $\Delta$h8 & $\Delta$h16 \\
 \midrule
-No patch & 0.000000 & 0.006250 & 0.009375 & 0.012500 & 0.000000 & 0.000000 & 0.000000 & 0.006250 \\
-Identity & 0.000000 & 0.006250 & 0.009375 & 0.012500 & 0.000000 & 0.000000 & 0.000000 & 0.006250 \\
-Broad MLP & 0.000000 & 0.012500 & 0.021875 & 0.017188 & 0.000000 & 0.000000 & 0.012500 & 0.012500 \\
-Targeted MLP @ layer16 & 0.000000 & 0.006250 & 0.009375 & 0.010937 & 0.000000 & 0.000000 & 0.000000 & 0.000000 \\
-Targeted MLP @ layer34 & 0.825000 & 0.212500 & 0.115625 & 0.067187 & 0.700000 & 0.175000 & 0.100000 & 0.056250 \\
-Oracle layer16 & 0.000000 & 0.012500 & 0.021875 & 0.017188 & 0.000000 & 0.000000 & 0.000000 & 0.000000 \\
-Oracle layer34 & 0.250000 & 0.068750 & 0.037500 & 0.028125 & 0.400000 & 0.100000 & 0.050000 & 0.031250 \\
-Oracle stride & 0.200000 & 0.081250 & 0.046875 & 0.031250 & 0.300000 & 0.075000 & 0.050000 & 0.025000 \\
+No patch & 0.037500 & 0.025000 & 0.017188 & 0.013281 & 0.000000 & 0.000000 \\
+Identity & 0.037500 & 0.025000 & 0.017188 & 0.013281 & 0.000000 & 0.000000 \\
+Broad MLP & 0.075000 & 0.028125 & 0.018750 & 0.010937 & 0.001562 & -0.002344 \\
+Targeted MLP @ layer16 & 0.012500 & 0.006250 & 0.010937 & 0.007812 & -0.006250 & -0.005469 \\
+Targeted MLP @ layer34 & 0.775000 & 0.203125 & 0.115625 & 0.064062 & 0.098438 & 0.050781 \\
+Reference layer16 & 0.012500 & 0.006250 & 0.014063 & 0.013281 & -0.003125 & 0.000000 \\
+Reference layer34 & 0.387500 & 0.103125 & 0.060937 & 0.036719 & 0.043750 & 0.023438 \\
+Reference stride & 0.237500 & 0.068750 & 0.046875 & 0.028906 & 0.029687 & 0.015625 \\
 \bottomrule
 \end{tabular}