Release FeatureLens v0.7.0

CHANGELOG.md

@@ -1,5 +1,30 @@
 # Changelog
 
+## v0.7.0
+
+### Research-instrument UI cleanup
+- Removed collapsible wrappers from the core **scale dose-response** and **contrastive preference** experiments so section headings are no longer duplicated by accordion titles.
+- Simplified the page header and removed visible footer/redundancy that did not help a reviewer use the tool.
+- Strengthened result-table title and column-header typography.
+- Replaced full-viewport plot stretching with a bounded top-centered **focus view**; exiting focus restores the prior page position.
+- Kept native plot export but rename downloads to descriptive `featurelens_<plot-name>.png` filenames instead of a generic chart name.
+
+### Candidate discovery and causal readiness
+- Reworked live concept-guided discovery around **Balanced selectivity**: `selectivity × target activation rate × log1p(target mean)`. This prevents very large but non-selective SAE coefficients from dominating the exploratory shortlist.
+- Retained **Raw mean difference** as an explicit comparison mode rather than silently changing the old ranking.
+- Candidate discovery now evaluates the current Workbench prompt in the same GPU batch and reports current-prompt maximum activation plus selected-token activation.
+- The candidate selector defaults to the highest-ranked displayed candidate active at the current Workbench token when one exists.
+- Candidate-table row selection is wired directly to the candidate selector, and the reuse action now confirms exactly which downstream feature selectors were updated.
+
+### Lexical / structural specificity
+- Added a **cue × context specificity matrix**: cross several prompt stems with the same completion cues in one batched forward and measure the selected feature at every resulting final token.
+- This extends the single-stem completion-cue test so a response to `is` can be separated from a broader completion-boundary or context-dependent response.
+
+### Controlled data and validation
+- Replaced the French-language control concept one-for-one with a **German-language** concept while preserving 224 balanced discovery prompts and 28 causal tasks.
+- Added regression coverage for German data, balanced/raw candidate ranking, current-Workbench candidate compatibility, cue × context scanning, candidate row selection, descriptive export naming, and focus-position preservation.
+- Kept the actual Gradio `launch()` smoke test, compile gate, release checker, and deferred comprehensive adversarial suite.
+
 ## v0.6.0
 
 ### UX / navigation

README.md

@@ -13,7 +13,7 @@ license: mit
 
 # FeatureLens — Causal Interpretability Workbench
 
-> **v0.6:** a plain-language **Start here** workflow, persistent experiment context, explicit per-experiment feature selectors, plot fullscreen/export controls, concept-guided candidate-feature discovery, and completion-cue sensitivity tests—while retaining the v0.5 causal, geometry, robustness, and prompt-wide evidence tools.
+> **v0.7:** a cleaner research-instrument UI, bounded focus mode for tables/plots, descriptive chart exports, German-language controls, stronger concept-candidate ranking with current-prompt compatibility, click-to-select candidate handoff, and a cue × context specificity diagnostic—while retaining the causal, robustness, geometry, and prompt-wide evidence tools.
 
 **Research question:**
 
@@ -21,7 +21,6 @@ license: mit
 
 FeatureLens uses **Qwen3-1.7B-Base** with **Qwen-Scope residual-stream sparse autoencoders (SAEs)**. It is not an SAE viewer clone: the project separates representation quality, held-out predictiveness, robustness, and causal intervention evidence.
 
-FeatureLens is independent of thesis code and thesis datasets.
 
 ## Evidence ladder
 
@@ -129,7 +128,7 @@ The difference is reported as an **interaction excess**. A non-zero excess means
 
 The **Feature evidence** tab now supports four complementary live questions:
 
-1. **Concept-guided candidate discovery** — choose one of the seven controlled concept groups and rank SAE features by the difference between that concept's prompt-wide mean maximum activation and the mean across the other six groups. The result is a candidate list, **not a semantic label**.
+1. **Concept-guided candidate discovery** — choose one of the seven controlled concept groups and rank SAE features with a selectivity-first score that combines target-vs-other contrast, target activation rate, and log activation magnitude; raw mean-difference ranking remains available for comparison. The result is a candidate list, **not a semantic label**.
 2. **Feature-token activation trace** — show exactly where a selected feature enters the SAE TopK support across every token of the current Workbench prompt.
 3. **Completion-cue sensitivity** — append controlled suffixes/cues to one prompt stem and measure the selected feature at the final token. This is useful for separating concept evidence from lexical/structural completion cues.
 4. **Prompt-wide controlled concept contrast** — evaluate one selected feature on a small balanced batch from:
@@ -138,17 +137,17 @@ The **Feature evidence** tab now supports four complementary live questions:
 - mathematics;
 - positive sentiment;
 - negative sentiment;
-- French language;
+- German language;
 - factual entities;
 - uncertainty.
 
-Each controlled prompt is summarized by the feature's **maximum activation across non-padding tokens**. Candidate discovery ranks by a transparent target-minus-other mean difference and reports activation rates/selectivity alongside it. The controlled contrast reports mean/median prompt-wide maxima, prompt activation rate, mean activation when active, and maximum activation by concept.
+Each controlled prompt is summarized by the feature's **maximum activation across non-padding tokens**. Candidate discovery defaults to **Balanced selectivity** — `positive selectivity × target activation rate × log1p(target mean activation)` — so globally large but weakly selective coefficients do not dominate purely by scale. A **Raw mean difference** mode remains available for comparison. The discovery table also reports current-Workbench prompt/token activation so candidates can be separated from candidates that are immediately causal-testable at the selected location. The controlled contrast reports mean/median prompt-wide maxima, prompt activation rate, mean activation when active, and maximum activation by concept.
 
 Both tools are exploratory live evidence. They never overwrite `Offline concept hint`; the held-out offline AUROC/F1 benchmark remains the stronger feature/concept claim.
 
 ### Contrastive continuation preference
 
-A probability increase for one target can reflect a broad distribution shift rather than selective behavioral control. v0.5 adds a stricter live test with two exact continuations, A and B. FeatureLens scores both teacher-forced under the same SAE intervention and random-control ensemble, then reports the causal change in:
+A probability increase for one target can reflect a broad distribution shift rather than selective behavioral control. FeatureLens includes a stricter live test with two exact continuations, A and B. FeatureLens scores both teacher-forced under the same SAE intervention and random-control ensemble, then reports the causal change in:
 
 ```text
 log P(A) - log P(B)
@@ -173,9 +172,9 @@ The prompt-wide view reduces sensitivity to accidentally comparing semantically
 
 The trajectory view compares SAE reconstruction and sparsity/concentration diagnostics at the same prompt token across layers 4, 14 and 26. It does **not** equate feature IDs across layer-specific SAE dictionaries.
 
-## v0.6 interface and navigation
+## Interface and navigation
 
-v0.6 is designed so a reviewer does not need prior mechanistic-interpretability vocabulary to find the controls:
+The interface is designed so a reviewer does not need prior mechanistic-interpretability vocabulary to find the controls:
 
 - a **Start here** tab explains the evidence workflow and glossary in plain language;
 - a persistent **Current Workbench context** banner states the active prompt, layer and token after inspection;
@@ -183,7 +182,7 @@ v0.6 is designed so a reviewer does not need prior mechanistic-interpretability
 - single-feature causal, dose-response, contrastive-preference, feature-set and evidence panels expose their own editable feature selectors instead of silently borrowing a hidden value;
 - concept-guided discovery gives users a route to candidate features without knowing feature IDs in advance;
 - headings now follow a conventional descending hierarchy and table headers use the same readable base scale as the surrounding interface;
-- every native plot exposes Gradio's **fullscreen** control for zoomed reading and **export** control for PNG output;
+- tables and plots use a bounded **focus mode** instead of stretching to the entire viewport; plot exports are renamed from the generic browser filename to descriptive `featurelens_*.png` names;
 - all action/copy buttons retain one muted-teal visual language and copy actions provide visible confirmation.
 
 The centered 1600 px research canvas, bounded result-table heights, dynamic resize observer and explicit bottom padding from v0.5 remain in place for Hugging Face's embedded Space layout.
@@ -192,7 +191,7 @@ The centered 1600 px research canvas, bounded result-table heights, dynamic resi
 
 Every result table that is intended for analysis has an explicit **Copy table with headers** button. The app serializes the displayed data as TSV before copying, so column names are preserved.
 
-The native Dataframe fullscreen control is retained, while the custom copy action avoids relying on browser text selection and gives immediate visual confirmation when clipboard writing succeeds.
+The Dataframe focus control is retained, while the custom copy action avoids relying on browser text selection and gives immediate visual confirmation when clipboard writing succeeds.
 
 ## Reconstruction-preserving edits
 
@@ -304,7 +303,7 @@ FEATURELENS_CONTRAST_PROMPTS_PER_CONCEPT=4
 
 ## UI design
 
-v0.6 keeps the interface deliberately closer to a conventional research instrument than a rounded dashboard:
+The interface deliberately stays closer to a conventional research instrument than a rounded dashboard:
 
 - Times / Liberation Serif-style formal typography;
 - restrained teal, umber, stone, plum, and muted red chart accents;
@@ -356,7 +355,7 @@ python scripts/ui_smoke.py
 python scripts/release_check.py
 ```
 
-See [`docs/VALIDATION.md`](docs/VALIDATION.md). The v0.6 validation guide uses the **exact labels shown in the UI**, focuses on new-version acceptance/regression tests, and intentionally defers the comprehensive adversarial suite to the final release.
+See [`docs/VALIDATION.md`](docs/VALIDATION.md). The v0.7 validation guide uses the **exact labels shown in the UI**, focuses on new-version acceptance/regression tests, and intentionally defers the comprehensive adversarial suite to the final release.
 
 ## Limitations
 
@@ -372,7 +371,7 @@ See [`docs/VALIDATION.md`](docs/VALIDATION.md). The v0.6 validation guide uses t
 ## Resume-ready description
 
 > **FeatureLens — Causal Interpretability Workbench** | PyTorch, Qwen3, Sparse Autoencoders, Mechanistic Interpretability, Gradio  
-> Built an SAE-based interpretability system for Qwen3-1.7B with held-out concept discovery, concept-guided candidate discovery, token/prompt-wide and completion-cue feature evidence, reconstruction-preserving single and multi-feature interventions, full-continuation and contrastive preference scoring, dose-response analysis, decoder-geometry/non-additivity diagnostics, and norm-matched random-control ensembles.
+> Built an SAE-based interpretability system for Qwen3-1.7B with held-out concept discovery, concept-guided candidate discovery, token/prompt-wide, completion-cue, and cue × context feature evidence, reconstruction-preserving single and multi-feature interventions, full-continuation and contrastive preference scoring, dose-response analysis, decoder-geometry/non-additivity diagnostics, and norm-matched random-control ensembles.
 
 ## Acknowledgements
 

app.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
-import pandas as pd
 import gradio as gr
+import pandas as pd
 
 from featurelens.config import SETTINGS
 from featurelens.hf_runtime import gpu
@@ -47,26 +47,18 @@ CSS = r"""
 .gradio-container h3 { font-size: 1.46rem; line-height: 1.28; }
 .gradio-container h4 { font-size: 1.24rem; line-height: 1.30; }
 .hero {
-  padding: 15px 2px 7px;
+  padding: 14px 2px 10px;
   border-bottom: 1px solid var(--border-color-primary);
-  margin-bottom: 16px;
+  margin-bottom: 14px;
 }
 .hero h1 {
   margin: 0;
-  font-size: 2.25rem;
+  font-size: 2.18rem;
   font-weight: 600;
   letter-spacing: 0;
 }
-.hero .subtitle { margin-top: 4px; font-size: 1.08rem; opacity: .78; }
-.hero .metadata { margin-top: 8px; font-size: .92rem; opacity: .68; letter-spacing: .01em; }
-.research-q {
-  margin: 14px 0 18px;
-  padding: 11px 15px;
-  border-left: 3px solid #708B86;
-  border-top: 1px solid var(--border-color-primary);
-  border-bottom: 1px solid var(--border-color-primary);
-  background: var(--background-fill-secondary);
-}
+.hero .subtitle { margin-top: 4px; font-size: 1.08rem; opacity: .82; }
+.hero .question { margin-top: 6px; font-size: 1rem; opacity: .72; }
 .section-rule {
   margin: 20px 0 11px;
   padding-top: 9px;
@@ -109,14 +101,7 @@ CSS = r"""
 .gradio-container textarea, .gradio-container input { border-radius: 2px !important; }
 .form-note { margin-top: -3px; margin-bottom: 9px; font-size: .92rem; opacity: .72; }
 .wide-table { width: 100% !important; }
-.classic-footer {
-  margin-top: 44px;
-  padding: 20px 2px 58px;
-  border-top: 1px solid var(--border-color-primary);
-  opacity: .70;
-  font-size: .92rem;
-  min-height: 110px;
-}
+.bottom-spacer { height: 110px; width: 100%; }
 .tabs, .tabitem { padding-bottom: 26px !important; }
 .context-card {
   border: 1px solid var(--border-color-primary);
@@ -135,8 +120,53 @@ CSS = r"""
   border-radius: 2px;
 }
 .start-card h3 { margin-top: 0; }
-.experiment-accordion > button { font-size: 1.18rem !important; font-weight: 700 !important; }
 .graph-note { font-size: .95rem; opacity: .72; margin-top: 2px; }
+.result-table .label-wrap,
+.result-table .label-wrap span,
+.result-table label,
+.result-table label span,
+.result-table [data-testid="block-label"],
+.result-table [data-testid="block-label"] span,
+.result-table .block-label,
+.result-table .block-title {
+  font-size: 19px !important;
+  font-weight: 700 !important;
+  line-height: 1.30 !important;
+}
+.result-table table thead th,
+.result-table table thead th *,
+.result-table [role="columnheader"],
+.result-table [role="columnheader"] * {
+  font-size: 16.75px !important;
+  font-weight: 700 !important;
+  line-height: 1.25 !important;
+}
+/* Native fullscreen is converted into a bounded focus surface. Tables may be wider; plots stay
+   narrower so chart labels can be read without scanning across an entire monitor. */
+.fl-plot.fullscreen {
+  top: 6vh !important;
+  left: 50% !important;
+  transform: translateX(-50%) !important;
+  width: min(90vw, 1120px) !important;
+  height: 88vh !important;
+  padding: 18px 22px 22px !important;
+  border-radius: 4px !important;
+  overflow: auto !important;
+  box-shadow: 0 0 0 100vmax rgba(0, 0, 0, .66), 0 14px 40px rgba(0, 0, 0, .35) !important;
+}
+.result-table.fullscreen {
+  top: 5vh !important;
+  left: 50% !important;
+  transform: translateX(-50%) !important;
+  width: min(94vw, 1420px) !important;
+  height: 90vh !important;
+  padding: 18px 22px 24px !important;
+  border-radius: 4px !important;
+  overflow: auto !important;
+  box-shadow: 0 0 0 100vmax rgba(0, 0, 0, .66), 0 14px 40px rgba(0, 0, 0, .35) !important;
+}
+.fl-plot.fullscreen svg { max-width: 100% !important; }
+.candidate-help { opacity: .78; font-size: .97rem; margin-top: -2px; }
 @media (max-width: 900px) {
   .gradio-container { width: 100% !important; padding-left: 12px !important; padding-right: 12px !important; }
 }
@@ -205,6 +235,56 @@ INSTALL_REFLOW_JS = r"""
   const mutation = new MutationObserver(kick);
   mutation.observe(root, {subtree: true, childList: true});
   window.__featurelens_mutation_observer = mutation;
+
+  // Keep the page anchored when entering/exiting native fullscreen. Gradio toggles fullscreen on the
+  // same component node, so FeatureLens can focus the component without navigating away from its source location.
+  document.addEventListener("click", (event) => {
+    const button = event.target && event.target.closest ? event.target.closest("button") : null;
+    if (!button) return;
+    const label = `${button.getAttribute("aria-label") || ""} ${button.getAttribute("title") || ""} ${button.textContent || ""}`.toLowerCase();
+    if (!label.includes("fullscreen")) return;
+    const block = button.closest(".fl-plot, .result-table");
+    if (!block) return;
+    const entering = !block.classList.contains("fullscreen");
+    if (entering) {
+      window.__featurelens_focus_scroll_y = window.scrollY;
+      window.__featurelens_focus_block = block;
+      window.setTimeout(() => {
+        block.scrollTop = 0;
+        if (window.__featurelens_focus_scroll_y != null) window.scrollTo(0, window.__featurelens_focus_scroll_y);
+      }, 80);
+    } else {
+      const y = window.__featurelens_focus_scroll_y;
+      window.setTimeout(() => {
+        if (y != null) window.scrollTo(0, y);
+        window.__featurelens_focus_scroll_y = null;
+        window.__featurelens_focus_block = null;
+      }, 80);
+    }
+  }, true);
+
+  // Gradio's native plot exporter currently emits chart.png. Remember which FeatureLens plot initiated
+  // the export, then rename the transient download anchor before the browser handles it.
+  document.addEventListener("click", (event) => {
+    const button = event.target && event.target.closest ? event.target.closest("button") : null;
+    if (button) {
+      const label = `${button.getAttribute("aria-label") || ""} ${button.getAttribute("title") || ""} ${button.textContent || ""}`.toLowerCase();
+      if (label.includes("export")) {
+        const block = button.closest(".fl-plot");
+        if (block) {
+          const id = block.id || "plot-featurelens-chart";
+          const stem = id.replace(/^plot-/, "").replace(/[^a-z0-9_-]+/gi, "-").replace(/-+/g, "-").replace(/^-|-$/g, "");
+          window.__featurelens_export_name = `featurelens_${stem || "chart"}.png`;
+        }
+      }
+    }
+    const anchor = event.target && event.target.closest ? event.target.closest('a[download="chart.png"]') : null;
+    if (anchor && window.__featurelens_export_name) {
+      anchor.setAttribute("download", window.__featurelens_export_name);
+      window.setTimeout(() => { window.__featurelens_export_name = null; }, 500);
+    }
+  }, true);
+
   kick();
   return [];
 }
@@ -439,12 +519,39 @@ def _discovery_metrics_markdown(result) -> str:
             f"No positively selective candidate features were found for **{result.concept}** at layer "
             f"**{result.layer}** in this small live batch. Increase the sample or try another layer."
         )
+    ranking = (
+        "balanced selectivity × coverage × log-magnitude"
+        if result.ranking_mode == "balanced_selectivity"
+        else "raw target-minus-other mean activation"
+    )
+    context = (
+        f"Current Workbench token **{result.current_token_index}** was checked in the same model batch. "
+        "The default candidate prefers the highest-ranked feature active at that token when one exists."
+        if result.current_context_available
+        else "No Workbench prompt was available, so current-prompt compatibility was not evaluated."
+    )
     return (
         f"Concept **{result.concept}** · layer **{result.layer}** · "
         f"{result.prompts_per_concept} prompts/concept.  \n"
-        f"Showing **{len(result.candidate_ids)}** features ranked by target-minus-other prompt-wide mean activation.  \n\n"
-        "This is **candidate discovery**, not a semantic label. Use the offline held-out AUROC/F1 pipeline "
-        "before describing a feature as concept-selective."
+        f"Showing **{len(result.candidate_ids)}** candidates ranked by **{ranking}**.  \n"
+        f"{context}  \n\n"
+        "This is **candidate discovery**, not a semantic label. The offline held-out AUROC/F1 pipeline remains "
+        "the standard for a concept-selectivity claim."
+    )
+
+
+def _cue_context_metrics_markdown(result) -> str:
+    active = ", ".join(
+        f"{cue}: {count}/{len(result.stems)} contexts"
+        for cue, count in result.cue_active_context_counts.items()
+    )
+    return (
+        f"Feature **{result.feature_id}**, layer **{result.layer}** · active in "
+        f"**{result.active_condition_count}/{result.condition_count}** stem × cue conditions.  \n"
+        f"Cue coverage — {active}.  \n\n"
+        "A cue that activates across unrelated stems looks more lexical/cue-specific; activation only in a subset "
+        "of semantically related stems is more consistent with context-sensitive completion structure. This is still "
+        "a diagnostic, not a semantic label."
     )
 
 
@@ -916,35 +1023,56 @@ def run_contrastive_causal(
 
 
 @gpu(duration=35)
-def run_concept_feature_discovery(concept: str, layer: int, prompts_per_concept: int, top_n: int):
+def run_concept_feature_discovery(
+    concept: str,
+    layer: int,
+    prompts_per_concept: int,
+    top_n: int,
+    ranking_label: str,
+    workbench_prompt: str,
+    workbench_token_index: int,
+):
     try:
+        ranking_mode = (
+            "balanced_selectivity"
+            if ranking_label == "Balanced selectivity"
+            else "raw_mean_difference"
+        )
         result = RUNTIME.concept_feature_discovery(
             concept=concept,
             layer=int(layer),
             prompts_per_concept=int(prompts_per_concept),
             top_n=int(top_n),
+            ranking_mode=ranking_mode,
+            current_text=workbench_prompt,
+            current_token_index=int(workbench_token_index),
         )
         columns = [
             "Rank",
             "Feature id",
+            "Candidate score",
             "Target mean max",
             "Other mean max",
             "Mean difference",
-            "Selectivity score",
+            "Selectivity",
             "Target activation rate",
             "Other activation rate",
-            "Target max activation",
+            "Current prompt max",
+            "Current token activation",
+            "Active at current token",
         ]
         table = pd.DataFrame(result.rows, columns=columns)
-        chart = pd.DataFrame(result.chart_rows, columns=["Feature", "Target − other mean"])
-        chart["Series"] = "Concept contrast"
+        chart = pd.DataFrame(result.chart_rows, columns=["Feature", "Candidate score"])
+        chart["Series"] = "Candidate score"
         choices = [str(fid) for fid in result.candidate_ids]
-        candidate_update = gr.update(choices=choices, value=choices[0] if choices else None)
+        default = str(result.default_candidate_id) if result.default_candidate_id is not None else (choices[0] if choices else None)
+        candidate_update = gr.update(choices=choices, value=default)
         return _discovery_metrics_markdown(result), table, chart, candidate_update, _tsv(table)
     except Exception as exc:
         _raise_ui_error(exc)
 
 
+
 @gpu(duration=25)
 def run_feature_cue_scan(feature_id: str, layer: int, prompt_stem: str, cue_text: str):
     try:
@@ -966,11 +1094,47 @@ def run_feature_cue_scan(feature_id: str, layer: int, prompt_stem: str, cue_text
         _raise_ui_error(exc)
 
 
+@gpu(duration=30)
+def run_feature_cue_context_scan(feature_id: str, layer: int, stems_text: str, cue_text: str):
+    try:
+        if feature_id is None or str(feature_id).strip() == "":
+            raise ValueError("Choose a feature id first.")
+        stems = [line for line in str(stems_text).splitlines() if line.strip()]
+        cues = [line for line in str(cue_text).splitlines() if line.strip()]
+        result = RUNTIME.feature_cue_context_scan(
+            feature_id=int(float(feature_id)),
+            layer=int(layer),
+            stems=stems,
+            cues=cues,
+        )
+        columns = ["Prompt stem", "Cue", "Full prompt", "Final token", "Activation", "Active in TopK"]
+        table = pd.DataFrame(result.rows, columns=columns)
+        chart = pd.DataFrame(result.chart_rows, columns=["Prompt stem", "Cue", "Activation"])
+        return _cue_context_metrics_markdown(result), table, chart, _tsv(table)
+    except Exception as exc:
+        _raise_ui_error(exc)
+
+
+def select_candidate_row(table: pd.DataFrame, evt: gr.SelectData):
+    if table is None or len(table) == 0:
+        return gr.update()
+    index = evt.index
+    row_index = int(index[0] if isinstance(index, (tuple, list)) else index)
+    if row_index < 0 or row_index >= len(table):
+        return gr.update()
+    value = str(int(float(table.iloc[row_index]["Feature id"])))
+    return gr.update(value=value)
+
+
 def use_candidate_feature(candidate_id: str):
     if candidate_id is None or str(candidate_id).strip() == "":
         raise gr.Error("Run concept-guided discovery and choose a candidate first.")
     value = str(int(float(candidate_id)))
-    return value, value, value, value
+    status = (
+        f"**Feature {value} loaded.** It is now selected in Single-feature causal test, "
+        "Scale dose-response, Contrastive preference, and Feature evidence. You can still change any selector independently."
+    )
+    return value, value, value, value, status
 
 
 def mode_help(mode: str):
@@ -990,13 +1154,10 @@ def set_mode_help(mode: str):
 with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_width=True) as demo:
     gr.HTML(
         '<header class="hero">'
-        '<h1>FeatureLens <span style="font-size:.48em;opacity:.58;font-weight:400">v0.6</span></h1>'
+        '<h1>FeatureLens <span style="font-size:.48em;opacity:.58;font-weight:400">v0.7</span></h1>'
         '<div class="subtitle">Causal Interpretability Workbench</div>'
-        '<div class="metadata">Qwen3-1.7B-Base · Qwen-Scope residual SAEs · layers 4 / 14 / 26 · '
-        '32,768 features · TopK=50 · ZeroGPU</div>'
+        '<div class="question">Discover sparse features, test robustness, and separate correlation from causal influence.</div>'
         '</header>'
-        '<div class="research-q"><strong>Research question.</strong> Do sparse features that predict a concept '
-        'also causally influence the model’s behaviour?</div>'
     )
 
     global_context = gr.Markdown(
@@ -1035,7 +1196,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
             "- **Ablate:** remove that feature's current contribution. **Scale:** multiply it.\n"
             "- **Random control:** an equally large residual edit in an unrelated direction.\n"
             "- **Teacher-forced score:** probability assigned to exact continuation text, even if greedy generation does not change.\n\n"
-            "**Tables:** use fullscreen/copy controls. **Plots:** use the top-right fullscreen button to zoom and the export button for PNG."
+            "**Tables:** use the top-right focus control or copy-with-headers action. **Plots:** use the top-right focus control for a bounded reading view; PNG exports use descriptive FeatureLens filenames."
         )
 
     with gr.Tab("Workbench"):
@@ -1052,7 +1213,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     examples=[
                         ["The derivative of x squared is"],
                         ["In Python, reverse a list using"],
-                        ["Je voudrais réserver une table pour"],
+                        ["Ich möchte einen Tisch für zwei reservieren."],
                         ["I am not fully certain, but the answer may be"],
                     ],
                     inputs=[prompt],
@@ -1087,7 +1248,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     label="Strongest active SAE features",
                     wrap=False,
                     max_height=380,
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                 )
                 feature_tsv = gr.Textbox(visible="hidden")
                 feature_copy = _copy_button()
@@ -1097,11 +1258,11 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     y="Activation",
                     color="Series",
                     color_map={"Activation": INK_TEAL},
-                    title="Activation profile",
+                    title="Activation profile", elem_id="plot-activation-profile",
                     x_title="Feature id",
                     y_title="Activation",
                     x_label_angle=-35,
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=330,
                 )
 
@@ -1149,7 +1310,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 token_prob_table = gr.Dataframe(
                     interactive=False,
                     label="Next-token distribution shift",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1158,7 +1319,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 target_token_table = gr.Dataframe(
                     interactive=False,
                     label="Target continuation token-by-token score",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1166,8 +1327,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 target_token_copy = _copy_button()
 
         gr.HTML('<div class="section-rule">III. Single-feature scale dose-response</div>')
-        with gr.Accordion("Scale dose-response", open=False, elem_classes=["experiment-accordion"]):
-            gr.Markdown("### Scale dose-response")
+        with gr.Group():
             gr.Markdown("**Uses current Workbench prompt / layer / token.** Choose the feature for this experiment below.")
             dose_feature_id = gr.Dropdown(
                 choices=[],
@@ -1187,7 +1347,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     dose_table = gr.Dataframe(
                         interactive=False,
                         label="Scale dose-response measurements",
-                        buttons=["fullscreen"],
+                        buttons=["fullscreen"], elem_classes=["result-table"],
                         wrap=False,
                     max_height=380,
                     )
@@ -1199,16 +1359,15 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         y="Δ mean log p/token",
                         color="Series",
                         color_map={"SAE feature": INK_TEAL},
-                        title="Scale dose-response",
+                        title="Scale dose-response", elem_id="plot-scale-dose-response",
                         x_title="Feature multiplier",
                         y_title="Δ mean log p/token",
-                        buttons=["fullscreen", "export"],
+                        buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                         height=330,
                     )
 
         gr.HTML('<div class="section-rule">IV. Contrastive continuation preference</div>')
-        with gr.Accordion("Contrastive causal preference test", open=False, elem_classes=["experiment-accordion"]):
-            gr.Markdown("### Contrastive causal preference test")
+        with gr.Group():
             gr.Markdown("**Uses current Workbench prompt / layer / token.** Select the feature explicitly below.")
             contrastive_feature_id = gr.Dropdown(
                 choices=[],
@@ -1246,7 +1405,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     contrastive_table = gr.Dataframe(
                         interactive=False,
                         label="Contrastive continuation scores",
-                        buttons=["fullscreen"],
+                        buttons=["fullscreen"], elem_classes=["result-table"],
                         wrap=False,
                         max_height=320,
                     )
@@ -1258,10 +1417,10 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         y="A−B sequence log-odds",
                         color="Series",
                         color_map={"Contrastive preference": INK_TEAL},
-                        title="Preference between exact continuations",
+                        title="Preference between exact continuations", elem_id="plot-contrastive-preference",
                         x_title="Execution condition",
                         y_title="Sequence log-odds A−B",
-                        buttons=["fullscreen", "export"],
+                        buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                         height=320,
                     )
 
@@ -1313,7 +1472,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 set_feature_table = gr.Dataframe(
                     interactive=False,
                     label="Joint intervention features",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1323,7 +1482,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 set_target_table = gr.Dataframe(
                     interactive=False,
                     label="Target continuation token-by-token score",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1343,7 +1502,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 set_sweep_table = gr.Dataframe(
                     interactive=False,
                     label="Feature-set size measurements",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1358,10 +1517,10 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         "Top-k SAE ablation": INK_TEAL,
                         "Random signed mean": INK_STONE,
                     },
-                    title="Effect vs feature-set size",
+                    title="Effect vs feature-set size", elem_id="plot-feature-set-size",
                     x_title="Number of jointly ablated features",
                     y_title="Δ mean log p/token",
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=330,
                 )
 
@@ -1379,7 +1538,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 interaction_table = gr.Dataframe(
                     interactive=False,
                     label="Individual and joint ablation measurements",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1391,18 +1550,17 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     y="Δ mean log p/token",
                     color="Series",
                     color_map={"Ablation effect": INK_UMBER},
-                    title="Individual vs joint effect",
+                    title="Individual vs joint effect", elem_id="plot-individual-vs-joint",
                     x_title="Intervention condition",
                     y_title="Δ mean log p/token",
                     x_label_angle=-25,
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=330,
                 )
 
 
         gr.HTML('<div class="section-rule">Decoder geometry</div>')
-        with gr.Accordion("Feature-set decoder geometry", open=False, elem_classes=["experiment-accordion"]):
-            gr.Markdown("### Feature-set decoder geometry")
+        with gr.Group():
             gr.Markdown(
                 "Joint causal effects can reflect both **decoder-direction geometry** and downstream non-linearity. "
                 "This zero-extra-generation diagnostic measures pairwise decoder cosines and compares the actual "
@@ -1417,7 +1575,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     geometry_table = gr.Dataframe(
                         interactive=False,
                         label="Pairwise decoder geometry",
-                        buttons=["fullscreen"],
+                        buttons=["fullscreen"], elem_classes=["result-table"],
                         wrap=False,
                         max_height=340,
                     )
@@ -1429,11 +1587,11 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         y="Decoder cosine",
                         color="Series",
                         color_map={"Decoder cosine": INK_PLUM},
-                        title="Pairwise SAE decoder cosine",
+                        title="Pairwise SAE decoder cosine", elem_id="plot-decoder-geometry",
                         x_title="Feature pair",
                         y_title="Cosine similarity",
                         x_label_angle=-30,
-                        buttons=["fullscreen", "export"],
+                        buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                         height=320,
                     )
 
@@ -1446,15 +1604,16 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
         )
         gr.HTML('<div class="section-rule">A. Concept-guided candidate discovery</div>')
         gr.Markdown(
-            "Don't know which feature id to inspect? Choose a controlled concept and let FeatureLens rank candidate "
-            "features whose **prompt-wide maximum activation** is higher for that concept than for the other six groups. "
-            "This is a live screening step, not a semantic label."
+            "Start from a concept instead of an anonymous feature id. The default **Balanced selectivity** ranking "
+            "downweights globally high-activation features and favors candidates that are both selective and consistently "
+            "active for the target concept. The table also checks whether each candidate is active in the current Workbench "
+            "prompt and selected token, so you can tell whether it is immediately usable for a causal test."
         )
         with gr.Row(equal_height=True):
             discovery_concept = gr.Dropdown(
                 choices=[
                     "code", "mathematics", "positive_sentiment", "negative_sentiment",
-                    "french_language", "factual_entities", "uncertainty"
+                    "german_language", "factual_entities", "uncertainty"
                 ],
                 value="mathematics",
                 label="Target concept",
@@ -1462,25 +1621,46 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
             discovery_layer = gr.Dropdown(choices=list(SETTINGS.layers), value=SETTINGS.layers[1], label="Residual layer")
             discovery_n = gr.Slider(2, 6, value=SETTINGS.contrast_prompts_per_concept, step=1, label="Prompts per concept")
             discovery_top_n = gr.Slider(5, 20, value=12, step=1, label="Candidate features")
+        with gr.Row(equal_height=True):
+            discovery_ranking = gr.Dropdown(
+                choices=["Balanced selectivity", "Raw mean difference"],
+                value="Balanced selectivity",
+                label="Candidate ranking",
+                info="Balanced selectivity is the recommended exploratory ranking; raw mean difference reproduces the earlier scale-dominated ordering.",
+                scale=2,
+            )
+            gr.Markdown(
+                "**Input source:** controlled concept prompts + the current Workbench prompt/token, evaluated in the same GPU batch.",
+                elem_classes=["instrument-note"],
+            )
         discovery_btn = gr.Button("Discover concept-associated candidates", variant="primary", elem_classes=["action-btn"])
         discovery_metrics = gr.Markdown()
         with gr.Row(equal_height=False):
             with gr.Column(scale=3):
                 discovery_table = gr.Dataframe(
-                    interactive=False, label="Candidate feature contrast", buttons=["fullscreen"], wrap=False, max_height=380
+                    interactive=False, label="Candidate feature evidence", buttons=["fullscreen"], elem_classes=["result-table"],
+                    wrap=False, max_height=420
                 )
                 discovery_tsv = gr.Textbox(visible="hidden")
                 discovery_copy = _copy_button()
             with gr.Column(scale=2):
                 discovery_plot = gr.BarPlot(
-                    x="Feature", y="Target − other mean", color="Series",
-                    color_map={"Concept contrast": INK_TEAL}, title="Candidate concept contrast",
-                    x_title="Feature id", y_title="Target − other mean max activation", x_label_angle=-35,
-                    buttons=["fullscreen", "export"], height=330
+                    x="Feature", y="Candidate score", color="Series",
+                    color_map={"Candidate score": INK_TEAL}, title="Candidate evidence score", elem_id="plot-candidate-discovery",
+                    x_title="Feature id", y_title="Exploratory ranking score", x_label_angle=-35,
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"], height=330
                 )
+        gr.Markdown(
+            "Click a row in **Candidate feature evidence** to place that feature in the selector below. "
+            "The selector defaults to the highest-ranked candidate active at the current Workbench token when available.",
+            elem_classes=["candidate-help"],
+        )
         with gr.Row(equal_height=True):
-            discovery_candidate = gr.Dropdown(choices=[], label="Discovered candidate feature id", allow_custom_value=True, scale=3)
-            use_candidate_btn = gr.Button("Use candidate in feature experiments", variant="primary", elem_classes=["action-btn"], scale=2)
+            discovery_candidate = gr.Dropdown(choices=[], label="Selected candidate feature id", allow_custom_value=True, scale=3)
+            use_candidate_btn = gr.Button(
+                "Use selected candidate across feature tests", variant="primary", elem_classes=["action-btn"], scale=2
+            )
+        candidate_use_status = gr.Markdown()
 
         gr.HTML('<div class="section-rule">B. Inspect one feature</div>')
         contrast_location = gr.Markdown(
@@ -1520,7 +1700,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 trace_table = gr.Dataframe(
                     interactive=False,
                     label="Feature activation by prompt token",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=340,
                 )
@@ -1532,11 +1712,11 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     y="Activation",
                     color="Series",
                     color_map={"Feature activation": INK_TEAL},
-                    title="Feature activation across prompt tokens",
+                    title="Feature activation across prompt tokens", elem_id="plot-feature-token-trace",
                     x_title="Prompt token",
                     y_title="Activation",
                     x_label_angle=-35,
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=320,
                 )
 
@@ -1553,17 +1733,64 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
         cue_metrics = gr.Markdown()
         with gr.Row(equal_height=False):
             with gr.Column(scale=3):
-                cue_table = gr.Dataframe(interactive=False, label="Feature response by completion cue", buttons=["fullscreen"], wrap=False, max_height=340)
+                cue_table = gr.Dataframe(interactive=False, label="Feature response by completion cue", buttons=["fullscreen"], elem_classes=["result-table"], wrap=False, max_height=340)
                 cue_tsv = gr.Textbox(visible="hidden")
                 cue_copy = _copy_button()
             with gr.Column(scale=2):
                 cue_plot = gr.BarPlot(
                     x="Cue", y="Activation", color="Series", color_map={"Cue response": INK_UMBER},
-                    title="Completion-cue feature response", x_title="Cue", y_title="Final-token activation",
-                    buttons=["fullscreen", "export"], height=320
+                    title="Completion-cue feature response", elem_id="plot-completion-cue-response", x_title="Cue", y_title="Final-token activation",
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"], height=320
+                )
+
+        gr.HTML('<div class="section-rule">D. Cue × context specificity</div>')
+        gr.Markdown(
+            "The single-stem cue scan can tell you that a feature prefers `is`, but not whether that preference is lexical "
+            "or context-sensitive. This matrix crosses several prompt stems with the same completion cues in one batch."
+        )
+        with gr.Row(equal_height=True):
+            cue_context_stems = gr.Textbox(
+                label="Prompt stems (one per line)",
+                value="The derivative of x squared\nThe capital of Germany\nThe weather today\nMy name",
+                lines=5,
+                scale=3,
+            )
+            cue_context_cues = gr.Textbox(
+                label="Completion cues (one per line)",
+                value="is\n=\n:\nequals\ntherefore",
+                lines=5,
+                scale=2,
+            )
+        cue_context_btn = gr.Button("Run cue × context scan", variant="primary", elem_classes=["action-btn"])
+        cue_context_metrics = gr.Markdown()
+        with gr.Row(equal_height=False):
+            with gr.Column(scale=3):
+                cue_context_table = gr.Dataframe(
+                    interactive=False,
+                    label="Cue × context feature response",
+                    buttons=["fullscreen"],
+                    elem_classes=["result-table"],
+                    wrap=False,
+                    max_height=420,
+                )
+                cue_context_tsv = gr.Textbox(visible="hidden")
+                cue_context_copy = _copy_button()
+            with gr.Column(scale=2):
+                cue_context_plot = gr.BarPlot(
+                    x="Prompt stem",
+                    y="Activation",
+                    color="Cue",
+                    title="Cue response across contexts",
+                    elem_id="plot-cue-context-matrix",
+                    x_title="Prompt stem",
+                    y_title="Final-token activation",
+                    x_label_angle=-25,
+                    buttons=["fullscreen", "export"],
+                    elem_classes=["fl-plot"],
+                    height=340,
                 )
 
-        gr.HTML('<div class="section-rule">D. Controlled concept contrast for selected feature</div>')
+        gr.HTML('<div class="section-rule">E. Controlled concept contrast for selected feature</div>')
         gr.Markdown("### Prompt-wide controlled concept contrast")
         contrast_btn = gr.Button("Run controlled concept contrast", variant="primary", elem_classes=["action-btn"])
         contrast_metrics = gr.Markdown()
@@ -1572,7 +1799,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 contrast_table = gr.Dataframe(
                     interactive=False,
                     label="Feature activation by controlled concept",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1584,11 +1811,11 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     y="Mean prompt-wide max",
                     color="Series",
                     color_map={"Prompt-wide max": INK_BLUEGREY},
-                    title="Prompt-wide controlled concept contrast",
+                    title="Prompt-wide controlled concept contrast", elem_id="plot-controlled-concept-contrast",
                     x_title="Concept",
                     y_title="Mean max activation",
                     x_label_angle=-25,
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=330,
                 )
 
@@ -1633,7 +1860,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 para_table = gr.Dataframe(
                     interactive=False,
                     label="Top-feature overlap at selected tokens",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1645,11 +1872,11 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     y="Activation",
                     color="Prompt",
                     color_map={"Original": INK_TEAL, "Paraphrase": INK_PLUM},
-                    title="Selected-token feature activations",
+                    title="Selected-token feature activations", elem_id="plot-paraphrase-selected-token",
                     x_title="Feature id",
                     y_title="Activation",
                     x_label_angle=-35,
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=330,
                 )
 
@@ -1680,7 +1907,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 trajectory_table = gr.Dataframe(
                     interactive=False,
                     label="Layer diagnostics",
-                    buttons=["fullscreen"],
+                    buttons=["fullscreen"], elem_classes=["result-table"],
                     wrap=False,
                     max_height=380,
                 )
@@ -1696,10 +1923,10 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         "Top-5 mass": INK_UMBER,
                         "Activation entropy": INK_RED,
                     },
-                    title="Representation trajectory",
+                    title="Representation trajectory", elem_id="plot-layer-trajectory",
                     x_title="Layer",
                     y_title="Normalized value",
-                    buttons=["fullscreen", "export"],
+                    buttons=["fullscreen", "export"], elem_classes=["fl-plot"],
                     height=330,
                 )
 
@@ -1708,7 +1935,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
         gr.Markdown(
             "The offline pipeline remains the source of held-out feature/concept AUROC + F1, reconstruction "
             "quality, paraphrase stability, dense residual linear probes, and causal intervention conclusions. "
-            "v0.6 strengthens the live workbench with batch-context null references and random-control ensembles; "
+            "The live workbench uses batch-context null references and random-control ensembles; "
             "the saved report should be generated only from real experiment artifacts."
         )
 
@@ -1729,7 +1956,7 @@ $$h' = h + \sum_{i \in S}\Delta z_i d_i.$$
 
 FeatureLens patches the delta into the **original residual**; it never replaces the residual with the complete SAE reconstruction.
 
-### v0.6 control discipline
+### Control discipline
 
 Batched experiments include an explicit **zero-edit row**. Causal effects are measured against that row rather than a separately executed baseline, which removes batch-vs-single floating-point drift from the measured effect. Random specificity uses an ensemble of norm-matched residual directions rather than one arbitrary seed.
 
@@ -1738,7 +1965,7 @@ Batched experiments include an explicit **zero-edit row**. Causal effects are me
 1. SAE reconstruction quality.
 2. Held-out feature/concept prediction.
 3. Concept-guided candidate discovery and selected-feature concept contrast.
-4. Token-local activation traces and completion-cue sensitivity.
+4. Token-local activation traces, completion-cue sensitivity, and cue × context specificity.
 5. Local and prompt-wide paraphrase robustness.
 6. Single-feature causal intervention and dose-response.
 7. Contrastive continuation preference under intervention.
@@ -1750,11 +1977,7 @@ Association, robustness, geometry, and intervention evidence remain separate cla
 """
         )
 
-    gr.HTML(
-        '<footer class="classic-footer">Built with PyTorch, Transformers, Qwen3-1.7B-Base and Qwen-Scope '
-        'residual-stream SAEs. FeatureLens is independent of thesis code and thesis datasets.<br><br>'
-        'End of workbench.</footer>'
-    )
+    gr.HTML('<div class="bottom-spacer" aria-hidden="true"></div>')
 
     demo.load(fn=None, js=INSTALL_REFLOW_JS, queue=False)
 
@@ -1837,13 +2060,22 @@ Association, robustness, geometry, and intervention evidence remain separate cla
     )
     discovery_btn.click(
         run_concept_feature_discovery,
-        inputs=[discovery_concept, discovery_layer, discovery_n, discovery_top_n],
+        inputs=[
+            discovery_concept, discovery_layer, discovery_n, discovery_top_n, discovery_ranking,
+            prompt, token_index,
+        ],
         outputs=[discovery_metrics, discovery_table, discovery_plot, discovery_candidate, discovery_tsv],
     )
+    discovery_table.select(
+        select_candidate_row,
+        inputs=[discovery_table],
+        outputs=[discovery_candidate],
+        queue=False,
+    )
     use_candidate_btn.click(
         use_candidate_feature,
         inputs=[discovery_candidate],
-        outputs=[feature_id, dose_feature_id, contrastive_feature_id, contrast_feature_id],
+        outputs=[feature_id, dose_feature_id, contrastive_feature_id, contrast_feature_id, candidate_use_status],
         queue=False,
     )
     cue_btn.click(
@@ -1851,6 +2083,11 @@ Association, robustness, geometry, and intervention evidence remain separate cla
         inputs=[contrast_feature_id, contrast_layer, cue_stem, cue_text],
         outputs=[cue_metrics, cue_table, cue_plot, cue_tsv],
     )
+    cue_context_btn.click(
+        run_feature_cue_context_scan,
+        inputs=[contrast_feature_id, contrast_layer, cue_context_stems, cue_context_cues],
+        outputs=[cue_context_metrics, cue_context_table, cue_context_plot, cue_context_tsv],
+    )
     para_btn.click(
         run_paraphrase_compare,
         inputs=[para_a, para_b, para_layer, para_idx_a, para_idx_b, para_top_n],
@@ -1877,6 +2114,7 @@ Association, robustness, geometry, and intervention evidence remain separate cla
         (contrast_copy, contrast_tsv),
         (discovery_copy, discovery_tsv),
         (cue_copy, cue_tsv),
+        (cue_context_copy, cue_context_tsv),
         (para_copy, para_tsv),
         (trajectory_copy, trajectory_tsv),
     ]:
@@ -1889,4 +2127,4 @@ if __name__ == "__main__":
         theme=THEME,
         ssr_mode=False,
         show_error=True,
-    )
+    )

data/causal_tasks.jsonl

@@ -14,10 +14,10 @@
 {"id": 13, "concept": "code", "prompt": "In Python, an exception handler begins with the keyword", "target": " except"}
 {"id": 14, "concept": "code", "prompt": "SQL keyword used to retrieve rows:", "target": " SELECT"}
 {"id": 15, "concept": "code", "prompt": "A Python conditional branch commonly starts with", "target": " if"}
-{"id": 16, "concept": "french_language", "prompt": "Translate 'hello' into French:", "target": " bonjour"}
-{"id": 17, "concept": "french_language", "prompt": "Translate 'thank you' into French:", "target": " merci"}
-{"id": 18, "concept": "french_language", "prompt": "Translate 'yes' into French:", "target": " oui"}
-{"id": 19, "concept": "french_language", "prompt": "Translate 'good evening' into French:", "target": " bonsoir"}
+{"id": 16, "concept": "german_language", "prompt": "Translate 'hello' into German:", "target": " hallo"}
+{"id": 17, "concept": "german_language", "prompt": "Translate 'thank you' into German:", "target": " danke"}
+{"id": 18, "concept": "german_language", "prompt": "Translate 'yes' into German:", "target": " ja"}
+{"id": 19, "concept": "german_language", "prompt": "Translate 'good evening' into German:", "target": " guten Abend"}
 {"id": 20, "concept": "factual_entities", "prompt": "The scientist associated with radium research, Marie", "target": " Curie"}
 {"id": 21, "concept": "factual_entities", "prompt": "The composer of the Fifth Symphony, Ludwig van", "target": " Beethoven"}
 {"id": 22, "concept": "factual_entities", "prompt": "The computer scientist known for the Turing machine, Alan", "target": " Turing"}

data/prompts.jsonl

@@ -126,38 +126,38 @@
 {"id": 125, "concept": "negative_sentiment", "pair_id": "negative_sentiment-14", "variant": 1, "text": "I thought the exhibition was dull and unrewarding."}
 {"id": 126, "concept": "negative_sentiment", "pair_id": "negative_sentiment-15", "variant": 0, "text": "The train journey was stressful and uncomfortable."}
 {"id": 127, "concept": "negative_sentiment", "pair_id": "negative_sentiment-15", "variant": 1, "text": "The trip by train felt frustrating, noisy, and unpleasant."}
-{"id": 128, "concept": "french_language", "pair_id": "french_language-00", "variant": 0, "text": "Bonjour, comment allez-vous aujourd’hui ?"}
-{"id": 129, "concept": "french_language", "pair_id": "french_language-00", "variant": 1, "text": "Salut, comment vas-tu aujourd’hui ?"}
-{"id": 130, "concept": "french_language", "pair_id": "french_language-01", "variant": 0, "text": "Je voudrais réserver une table pour deux personnes."}
-{"id": 131, "concept": "french_language", "pair_id": "french_language-01", "variant": 1, "text": "Puis-je réserver une table pour deux, s’il vous plaît ?"}
-{"id": 132, "concept": "french_language", "pair_id": "french_language-02", "variant": 0, "text": "La bibliothèque ferme à dix-huit heures."}
-{"id": 133, "concept": "french_language", "pair_id": "french_language-02", "variant": 1, "text": "La bibliothèque est ouverte jusqu’à dix-huit heures."}
-{"id": 134, "concept": "french_language", "pair_id": "french_language-03", "variant": 0, "text": "Ce livre raconte une histoire très intéressante."}
-{"id": 135, "concept": "french_language", "pair_id": "french_language-03", "variant": 1, "text": "L’histoire racontée dans ce livre est vraiment intéressante."}
-{"id": 136, "concept": "french_language", "pair_id": "french_language-04", "variant": 0, "text": "Nous allons prendre le train demain matin."}
-{"id": 137, "concept": "french_language", "pair_id": "french_language-04", "variant": 1, "text": "Demain matin, nous voyagerons en train."}
-{"id": 138, "concept": "french_language", "pair_id": "french_language-05", "variant": 0, "text": "Pouvez-vous m’indiquer le chemin vers la gare ?"}
-{"id": 139, "concept": "french_language", "pair_id": "french_language-05", "variant": 1, "text": "Comment puis-je aller jusqu’à la gare ?"}
-{"id": 140, "concept": "french_language", "pair_id": "french_language-06", "variant": 0, "text": "J’aime apprendre de nouvelles langues."}
-{"id": 141, "concept": "french_language", "pair_id": "french_language-06", "variant": 1, "text": "Apprendre des langues nouvelles me plaît beaucoup."}
-{"id": 142, "concept": "french_language", "pair_id": "french_language-07", "variant": 0, "text": "Le temps est magnifique au bord de la mer."}
-{"id": 143, "concept": "french_language", "pair_id": "french_language-07", "variant": 1, "text": "Il fait très beau près de la mer."}
-{"id": 144, "concept": "french_language", "pair_id": "french_language-08", "variant": 0, "text": "Elle prépare le dîner dans la cuisine."}
-{"id": 145, "concept": "french_language", "pair_id": "french_language-08", "variant": 1, "text": "Dans la cuisine, elle est en train de préparer le repas du soir."}
-{"id": 146, "concept": "french_language", "pair_id": "french_language-09", "variant": 0, "text": "Nous avons visité un musée pendant le week-end."}
-{"id": 147, "concept": "french_language", "pair_id": "french_language-09", "variant": 1, "text": "Ce week-end, nous sommes allés voir un musée."}
-{"id": 148, "concept": "french_language", "pair_id": "french_language-10", "variant": 0, "text": "La réunion commence à neuf heures précises."}
-{"id": 149, "concept": "french_language", "pair_id": "french_language-10", "variant": 1, "text": "Le rendez-vous débute exactement à neuf heures."}
-{"id": 150, "concept": "french_language", "pair_id": "french_language-11", "variant": 0, "text": "Mon ordinateur ne fonctionne plus correctement."}
-{"id": 151, "concept": "french_language", "pair_id": "french_language-11", "variant": 1, "text": "J’ai un problème : mon ordinateur marche mal maintenant."}
-{"id": 152, "concept": "french_language", "pair_id": "french_language-12", "variant": 0, "text": "Cette ville possède de nombreux bâtiments historiques."}
-{"id": 153, "concept": "french_language", "pair_id": "french_language-12", "variant": 1, "text": "On trouve beaucoup d’édifices historiques dans cette ville."}
-{"id": 154, "concept": "french_language", "pair_id": "french_language-13", "variant": 0, "text": "Il faut acheter du pain et des légumes."}
-{"id": 155, "concept": "french_language", "pair_id": "french_language-13", "variant": 1, "text": "Nous devons prendre du pain ainsi que des légumes."}
-{"id": 156, "concept": "french_language", "pair_id": "french_language-14", "variant": 0, "text": "Je cherche un appartement près de l’université."}
-{"id": 157, "concept": "french_language", "pair_id": "french_language-14", "variant": 1, "text": "Je voudrais trouver un logement proche de l’université."}
-{"id": 158, "concept": "french_language", "pair_id": "french_language-15", "variant": 0, "text": "Merci beaucoup pour votre aide."}
-{"id": 159, "concept": "french_language", "pair_id": "french_language-15", "variant": 1, "text": "Je vous remercie sincèrement de votre aide."}
+{"id": 128, "concept": "german_language", "pair_id": "german_language-00", "variant": 0, "text": "Guten Tag, wie geht es Ihnen heute?"}
+{"id": 129, "concept": "german_language", "pair_id": "german_language-00", "variant": 1, "text": "Hallo, wie geht es dir heute?"}
+{"id": 130, "concept": "german_language", "pair_id": "german_language-01", "variant": 0, "text": "Ich möchte einen Tisch für zwei Personen reservieren."}
+{"id": 131, "concept": "german_language", "pair_id": "german_language-01", "variant": 1, "text": "Kann ich bitte einen Tisch für zwei reservieren?"}
+{"id": 132, "concept": "german_language", "pair_id": "german_language-02", "variant": 0, "text": "Die Bibliothek schließt um achtzehn Uhr."}
+{"id": 133, "concept": "german_language", "pair_id": "german_language-02", "variant": 1, "text": "Die Bibliothek ist bis achtzehn Uhr geöffnet."}
+{"id": 134, "concept": "german_language", "pair_id": "german_language-03", "variant": 0, "text": "Dieses Buch erzählt eine sehr interessante Geschichte."}
+{"id": 135, "concept": "german_language", "pair_id": "german_language-03", "variant": 1, "text": "Die Geschichte in diesem Buch ist wirklich interessant."}
+{"id": 136, "concept": "german_language", "pair_id": "german_language-04", "variant": 0, "text": "Wir fahren morgen früh mit dem Zug."}
+{"id": 137, "concept": "german_language", "pair_id": "german_language-04", "variant": 1, "text": "Morgen früh werden wir mit dem Zug reisen."}
+{"id": 138, "concept": "german_language", "pair_id": "german_language-05", "variant": 0, "text": "Können Sie mir den Weg zum Bahnhof zeigen?"}
+{"id": 139, "concept": "german_language", "pair_id": "german_language-05", "variant": 1, "text": "Wie komme ich zum Bahnhof?"}
+{"id": 140, "concept": "german_language", "pair_id": "german_language-06", "variant": 0, "text": "Ich lerne gern neue Sprachen."}
+{"id": 141, "concept": "german_language", "pair_id": "german_language-06", "variant": 1, "text": "Neue Sprachen zu lernen macht mir viel Spaß."}
+{"id": 142, "concept": "german_language", "pair_id": "german_language-07", "variant": 0, "text": "Das Wetter am Meer ist wunderschön."}
+{"id": 143, "concept": "german_language", "pair_id": "german_language-07", "variant": 1, "text": "Am Meer ist das Wetter heute sehr schön."}
+{"id": 144, "concept": "german_language", "pair_id": "german_language-08", "variant": 0, "text": "Sie bereitet das Abendessen in der Küche zu."}
+{"id": 145, "concept": "german_language", "pair_id": "german_language-08", "variant": 1, "text": "In der Küche kocht sie gerade das Abendessen."}
+{"id": 146, "concept": "german_language", "pair_id": "german_language-09", "variant": 0, "text": "Wir haben am Wochenende ein Museum besucht."}
+{"id": 147, "concept": "german_language", "pair_id": "german_language-09", "variant": 1, "text": "Am Wochenende waren wir in einem Museum."}
+{"id": 148, "concept": "german_language", "pair_id": "german_language-10", "variant": 0, "text": "Die Besprechung beginnt genau um neun Uhr."}
+{"id": 149, "concept": "german_language", "pair_id": "german_language-10", "variant": 1, "text": "Das Treffen fängt pünktlich um neun Uhr an."}
+{"id": 150, "concept": "german_language", "pair_id": "german_language-11", "variant": 0, "text": "Mein Computer funktioniert nicht mehr richtig."}
+{"id": 151, "concept": "german_language", "pair_id": "german_language-11", "variant": 1, "text": "Mit meinem Computer stimmt etwas nicht; er läuft nicht richtig."}
+{"id": 152, "concept": "german_language", "pair_id": "german_language-12", "variant": 0, "text": "Diese Stadt hat viele historische Gebäude."}
+{"id": 153, "concept": "german_language", "pair_id": "german_language-12", "variant": 1, "text": "In dieser Stadt gibt es zahlreiche historische Bauwerke."}
+{"id": 154, "concept": "german_language", "pair_id": "german_language-13", "variant": 0, "text": "Wir müssen Brot und Gemüse kaufen."}
+{"id": 155, "concept": "german_language", "pair_id": "german_language-13", "variant": 1, "text": "Wir sollten noch Brot sowie Gemüse einkaufen."}
+{"id": 156, "concept": "german_language", "pair_id": "german_language-14", "variant": 0, "text": "Ich suche eine Wohnung in der Nähe der Universität."}
+{"id": 157, "concept": "german_language", "pair_id": "german_language-14", "variant": 1, "text": "Ich möchte eine Unterkunft nahe der Universität finden."}
+{"id": 158, "concept": "german_language", "pair_id": "german_language-15", "variant": 0, "text": "Vielen Dank für Ihre Hilfe."}
+{"id": 159, "concept": "german_language", "pair_id": "german_language-15", "variant": 1, "text": "Ich danke Ihnen herzlich für Ihre Unterstützung."}
 {"id": 160, "concept": "factual_entities", "pair_id": "factual_entities-00", "variant": 0, "text": "Tell me about Marie Curie and her scientific work."}
 {"id": 161, "concept": "factual_entities", "pair_id": "factual_entities-00", "variant": 1, "text": "Summarize the scientific contributions of Marie Curie."}
 {"id": 162, "concept": "factual_entities", "pair_id": "factual_entities-01", "variant": 0, "text": "What is notable about the city of Kyoto?"}

docs/HF_DEPLOY.md

@@ -15,11 +15,11 @@ The live app does not need every SAE layer from the full repository.
 
 ## GPU-decorated actions
 
-Current v0.6 live actions include:
+Current live actions include:
 
 - concept-guided candidate feature discovery and candidate reuse;
 - completion-cue sensitivity scans;
-- plot fullscreen/export controls and a persistent Workbench context banner;
+- bounded table/plot focus controls, descriptive PNG exports, and a persistent Workbench context banner;
 - **Inspect sparse features**;
 - **Run single-feature causal test**;
 - **Run scale dose-response**;
@@ -73,7 +73,7 @@ The app uses:
 - restrained serif typography with normalized control/table sizes;
 - consistent muted-teal action and copy buttons;
 - bounded result-table heights;
-- explicit bottom padding and a visible `End of workbench.` footer;
+- explicit bottom padding with no visible footer clutter;
 - a browser-side ResizeObserver/MutationObserver that requests a resize reflow after dynamic result-height changes.
 
 These changes reduce wasted horizontal space and mitigate the embedded-Space case where the outer page stopped extending after a large dynamic result. Hugging Face still owns the outer embedding frame, so compare with the direct `*.hf.space` URL if the parent page ever behaves differently.

docs/METHODOLOGY.md

@@ -29,7 +29,7 @@ The controlled discovery benchmark contains seven concept groups:
 - mathematics;
 - positive sentiment;
 - negative sentiment;
-- French language;
+- German language;
 - factual entities;
 - uncertainty.
 
@@ -348,9 +348,9 @@ A large point-estimate ratio alone is not sufficient for a strong causal-specifi
 FeatureLens keeps the raw rows even when the resulting conclusion is null, mixed, or contrary to the original hypothesis.
 
 
-## v0.6 live candidate discovery
+## Live candidate discovery
 
-The selected-feature concept contrast asks **given a feature, where does it activate?** v0.6 adds the reverse live question: **given a controlled concept, which SAE features are plausible candidates to investigate?**
+The selected-feature concept contrast asks **given a feature, where does it activate?** FeatureLens also asks the reverse live question: **given a controlled concept, which SAE features are plausible candidates to investigate?**
 
 For a balanced batch with `n` prompts from each of the seven controlled groups, FeatureLens encodes every non-padding token and forms a prompt-wide feature profile by taking the maximum TopK activation of each SAE feature over the prompt. For target concept `c` and feature `f`, it reports:
 
@@ -361,9 +361,9 @@ mean_difference(f) = target_mean(f) - other_mean(f)
 selectivity(f) = mean_difference(f) / (target_mean(f) + other_mean(f) + eps)
 ```
 
-Candidates are ranked by positive `mean_difference`. The table also reports target/other activation rates and target maximum activation. This is intentionally **not a held-out labeler**: the same small batch is used for live screening. Semantic claims still require the offline grouped train/test feature evaluation.
+The default live ranking uses the exploratory score `max(0, selectivity) × target_activation_rate × log1p(target_mean)`, which prevents very large but broadly active SAE coefficients from dominating merely because of scale. A raw positive `mean_difference` ordering remains available for comparison. The table also reports target/other activation rates plus current-Workbench prompt-wide and selected-token activation. This is intentionally **not a held-out labeler**: the same small controlled batch is used for live screening. Semantic claims still require the offline grouped train/test feature evaluation.
 
-## v0.6 completion-cue sensitivity
+## Completion-cue sensitivity
 
 The v0.5 math example showed feature `22632` active only at the final `is` token. That pattern motivates a lexical/structural control. For one prompt stem and a user-supplied list of completion cues, FeatureLens appends each cue, encodes the resulting prompt, and measures the selected feature at the final non-padding token.
 
@@ -375,3 +375,19 @@ The cue scan is intended to distinguish hypotheses such as:
 - broad activation across several continuation cues.
 
 It is a controlled diagnostic only. A cue response cannot establish the feature's complete semantics.
+
+## Cue × context specificity
+
+A single completion-cue scan can show that a feature responds to a token such as `is`, but cannot distinguish a lexical cue feature from a context-dependent completion-boundary feature. v0.7 therefore crosses several prompt stems with the same cue set in one batched forward pass.
+
+For feature $f$, stem $s$, and cue $c$, the diagnostic records the final-token SAE activation $z_f(s+c)$. A cue that activates across unrelated stems is more consistent with lexical/cue specificity; activation restricted to a subset of semantically related stems is more consistent with context-sensitive completion structure. This remains a diagnostic rather than a semantic label.
+
+## Live candidate ranking and causal readiness
+
+Large SAE coefficients can dominate a raw mean-difference ranking even when they are common across many concepts. The default candidate score is therefore
+
+$$
+S_f = \max(0,\mathrm{selectivity}_f) \cdot \mathrm{target\ rate}_f \cdot \log(1 + \mathrm{target\ mean}_f).
+$$
+
+This ranking is still exploratory. It is designed to triage candidates, not replace held-out AUROC/F1. The same forward pass optionally includes the current Workbench prompt, allowing the candidate table to report current-prompt maximum activation and selected-token activation. A candidate can therefore be concept-associated in the live batch but visibly inactive at the current causal location.

docs/VALIDATION.md

@@ -1,8 +1,6 @@
-# FeatureLens v0.6 validation
+# FeatureLens v0.7 validation
 
-This guide tests the **new v0.6 behavior** plus a few high-value regressions. The comprehensive edge/adversarial suite remains deferred until the final hardening release, as agreed.
-
-Use the exact UI labels below.
+This guide tests the **v0.7 changes** plus a small number of high-value regressions. Use the exact UI labels below. The comprehensive edge/adversarial suite remains deferred until final hardening.
 
 ## A. Local release gates
 
@@ -18,7 +16,7 @@ python3 scripts/release_check.py
 Expected:
 
 ```text
-40 passed
+45 passed
 FeatureLens UI launch smoke: PASS
 FeatureLens release check: PASS
   discovery prompts: 224
@@ -26,21 +24,32 @@ FeatureLens release check: PASS
   layers: [4, 14, 26]
   feature-set sizes: [1, 3, 5]
   random controls: 8
-  release: v0.6.0
+  release: v0.7.0
 ```
 
-## B. Navigation / typography / plot controls
+## B. UI cleanup, table typography, focus mode, and export names
+
+1. Open **Workbench**.
+2. Confirm these are normal always-visible sections, **not collapsible dropdowns/accordions**:
+   - **III. Single-feature scale dose-response**
+   - **IV. Contrastive continuation preference**
+3. Confirm there is only one experiment heading before each section. A plot title such as `Scale dose-response` is fine; there should not be a second accordion heading with the same text.
+4. After any result table appears, confirm its table title (for example **Strongest active SAE features**) is visually larger/bolder than ordinary body text, and column headings are clearly readable.
+5. Produce a plot and click its top-right **fullscreen/focus** icon.
+   - The focused plot should open near the top-center of the viewport.
+   - It should be bounded rather than stretching across the entire monitor.
+   - Axis labels should remain readable without needing to pan horizontally across a huge plot.
+6. Exit focus mode. You should return to approximately the same page location you were at before opening it.
+7. Repeat the focus test on one wide result table.
+8. Click the plot **export** icon. Verify the downloaded filename is descriptive, for example:
 
-1. Open **Start here**.
-2. Confirm the three workflow cards and the plain-language glossary are readable without tiny heading text.
-3. Open **Workbench**.
-4. Confirm the hierarchy is visually descending: section title > subsection heading > explanatory body text.
-5. Confirm result-table column headers are not noticeably smaller than normal body text.
-6. After any plot is produced, use its top-right **fullscreen** button. The plot should open large enough for labels to be read comfortably.
-7. Confirm the adjacent **export** control produces a PNG.
-8. Confirm copy buttons still show `✓ Copied with headers` briefly.
+```text
+featurelens_activation-profile.png
+featurelens_scale-dose-response.png
+```
 
-Pass condition: no heading level appears smaller than the body text it introduces, tables remain readable, and every native plot has fullscreen/export controls.
+It should not be the generic `chart.png`.
+9. Click **Copy table with headers** and confirm the temporary `✓ Copied with headers` feedback still appears.
 
 ## C. Establish the shared Workbench context
 
@@ -65,108 +74,141 @@ NMSE: 0.1874
 Top-5 activation mass: 26.7%
 ```
 
-The persistent **Current Workbench context** banner above the tabs must now state the inspected prompt, layer 14, and token 5. Switch to **Feature sets** and **Feature evidence**; the banner should remain visible.
-
-## D. Explicit experiment feature selectors
+The persistent **Current Workbench context** banner must state the inspected prompt, layer 14, and token 5. It should remain visible when switching tabs.
 
-### D1. Scale dose-response
+## D. German-language control regression
 
-Open **Workbench → III. Single-feature scale dose-response → Scale dose-response**.
+Open **Feature evidence → A. Concept-guided candidate discovery** and inspect **Target concept**.
 
-Confirm there is a visible field named:
-
-```text
-Dose-response feature id
-```
+Pass conditions:
 
-It should be populated from the latest inspection but editable independently. Set it to `22632` and run with target continuation `2x` from the single-feature causal panel.
+- `german_language` is available;
+- `french_language` is not available.
 
-The `1.0` multiplier row must remain the exact numerical null:
+The controlled examples in **Workbench** should also include a German example such as:
 
 ```text
-Δ feature coefficient = 0
-Perturbation L2 = 0
-Δ mean log p/token = 0
-Δ sequence log p = 0
-Next-token JS = 0
+Ich möchte einen Tisch für zwei reservieren.
 ```
 
-### D2. Contrastive continuation preference
+No GPU run is required for this check.
 
-Open **Workbench → IV. Contrastive continuation preference → Contrastive causal preference test**.
+## E. Balanced concept-guided candidate discovery
 
-Confirm there is now a visible field named:
+Open **Feature evidence → A. Concept-guided candidate discovery**.
+
+Set:
 
 ```text
-Contrastive feature id
+Target concept: mathematics
+Residual layer: 14
+Prompts per concept: 4
+Candidate features: 12
+Candidate ranking: Balanced selectivity
 ```
 
-Set:
+Keep the Workbench context from Test C and click **Discover concept-associated candidates**.
+
+Verify that **Candidate feature evidence** includes these fields:
 
 ```text
-Contrastive feature id: 22632
-Continuation A (preferred): 2x
-Continuation B (comparison): x
-Contrastive intervention: ablate
+Rank
+Feature id
+Candidate score
+Target mean max
+Other mean max
+Mean difference
+Selectivity
+Target activation rate
+Other activation rate
+Current prompt max
+Current token activation
+Active at current token
 ```
 
-Click **Run contrastive preference test**.
+Interpretation check:
 
-The v0.5 reference result was a causal sequence-log-odds shift of about `-0.0882`; small numerical variation is acceptable. The important regression is that the selected feature is explicit and the panel no longer silently inherits a hidden feature id.
+- `Candidate score` is an exploratory balanced score, not a semantic label.
+- Globally huge but weakly selective features should no longer automatically dominate merely because their raw coefficients are enormous.
+- `Current prompt max` and `Current token activation` show whether a candidate can immediately be tested in the current Workbench context.
+- **Selected candidate feature id** should default to the highest-ranked displayed candidate active at the current Workbench token when such a candidate exists; otherwise it may default to the top-ranked candidate.
 
-## E. Concept-guided candidate feature discovery
+Record the top 12 rows. These are the most scientifically useful new v0.7 results.
 
-Open **Feature evidence → A. Concept-guided candidate discovery**.
+### E2. Raw-ranking comparison
 
-Set:
+Without changing the other fields, set:
 
 ```text
-Target concept: mathematics
-Residual layer: 14
-Prompts per concept: 4
-Candidate features: 12
+Candidate ranking: Raw mean difference
 ```
 
-Click **Discover concept-associated candidates**.
+and rerun **Discover concept-associated candidates**.
+
+This should reproduce the scale-sensitive style of ranking used previously. The purpose of this test is to verify that v0.7 exposes the methodological choice explicitly instead of silently changing it.
+
+You do **not** need to investigate all 12 raw-ranking features causally.
+
+## F. Candidate row selection and explicit reuse
+
+After Test E:
 
-Verify:
+1. Click a row in **Candidate feature evidence**. The row's feature id should appear in **Selected candidate feature id**.
+2. Click **Use selected candidate across feature tests**.
+3. A visible confirmation should say that the feature was loaded.
+4. Verify the same id appears in:
+   - **Workbench → Single feature id**;
+   - **Workbench → Dose-response feature id**;
+   - **Workbench → Contrastive feature id**;
+   - **Feature evidence → Feature id**.
 
-- the result is a table named **Candidate feature contrast**;
-- rows contain `Feature id`, target/other prompt-wide means, mean difference, selectivity score, and activation rates;
-- candidates are ordered by positive `Target − other mean` evidence;
-- the UI calls them **candidates**, not semantic labels;
-- **Discovered candidate feature id** is populated when candidates exist;
-- the plot has fullscreen/export controls.
+This is a state-routing operation and should not consume another GPU call.
 
-An empty candidate list is not itself a software failure if the UI explicitly reports that no positively selective candidate was found. It must not invent a label.
+## G. Completion-cue regression for feature 22632
 
-## F. Reuse a discovered candidate
+Open **Feature evidence → C. Completion-cue sensitivity**.
 
-After Test E, choose any value in **Discovered candidate feature id** and click:
+Set:
 
 ```text
-Use candidate in feature experiments
+Feature id: 22632
+Residual layer: 14
+Prompt stem: The derivative of x squared
+Completion cues:
+is
+=
+:
+equals
+therefore
 ```
 
-Verify that the same id appears in:
+Click **Run completion-cue scan**.
+
+The `is` condition reconstructs the original prompt and should remain approximately:
 
-- **Workbench → Single feature id**;
-- **Workbench → Dose-response feature id**;
-- **Workbench → Contrastive feature id**;
-- **Feature evidence → Feature id**.
+```text
+activation: 32.90625
+Active in TopK: True
+```
 
-This is a UI/state-routing test only; it should not invoke the GPU.
+The previous reference result had zero activation for `=`, `:`, `equals`, and `therefore`. Treat that as a regression reference, not as a hard scientific requirement if tiny numerical differences appear.
 
-## G. Completion-cue sensitivity
+## H. Cue × context specificity
 
-Open **Feature evidence → C. Completion-cue sensitivity**.
+Open **Feature evidence → D. Cue × context specificity**.
 
 Use:
 
 ```text
 Feature id: 22632
 Residual layer: 14
-Prompt stem: The derivative of x squared
+
+Prompt stems:
+The derivative of x squared
+The capital of Germany
+The weather today
+My name
+
 Completion cues:
 is
 =
@@ -175,40 +217,56 @@ equals
 therefore
 ```
 
-Click **Run completion-cue scan**.
+Click **Run cue × context scan**.
 
-The `is` row reconstructs the original prompt `The derivative of x squared is`, so its final-token activation should be approximately `32.90625` and `Active in TopK = True`.
+Expected output:
 
-Record the remaining cue activations. Their scientific purpose is to test whether feature 22632 is strongly tied to a completion cue such as `is`, rather than to mathematics generally. Do **not** require the other cues to be zero; that is an empirical result.
+- 20 rows = 4 stems × 5 cues;
+- columns **Prompt stem**, **Cue**, **Full prompt**, **Final token**, **Activation**, **Active in TopK**;
+- a **Cue response across contexts** plot.
 
-## H. Prompt-wide selected-feature contrast regression
+Scientific question: does feature `22632` fire specifically on the literal `is`, on `is` across many contexts, on several completion-boundary cues, or only in the original mathematics context?
 
-Still in **Feature evidence**, use feature `22632`, layer `14`, 4 prompts/concept and click **Run controlled concept contrast**.
+Send the active rows and the per-cue context summary. This is the other key new v0.7 result.
 
-Reference v0.5 behavior:
+## I. Prompt-wide selected-feature contrast with German control
+
+Still in **Feature evidence**, set:
 
 ```text
-active in 7/28 prompts
-uncertainty mean ≈ 7.55
-mathematics mean ≈ 5.42
-french_language mean = 0
+Feature id: 22632
+Residual layer: 14
+Prompts per concept: 4
 ```
 
-Small numerical variation is acceptable. This verifies the existing prompt-wide selected-feature scan remains unchanged while the new candidate-discovery scan answers the reverse question: *given a concept, which features should I investigate?*
+Click **Run controlled concept contrast**.
+
+The result should contain `german_language` rather than `french_language` while preserving seven balanced concepts and 28 sampled prompts total.
+
+Do not expect the exact old French-based concept means to remain identical because one control group has intentionally changed.
 
-## I. Quick existing regressions
+## J. High-value existing regressions
 
-You do not need to resend full tables unless something changes unexpectedly.
+Only confirm these run; do not resend full tables unless something changes unexpectedly.
 
-1. **Paraphrase robustness**: default pair should remain roughly selected-token cosine `0.301`, prompt-wide cosine `0.981`.
-2. **Identical-prompt control**: all four robustness metrics should remain `1.000`.
-3. **Layer trajectory**: layer-14 reconstruction should remain the weakest of layers 4/14/26 for the default prompt.
-4. **Feature sets → Set-size sensitivity → Run 1/3/5-feature ablation sweep**: this is the “top-k sweep.” It is **not** the basic Workbench inspection. Only confirm the experiment still runs and produces k = 1, 3, 5 rows.
-5. Scroll all the way to **End of workbench** after dynamic outputs expand; no tab switch should be necessary.
+1. **Workbench → III. Single-feature scale dose-response** with feature `22632`, target `2x`:
+   - multiplier `1.0` must remain the exact same-batch numerical null with all causal deltas equal to zero.
+2. **Paraphrase robustness** default pair:
+   - selected-token cosine should remain around `0.301`;
+   - prompt-wide cosine should remain around `0.981`.
+3. **Paraphrase robustness** identical-prompt control:
+   - all four robustness metrics should remain `1.000`.
+4. **Layer trajectory**:
+   - layer 14 should remain the weakest reconstruction of layers 4/14/26 on the default prompt.
+5. **Feature sets → Set-size sensitivity → Run 1/3/5-feature ablation sweep**:
+   - this is the top-k causal sweep;
+   - it is **not** the Workbench `Strongest active SAE features` table;
+   - confirm k = 1, 3, 5 rows appear.
+6. After dynamic results expand, scroll to the actual bottom without switching tabs. The v0.5 reflow fix should remain intact.
 
-## J. Deferred final hardening
+## K. Deferred final hardening
 
-Do not spend ZeroGPU quota on the complete adversarial suite yet. Keep these for the final release:
+Do not spend ZeroGPU quota on the complete adversarial suite yet. Preserve these for the final hardening release:
 
 - empty prompts;
 - out-of-range token indices;
@@ -221,4 +279,4 @@ Do not spend ZeroGPU quota on the complete adversarial suite yet. Keep these for
 - long-prompt truncation;
 - rapid repeated clicks / queue behavior;
 - narrow/mobile-width layout;
-- final table-copy and plot fullscreen/export sweep.
+- final copy/focus/export sweep.

experiments/build_dataset.py

@@ -79,23 +79,23 @@ PAIRS: dict[str, list[tuple[str, str]]] = {
         ('The museum exhibition was painfully boring.', 'I thought the exhibition was dull and unrewarding.'),
         ('The train journey was stressful and uncomfortable.', 'The trip by train felt frustrating, noisy, and unpleasant.'),
     ],
-    'french_language': [
-        ('Bonjour, comment allez-vous aujourd’hui ?', 'Salut, comment vas-tu aujourd’hui ?'),
-        ('Je voudrais réserver une table pour deux personnes.', 'Puis-je réserver une table pour deux, s’il vous plaît ?'),
-        ('La bibliothèque ferme à dix-huit heures.', 'La bibliothèque est ouverte jusqu’à dix-huit heures.'),
-        ('Ce livre raconte une histoire très intéressante.', 'L’histoire racontée dans ce livre est vraiment intéressante.'),
-        ('Nous allons prendre le train demain matin.', 'Demain matin, nous voyagerons en train.'),
-        ('Pouvez-vous m’indiquer le chemin vers la gare ?', 'Comment puis-je aller jusqu’à la gare ?'),
-        ('J’aime apprendre de nouvelles langues.', 'Apprendre des langues nouvelles me plaît beaucoup.'),
-        ('Le temps est magnifique au bord de la mer.', 'Il fait très beau près de la mer.'),
-        ('Elle prépare le dîner dans la cuisine.', 'Dans la cuisine, elle est en train de préparer le repas du soir.'),
-        ('Nous avons visité un musée pendant le week-end.', 'Ce week-end, nous sommes allés voir un musée.'),
-        ('La réunion commence à neuf heures précises.', 'Le rendez-vous débute exactement à neuf heures.'),
-        ('Mon ordinateur ne fonctionne plus correctement.', 'J’ai un problème : mon ordinateur marche mal maintenant.'),
-        ('Cette ville possède de nombreux bâtiments historiques.', 'On trouve beaucoup d’édifices historiques dans cette ville.'),
-        ('Il faut acheter du pain et des légumes.', 'Nous devons prendre du pain ainsi que des légumes.'),
-        ('Je cherche un appartement près de l’université.', 'Je voudrais trouver un logement proche de l’université.'),
-        ('Merci beaucoup pour votre aide.', 'Je vous remercie sincèrement de votre aide.'),
+    'german_language': [
+        ('Guten Tag, wie geht es Ihnen heute?', 'Hallo, wie geht es dir heute?'),
+        ('Ich möchte einen Tisch für zwei Personen reservieren.', 'Kann ich bitte einen Tisch für zwei reservieren?'),
+        ('Die Bibliothek schließt um achtzehn Uhr.', 'Die Bibliothek ist bis achtzehn Uhr geöffnet.'),
+        ('Dieses Buch erzählt eine sehr interessante Geschichte.', 'Die Geschichte in diesem Buch ist wirklich interessant.'),
+        ('Wir fahren morgen früh mit dem Zug.', 'Morgen früh werden wir mit dem Zug reisen.'),
+        ('Können Sie mir den Weg zum Bahnhof zeigen?', 'Wie komme ich zum Bahnhof?'),
+        ('Ich lerne gern neue Sprachen.', 'Neue Sprachen zu lernen macht mir viel Spaß.'),
+        ('Das Wetter am Meer ist wunderschön.', 'Am Meer ist das Wetter heute sehr schön.'),
+        ('Sie bereitet das Abendessen in der Küche zu.', 'In der Küche kocht sie gerade das Abendessen.'),
+        ('Wir haben am Wochenende ein Museum besucht.', 'Am Wochenende waren wir in einem Museum.'),
+        ('Die Besprechung beginnt genau um neun Uhr.', 'Das Treffen fängt pünktlich um neun Uhr an.'),
+        ('Mein Computer funktioniert nicht mehr richtig.', 'Mit meinem Computer stimmt etwas nicht; er läuft nicht richtig.'),
+        ('Diese Stadt hat viele historische Gebäude.', 'In dieser Stadt gibt es zahlreiche historische Bauwerke.'),
+        ('Wir müssen Brot und Gemüse kaufen.', 'Wir sollten noch Brot sowie Gemüse einkaufen.'),
+        ('Ich suche eine Wohnung in der Nähe der Universität.', 'Ich möchte eine Unterkunft nahe der Universität finden.'),
+        ('Vielen Dank für Ihre Hilfe.', 'Ich danke Ihnen herzlich für Ihre Unterstützung.'),
     ],
     'factual_entities': [
         ('Tell me about Marie Curie and her scientific work.', 'Summarize the scientific contributions of Marie Curie.'),
@@ -152,10 +152,10 @@ CAUSAL_TASKS = [
     {'concept': 'code', 'prompt': 'In Python, an exception handler begins with the keyword', 'target': ' except'},
     {'concept': 'code', 'prompt': 'SQL keyword used to retrieve rows:', 'target': ' SELECT'},
     {'concept': 'code', 'prompt': 'A Python conditional branch commonly starts with', 'target': ' if'},
-    {'concept': 'french_language', 'prompt': "Translate 'hello' into French:", 'target': ' bonjour'},
-    {'concept': 'french_language', 'prompt': "Translate 'thank you' into French:", 'target': ' merci'},
-    {'concept': 'french_language', 'prompt': "Translate 'yes' into French:", 'target': ' oui'},
-    {'concept': 'french_language', 'prompt': "Translate 'good evening' into French:", 'target': ' bonsoir'},
+    {'concept': 'german_language', 'prompt': "Translate 'hello' into German:", 'target': ' hallo'},
+    {'concept': 'german_language', 'prompt': "Translate 'thank you' into German:", 'target': ' danke'},
+    {'concept': 'german_language', 'prompt': "Translate 'yes' into German:", 'target': ' ja'},
+    {'concept': 'german_language', 'prompt': "Translate 'good evening' into German:", 'target': ' guten Abend'},
     {'concept': 'factual_entities', 'prompt': 'The scientist associated with radium research, Marie', 'target': ' Curie'},
     {'concept': 'factual_entities', 'prompt': 'The composer of the Fifth Symphony, Ludwig van', 'target': ' Beethoven'},
     {'concept': 'factual_entities', 'prompt': 'The computer scientist known for the Turing machine, Alan', 'target': ' Turing'},

experiments/collect_activations.py

@@ -40,6 +40,19 @@ def _build_sparse(encodings: list, n_rows: int, width: int) -> sp.csr_matrix:
     return sp.csr_matrix((values, (row_ids, col_ids)), shape=(n_rows, width), dtype=np.float32)
 
 
+def _make_capture_hook(
+    captured: dict[int, torch.Tensor],
+    layer: int,
+):
+    """Bind the capture dictionary and layer before registering the hook."""
+
+    def hook(_module, _inputs, output):
+        hidden = output[0] if isinstance(output, tuple) else output
+        captured[layer] = hidden.detach()
+
+    return hook
+
+
 @torch.inference_mode()
 def main() -> None:
     args = parse_args()
@@ -85,15 +98,12 @@ def main() -> None:
         captured: dict[int, torch.Tensor] = {}
         handles = []
 
-        def make_hook(layer: int):
-            def hook(_module, _inputs, output):
-                hidden = output[0] if isinstance(output, tuple) else output
-                captured[layer] = hidden.detach()
-
-            return hook
-
         for layer in args.layers:
-            handles.append(model.model.layers[layer].register_forward_hook(make_hook(layer)))
+            handles.append(
+                model.model.layers[layer].register_forward_hook(
+                    _make_capture_hook(captured, layer)
+                )
+            )
         model(**batch, use_cache=False)
         for handle in handles:
             handle.remove()
@@ -148,4 +158,4 @@ def main() -> None:
 
 
 if __name__ == '__main__':
-    main()
+    main()

experiments/run_causal.py

@@ -54,6 +54,37 @@ def replace_hidden(output, hidden):
     return (hidden, *output[1:]) if isinstance(output, tuple) else hidden
 
 
+def _make_capture_hook(capture: dict):
+    """Bind a per-task capture dictionary before registering the hook."""
+
+    def capture_hook(_module, _inp, output):
+        if 'hidden' not in capture:
+            capture['hidden'] = hidden_from_output(output).detach()
+
+    return capture_hook
+
+
+def _make_batch_edit_hook(
+    applied: dict[str, bool],
+    prompt_len: int,
+    deltas: torch.Tensor,
+):
+    """Bind per-task edit state so hooks cannot capture a later loop iteration."""
+
+    def batch_edit_hook(_module, _inp, output):
+        if applied['done']:
+            return output
+        hidden = hidden_from_output(output)
+        modified = hidden.clone()
+        modified[:, prompt_len - 1, :] = (
+            modified[:, prompt_len - 1, :] + deltas.to(hidden.device, hidden.dtype)
+        )
+        applied['done'] = True
+        return replace_hidden(output, modified)
+
+    return batch_edit_hook
+
+
 def append_target(inputs: dict[str, torch.Tensor], target_ids: list[int]) -> dict[str, torch.Tensor]:
     prompt_ids = inputs['input_ids']
     target = torch.tensor(target_ids, dtype=prompt_ids.dtype, device=prompt_ids.device).unsqueeze(0)
@@ -121,12 +152,9 @@ def main() -> None:
         target_ids = [int(x) for x in target_ids]
         full_inputs = append_target(prompt_inputs, target_ids)
         capture: dict = {}
-
-        def capture_hook(_module, _inp, output):
-            if 'hidden' not in capture:
-                capture['hidden'] = hidden_from_output(output).detach()
-
-        handle = model.model.layers[layer].register_forward_hook(capture_hook)
+        handle = model.model.layers[layer].register_forward_hook(
+            _make_capture_hook(capture)
+        )
         single_baseline_out = model(**full_inputs, use_cache=False)
         handle.remove()
         single_baseline_logits = single_baseline_out.logits[0]
@@ -178,19 +206,9 @@ def main() -> None:
         deltas = torch.stack([zero, *[item[4] for item in condition_meta]], dim=0)
         repeated = {key: value.repeat(deltas.shape[0], 1) for key, value in full_inputs.items()}
         applied = {'done': False}
-
-        def batch_edit_hook(_module, _inp, output):
-            if applied['done']:
-                return output
-            hidden = hidden_from_output(output)
-            modified = hidden.clone()
-            modified[:, prompt_len - 1, :] = (
-                modified[:, prompt_len - 1, :] + deltas.to(hidden.device, hidden.dtype)
-            )
-            applied['done'] = True
-            return replace_hidden(output, modified)
-
-        hook = model.model.layers[layer].register_forward_hook(batch_edit_hook)
+        hook = model.model.layers[layer].register_forward_hook(
+            _make_batch_edit_hook(applied, prompt_len, deltas)
+        )
         edited_out = model(**repeated, use_cache=False)
         hook.remove()
 
@@ -210,7 +228,7 @@ def main() -> None:
         baseline_rank = int((baseline_next > baseline_next[target_id]).sum().item()) + 1
         baseline_top1 = int(torch.argmax(baseline_next).item())
 
-        for row_idx, (intervention_name, condition, control_id, spec, applied_delta, delta_activation) in enumerate(
+        for row_idx, (intervention_name, condition, control_id, _spec, applied_delta, delta_activation) in enumerate(
             condition_meta,
             start=1,
         ):

experiments/run_feature_sets.py

@@ -9,7 +9,11 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from experiments.common import ARTIFACT_DIR, DATA_DIR, load_jsonl, set_seed
 from featurelens.config import SETTINGS
-from featurelens.interventions import InterventionSpec, joint_residual_delta, normalized_random_control
+from featurelens.interventions import (
+    InterventionSpec,
+    joint_residual_delta,
+    normalized_random_control,
+)
 from featurelens.metrics import js_divergence_from_logits, sequence_logprob_summary
 from featurelens.sae import SAEStore
 from featurelens.selection import load_feature_sets
@@ -34,6 +38,37 @@ def replace_hidden(output, hidden):
     return (hidden, *output[1:]) if isinstance(output, tuple) else hidden
 
 
+def _make_capture_hook(capture: dict):
+    """Bind a per-task capture dictionary before registering the hook."""
+
+    def capture_hook(_module, _inp, output):
+        if 'hidden' not in capture:
+            capture['hidden'] = hidden_from_output(output).detach()
+
+    return capture_hook
+
+
+def _make_batch_edit_hook(
+    applied: dict[str, bool],
+    prompt_len: int,
+    deltas: torch.Tensor,
+):
+    """Bind per-task edit state so hooks cannot capture a later loop iteration."""
+
+    def edit_hook(_module, _inp, output):
+        if applied['done']:
+            return output
+        hidden = hidden_from_output(output)
+        modified = hidden.clone()
+        modified[:, prompt_len - 1, :] = (
+            modified[:, prompt_len - 1, :] + deltas.to(hidden.device, hidden.dtype)
+        )
+        applied['done'] = True
+        return replace_hidden(output, modified)
+
+    return edit_hook
+
+
 def append_target(inputs: dict[str, torch.Tensor], target_ids: list[int]) -> dict[str, torch.Tensor]:
     prompt_ids = inputs['input_ids']
     target = torch.tensor(target_ids, dtype=prompt_ids.dtype, device=prompt_ids.device).unsqueeze(0)
@@ -103,12 +138,9 @@ def main() -> None:
         target_ids = [int(x) for x in target_ids]
         full_inputs = append_target(prompt_inputs, target_ids)
         capture: dict = {}
-
-        def capture_hook(_module, _inp, output):
-            if 'hidden' not in capture:
-                capture['hidden'] = hidden_from_output(output).detach()
-
-        handle = model.model.layers[layer].register_forward_hook(capture_hook)
+        handle = model.model.layers[layer].register_forward_hook(
+            _make_capture_hook(capture)
+        )
         single_baseline_out = model(**full_inputs, use_cache=False)
         handle.remove()
         single_logits = single_baseline_out.logits[0]
@@ -146,19 +178,9 @@ def main() -> None:
         deltas = torch.stack([zero, *[item[4] for item in condition_meta]], dim=0)
         repeated = {key: value.repeat(deltas.shape[0], 1) for key, value in full_inputs.items()}
         applied = {'done': False}
-
-        def edit_hook(_module, _inp, output):
-            if applied['done']:
-                return output
-            hidden = hidden_from_output(output)
-            modified = hidden.clone()
-            modified[:, prompt_len - 1, :] = (
-                modified[:, prompt_len - 1, :] + deltas.to(hidden.device, hidden.dtype)
-            )
-            applied['done'] = True
-            return replace_hidden(output, modified)
-
-        hook = model.model.layers[layer].register_forward_hook(edit_hook)
+        hook = model.model.layers[layer].register_forward_hook(
+            _make_batch_edit_hook(applied, prompt_len, deltas)
+        )
         edited_out = model(**repeated, use_cache=False)
         hook.remove()
 
@@ -223,4 +245,4 @@ def main() -> None:
 
 
 if __name__ == '__main__':
-    main()
+    main()

featurelens/interventions.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
-from dataclasses import dataclass
 from collections.abc import Sequence
+from dataclasses import dataclass
 
 import torch
 

featurelens/runtime.py

@@ -2,13 +2,13 @@ from __future__ import annotations
 
 import hashlib
 import html
+import json
 import math
 import os
-import json
-from pathlib import Path
 from collections.abc import Iterator, Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
+from pathlib import Path
 
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -233,9 +233,13 @@ class ConceptFeatureDiscoveryResult:
     layer: int
     prompts_per_concept: int
     top_n: int
+    ranking_mode: str
     rows: list[list[object]]
     chart_rows: list[list[object]]
     candidate_ids: list[int]
+    default_candidate_id: int | None
+    current_context_available: bool
+    current_token_index: int | None
 
 
 @dataclass
@@ -249,6 +253,19 @@ class FeatureCueScanResult:
     cue_count: int
 
 
+@dataclass
+class FeatureCueContextResult:
+    feature_id: int
+    layer: int
+    stems: list[str]
+    cues: list[str]
+    rows: list[list[object]]
+    chart_rows: list[list[object]]
+    active_condition_count: int
+    condition_count: int
+    cue_active_context_counts: dict[str, int]
+
+
 @dataclass
 class ParaphraseResult:
     tokens_a: list[str]
@@ -474,7 +491,7 @@ class FeatureLensRuntime:
 
     @staticmethod
     def _control_seed(text: str, layer: int, key: str, mode: str, coefficient: float) -> int:
-        payload = f'{text}\0{layer}\0{key}\0{mode}\0{coefficient:.8g}'.encode('utf-8')
+        payload = f'{text}\0{layer}\0{key}\0{mode}\0{coefficient:.8g}'.encode()
         return int.from_bytes(hashlib.sha256(payload).digest()[:4], 'big', signed=False)
 
     @staticmethod
@@ -1778,8 +1795,18 @@ class FeatureLensRuntime:
         layer: int,
         prompts_per_concept: int | None = None,
         top_n: int = 12,
+        ranking_mode: str = 'balanced_selectivity',
+        current_text: str | None = None,
+        current_token_index: int = -1,
     ) -> ConceptFeatureDiscoveryResult:
-        """Find SAE features whose prompt-wide activation is higher for one controlled concept."""
+        """Find exploratory concept candidates and show whether they are usable in the current Workbench context.
+
+        ``balanced_selectivity`` downweights globally high-activation features by combining target selectivity,
+        target coverage, and target activation magnitude. ``raw_mean_difference`` preserves the simpler raw
+        mean-difference ranking for comparison. If a Workbench prompt is supplied it is appended to the same
+        model batch, so
+        current-prompt compatibility does not require another forward pass.
+        """
         self.ensure_ready(preload_saes=False)
         assert self.model is not None and self.tokenizer is not None and self.sae_store is not None
         concept = str(concept).strip()
@@ -1791,12 +1818,20 @@ class FeatureLensRuntime:
         top_n = int(top_n)
         if top_n < 1 or top_n > 25:
             raise ValueError('Number of candidate features must be between 1 and 25.')
+        if ranking_mode not in {'balanced_selectivity', 'raw_mean_difference'}:
+            raise ValueError("ranking_mode must be 'balanced_selectivity' or 'raw_mean_difference'.")
+
         rows = self._contrast_prompt_rows(n)
         available = sorted({str(row['concept']) for row in rows})
         if concept not in available:
             raise ValueError(f'Concept must be one of {available}.')
 
+        current_text = str(current_text or '').strip()
+        current_context_available = bool(current_text)
         texts = [str(row['text']) for row in rows]
+        if current_context_available:
+            texts.append(current_text)
+
         batch = self.tokenizer(
             texts,
             return_tensors='pt',
@@ -1826,14 +1861,17 @@ class FeatureLensRuntime:
             reduce='amax',
             include_self=True,
         )
+
+        controlled_count = len(rows)
+        controlled_dense = dense[:controlled_count]
         target_mask = torch.tensor(
             [str(row['concept']) == concept for row in rows],
             device=dense.device,
             dtype=torch.bool,
         )
         other_mask = ~target_mask
-        target = dense[target_mask]
-        other = dense[other_mask]
+        target = controlled_dense[target_mask]
+        other = controlled_dense[other_mask]
         target_mean = target.mean(dim=0)
         other_mean = other.mean(dim=0)
         target_rate = (target > 0).float().mean(dim=0)
@@ -1841,6 +1879,31 @@ class FeatureLensRuntime:
         mean_diff = target_mean - other_mean
         selectivity = mean_diff / (target_mean + other_mean + 1e-8)
 
+        # A scale-aware but selectivity-first exploratory score. log1p prevents very large SAE coefficients from
+        # overwhelming features that are much more exclusive to the target concept.
+        balanced_score = selectivity.clamp_min(0) * target_rate * torch.log1p(target_mean.clamp_min(0))
+
+        current_prompt_max = torch.zeros(self.settings.sae_width, device=dense.device, dtype=torch.float32)
+        current_token_dense = torch.zeros_like(current_prompt_max)
+        resolved_current_idx: int | None = None
+        if current_context_available:
+            current_row = controlled_count
+            current_prompt_max = dense[current_row]
+            valid_positions = torch.nonzero(attention[current_row], as_tuple=False).flatten()
+            prompt_len = int(valid_positions.numel())
+            if prompt_len:
+                resolved_current_idx = self._resolve_index(int(current_token_index), prompt_len)
+                padded_position = int(valid_positions[resolved_current_idx].item())
+                token_ids = encoding.indices[current_row, padded_position]
+                token_values = encoding.values[current_row, padded_position].float().clamp_min(0)
+                current_token_dense.scatter_reduce_(
+                    0,
+                    token_ids,
+                    token_values,
+                    reduce='amax',
+                    include_self=True,
+                )
+
         eligible = (target_mean > 0) & (mean_diff > 0)
         candidate_idx = torch.nonzero(eligible, as_tuple=False).flatten()
         if candidate_idx.numel() == 0:
@@ -1849,38 +1912,61 @@ class FeatureLensRuntime:
                 layer=int(layer),
                 prompts_per_concept=n,
                 top_n=top_n,
+                ranking_mode=ranking_mode,
                 rows=[],
                 chart_rows=[],
                 candidate_ids=[],
+                default_candidate_id=None,
+                current_context_available=current_context_available,
+                current_token_index=resolved_current_idx,
             )
-        order = torch.argsort(mean_diff[candidate_idx], descending=True)
+
+        ranking_values = balanced_score if ranking_mode == 'balanced_selectivity' else mean_diff
+        order = torch.argsort(ranking_values[candidate_idx], descending=True)
         candidate_idx = candidate_idx[order[:top_n]]
 
         table_rows: list[list[object]] = []
         chart_rows: list[list[object]] = []
+        default_candidate_id: int | None = None
         for rank, feature_tensor in enumerate(candidate_idx, start=1):
             fid = int(feature_tensor.item())
+            current_max = float(current_prompt_max[fid].item()) if current_context_available else 0.0
+            current_token = float(current_token_dense[fid].item()) if current_context_available else 0.0
+            if default_candidate_id is None and current_token > 0:
+                default_candidate_id = fid
+            score = float(ranking_values[fid].item())
             row = [
                 rank,
                 fid,
+                score,
                 float(target_mean[fid].item()),
                 float(other_mean[fid].item()),
                 float(mean_diff[fid].item()),
                 float(selectivity[fid].item()),
                 float(target_rate[fid].item()),
                 float(other_rate[fid].item()),
-                float(target[:, fid].max().item()),
+                current_max,
+                current_token,
+                bool(current_token > 0),
             ]
             table_rows.append(row)
-            chart_rows.append([str(fid), float(mean_diff[fid].item())])
+            chart_rows.append([str(fid), score])
+
+        if default_candidate_id is None and candidate_idx.numel():
+            default_candidate_id = int(candidate_idx[0].item())
+
         return ConceptFeatureDiscoveryResult(
             concept=concept,
             layer=int(layer),
             prompts_per_concept=n,
             top_n=top_n,
+            ranking_mode=ranking_mode,
             rows=table_rows,
             chart_rows=chart_rows,
             candidate_ids=[int(x.item()) for x in candidate_idx],
+            default_candidate_id=default_candidate_id,
+            current_context_available=current_context_available,
+            current_token_index=resolved_current_idx,
         )
 
     @staticmethod
@@ -1964,6 +2050,102 @@ class FeatureLensRuntime:
             cue_count=len(cue_list),
         )
 
+    @torch.inference_mode()
+    def feature_cue_context_scan(
+        self,
+        feature_id: int,
+        layer: int,
+        stems: Sequence[str],
+        cues: Sequence[str],
+    ) -> FeatureCueContextResult:
+        """Cross completion cues with multiple prompt stems in one batch.
+
+        This distinguishes a cue-specific response (for example a feature that responds to ``is`` everywhere)
+        from a context-sensitive completion-boundary response.
+        """
+        stem_list: list[str] = []
+        seen_stems: set[str] = set()
+        for raw in stems:
+            stem = str(raw).strip()
+            if not stem or stem in seen_stems:
+                continue
+            seen_stems.add(stem)
+            stem_list.append(stem)
+        cue_list: list[str] = []
+        seen_cues: set[str] = set()
+        for raw in cues:
+            cue = str(raw).strip()
+            if not cue or cue in seen_cues:
+                continue
+            seen_cues.add(cue)
+            cue_list.append(cue)
+        if not stem_list:
+            raise ValueError('Enter at least one prompt stem.')
+        if not cue_list:
+            raise ValueError('Enter at least one completion cue.')
+        if len(stem_list) > 8:
+            raise ValueError('Cue-context scan supports at most 8 prompt stems.')
+        if len(cue_list) > 8:
+            raise ValueError('Cue-context scan supports at most 8 completion cues.')
+        if len(stem_list) * len(cue_list) > 40:
+            raise ValueError('Cue-context scan supports at most 40 stem × cue conditions per run.')
+        if int(feature_id) < 0 or int(feature_id) >= self.settings.sae_width:
+            raise ValueError(f'Feature id must be in [0, {self.settings.sae_width - 1}].')
+
+        self.ensure_ready(preload_saes=False)
+        assert self.model is not None and self.tokenizer is not None and self.sae_store is not None
+        conditions: list[tuple[str, str, str]] = []
+        for stem in stem_list:
+            for cue in cue_list:
+                conditions.append((stem, cue, self._cue_prompt(stem, cue)))
+        prompts = [full for _, _, full in conditions]
+        batch = self.tokenizer(
+            prompts,
+            return_tensors='pt',
+            padding=True,
+            truncation=True,
+            max_length=self.settings.max_prompt_tokens,
+        )
+        batch = {key: value.to(self.device) for key, value in batch.items()}
+        capture: dict = {}
+        with self._capture_hook(int(layer), capture):
+            self.model(**batch, use_cache=False)
+        sae = self.sae_store.get(int(layer))
+        encoding = sae.encode(capture['hidden'])
+        attention = batch.get('attention_mask', torch.ones_like(batch['input_ids'])).bool()
+
+        rows: list[list[object]] = []
+        chart_rows: list[list[object]] = []
+        active_count = 0
+        cue_active_context_counts = {cue: 0 for cue in cue_list}
+        for row_idx, (stem, cue, full_prompt) in enumerate(conditions):
+            valid_positions = torch.nonzero(attention[row_idx], as_tuple=False).flatten()
+            final_pos = int(valid_positions[-1].item())
+            indices = encoding.indices[row_idx, final_pos]
+            values = encoding.values[row_idx, final_pos]
+            mask = indices == int(feature_id)
+            activation = float(values[mask][0].item()) if bool(mask.any()) else 0.0
+            active = activation > 0
+            active_count += int(active)
+            cue_active_context_counts[cue] += int(active)
+            token_id = int(batch['input_ids'][row_idx, final_pos].item())
+            final_token = self.tokenizer.decode([token_id])
+            short_stem = stem if len(stem) <= 42 else stem[:39] + '…'
+            rows.append([stem, cue, full_prompt, repr(final_token), activation, active])
+            chart_rows.append([short_stem, cue, activation])
+
+        return FeatureCueContextResult(
+            feature_id=int(feature_id),
+            layer=int(layer),
+            stems=stem_list,
+            cues=cue_list,
+            rows=rows,
+            chart_rows=chart_rows,
+            active_condition_count=active_count,
+            condition_count=len(conditions),
+            cue_active_context_counts=cue_active_context_counts,
+        )
+
 
 RUNTIME = FeatureLensRuntime()
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "featurelens"
-version = "0.6.0"
+version = "0.7.0"
 description = "Causal sparse-feature interpretability workbench for Qwen3 and Qwen-Scope SAEs"
 requires-python = ">=3.10"
 

research_config.json

@@ -14,7 +14,7 @@
     "mathematics",
     "positive_sentiment",
     "negative_sentiment",
-    "french_language",
+    "german_language",
     "factual_entities",
     "uncertainty"
   ],
@@ -90,7 +90,7 @@
     "contrastive_continuation_preference_test",
     "feature_decoder_geometry"
   ],
-  "concept_candidate_discovery_metric": "prompt-wide target-minus-other mean max activation; exploratory only",
+  "concept_candidate_discovery_metric": "balanced exploratory score = selectivity × target activation rate × log1p(target mean), with raw-contrast alternative and current-Workbench compatibility",
   "completion_cue_scan": "final-token feature activation after controlled suffix/cue substitution",
   "live_features_v0_6": [
     "start_here_plain_language_onboarding",
@@ -100,5 +100,14 @@
     "consistent_heading_and_table_typography",
     "concept_guided_candidate_feature_discovery",
     "completion_cue_sensitivity_scan"
+  ],
+  "live_features_v0_7": [
+    "cleaned_nonaccordion_experiment_layout",
+    "focused_fullscreen_modal_for_tables_and_plots",
+    "descriptive_plot_export_filenames",
+    "german_language_control_concept",
+    "balanced_candidate_ranking_and_current_prompt_compatibility",
+    "click_to_select_candidate_rows",
+    "completion_cue_context_matrix"
   ]
 }

scripts/release_check.py

@@ -5,7 +5,6 @@ import subprocess
 from collections import Counter
 from pathlib import Path
 
-
 ROOT = Path(__file__).resolve().parents[1]
 MAX_FILE_SIZE_BYTES = 5_000_000  # 5 MB
 
@@ -136,6 +135,24 @@ def check_config(config: dict) -> None:
             f'{sorted(actual_live_v06)}'
         )
 
+    required_live_v07 = {
+        'cleaned_nonaccordion_experiment_layout',
+        'focused_fullscreen_modal_for_tables_and_plots',
+        'descriptive_plot_export_filenames',
+        'german_language_control_concept',
+        'balanced_candidate_ranking_and_current_prompt_compatibility',
+        'click_to_select_candidate_rows',
+        'completion_cue_context_matrix',
+    }
+    actual_live_v07 = set(config.get('live_features_v0_7', []))
+    if actual_live_v07 != required_live_v07:
+        raise SystemExit(
+            'research_config.json live_features_v0_7 mismatch: '
+            f'{sorted(actual_live_v07)}'
+        )
+    if 'german_language' not in config.get('concepts', []) or 'french_language' in config.get('concepts', []):
+        raise SystemExit('research_config.json must use german_language and must not contain french_language.')
+
 
 def check_datasets(config: dict) -> tuple[list[dict], list[dict]]:
     prompts = load_jsonl(ROOT / 'data' / 'prompts.jsonl')
@@ -201,17 +218,20 @@ def check_readme() -> None:
         'token activation',
         'concept-guided candidate',
         'completion-cue',
+        'cue × context',
+        'balanced selectivity',
+        'german',
         'start here',
     ]
     missing = [value for value in required_strings if value.lower() not in readme.lower()]
     if missing:
-        raise SystemExit(f'README.md is missing required v0.6 content: {missing}')
+        raise SystemExit(f'README.md is missing required v0.7 content: {missing}')
 
 
 def check_pyproject() -> None:
     text = (ROOT / 'pyproject.toml').read_text(encoding='utf-8')
-    if 'version = "0.6.0"' not in text:
-        raise SystemExit('pyproject.toml must declare version 0.6.0.')
+    if 'version = "0.7.0"' not in text:
+        raise SystemExit('pyproject.toml must declare version 0.7.0.')
 
 
 def main() -> None:
@@ -229,7 +249,7 @@ def main() -> None:
     print(f'  layers: {config["layers"]}')
     print(f'  feature-set sizes: {config["feature_set_sizes"]}')
     print(f'  random controls: {config["live_random_controls"]}')
-    print('  release: v0.6.0')
+    print('  release: v0.7.0')
 
 
 if __name__ == '__main__':

tests/test_data.py

@@ -16,6 +16,11 @@ def test_discovery_dataset_is_balanced_and_paired() -> None:
     assert len(rows) == 224
     concept_counts = Counter(row['concept'] for row in rows)
     assert set(concept_counts.values()) == {32}
+    assert 'german_language' in concept_counts
+    assert 'french_language' not in concept_counts
+    german = [row for row in rows if row['concept'] == 'german_language']
+    assert len(german) == 32
+    assert any('Tisch' in row['text'] or 'Deutsch' in row['text'] or 'Hamburg' in row['text'] for row in german)
     pair_counts = Counter(row['pair_id'] for row in rows)
     assert set(pair_counts.values()) == {2}
 
@@ -25,3 +30,6 @@ def test_causal_dataset_covers_every_discovery_concept() -> None:
     causal = read_jsonl(ROOT / 'data' / 'causal_tasks.jsonl')
     assert len(causal) == 28
     assert {row['concept'] for row in causal} == {row['concept'] for row in discovery}
+    german = [row for row in causal if row['concept'] == 'german_language']
+    assert len(german) == 4
+    assert all(row['target'].strip() in {'hallo', 'danke', 'ja', 'guten Abend'} for row in german)

tests/test_live_runtime_helpers.py

@@ -219,12 +219,41 @@ def test_concept_contrast_promptwide_scan_runs_on_toy_runtime() -> None:
 def test_concept_feature_discovery_runs_on_toy_runtime() -> None:
     runtime = make_runtime()
     result = runtime.concept_feature_discovery(
-        concept='mathematics', layer=0, prompts_per_concept=1, top_n=3
+        concept='mathematics',
+        layer=0,
+        prompts_per_concept=1,
+        top_n=3,
+        ranking_mode='balanced_selectivity',
+        current_text='abc',
+        current_token_index=-1,
     )
     assert result.concept == 'mathematics'
+    assert result.ranking_mode == 'balanced_selectivity'
+    assert result.current_context_available is True
+    assert result.current_token_index == 2
     assert len(result.rows) <= 3
     assert result.candidate_ids == [int(row[1]) for row in result.rows]
-    assert all(len(row) == 9 for row in result.rows)
+    assert all(len(row) == 12 for row in result.rows)
+    if result.rows:
+        assert result.default_candidate_id in result.candidate_ids
+        assert all(math.isfinite(float(row[2])) for row in result.rows)
+        assert all(float(row[9]) >= 0 for row in result.rows)  # current prompt max
+        assert all(float(row[10]) >= 0 for row in result.rows)  # current token activation
+
+
+def test_concept_feature_discovery_supports_raw_mean_difference() -> None:
+    runtime = make_runtime()
+    result = runtime.concept_feature_discovery(
+        concept='mathematics',
+        layer=0,
+        prompts_per_concept=1,
+        top_n=3,
+        ranking_mode='raw_mean_difference',
+    )
+    assert result.ranking_mode == 'raw_mean_difference'
+    assert result.current_context_available is False
+    assert result.current_token_index is None
+    assert all(len(row) == 12 for row in result.rows)
 
 
 def test_feature_cue_scan_runs_on_toy_runtime() -> None:
@@ -239,3 +268,19 @@ def test_feature_cue_scan_runs_on_toy_runtime() -> None:
     assert len(result.rows) == 3
     assert all(len(row) == 5 for row in result.rows)
     assert 0 <= result.active_cue_count <= result.cue_count
+
+
+def test_feature_cue_context_scan_runs_on_toy_runtime() -> None:
+    runtime = make_runtime()
+    result = runtime.feature_cue_context_scan(
+        feature_id=0,
+        layer=0,
+        stems=['abc', 'xyz'],
+        cues=['is', '=', ':'],
+    )
+    assert result.condition_count == 6
+    assert len(result.rows) == 6
+    assert all(len(row) == 6 for row in result.rows)
+    assert 0 <= result.active_condition_count <= result.condition_count
+    assert set(result.cue_active_context_counts) == {'is', '=', ':'}
+    assert len(result.chart_rows) == 6

tests/test_ui_helpers.py

@@ -42,3 +42,31 @@ def test_tsv_copy_payload_keeps_headers() -> None:
     payload = app._tsv(frame)
     assert payload.startswith('Feature id\tActivation\n')
     assert payload.endswith('1\t2.0\n')
+
+
+def test_use_candidate_feature_returns_explicit_handoff_status() -> None:
+    app = _import_app()
+    outputs = app.use_candidate_feature('21885')
+    assert outputs[:4] == ('21885', '21885', '21885', '21885')
+    assert 'Feature 21885 loaded' in outputs[4]
+    assert 'Single-feature causal test' in outputs[4]
+
+
+def test_select_candidate_row_uses_feature_id_column() -> None:
+    app = _import_app()
+    table = pd.DataFrame(
+        [[1, 21885, 3.4], [2, 445, 3.3]],
+        columns=['Rank', 'Feature id', 'Candidate score'],
+    )
+    event = SimpleNamespace(index=(1, 0))
+    update = app.select_candidate_row(table, event)
+    # Gradio returns an update dictionary-like object in current releases.
+    assert update['value'] == '445'
+
+
+def test_frontend_helpers_name_exports_and_preserve_focus_position() -> None:
+    app = _import_app()
+    assert 'featurelens_${stem' in app.INSTALL_REFLOW_JS
+    assert 'chart.png' in app.INSTALL_REFLOW_JS
+    assert '__featurelens_focus_scroll_y' in app.INSTALL_REFLOW_JS
+    assert 'window.scrollTo' in app.INSTALL_REFLOW_JS