Release FeatureLens v0.6.0

CHANGELOG.md

@@ -1,5 +1,25 @@
 # Changelog
 
+## v0.6.0
+
+### UX / navigation
+- Added a plain-language **Start here** tab with a three-step workflow and glossary for non-specialist reviewers.
+- Added a persistent **Current Workbench context** banner so inherited prompt/layer/token state is visible from every tab.
+- Added explicit editable feature selectors for **scale dose-response** and **contrastive continuation preference** instead of silently reusing the single-feature selector.
+- Clarified state provenance in Feature Sets and Feature Evidence; experiment text now states whether it inherits Workbench context or uses an independent prompt set.
+- Normalized heading hierarchy and increased table/header typography for readability.
+- Added native plot **fullscreen** and **export PNG** controls to every BarPlot/LinePlot.
+
+### Live research tools
+- Added **concept-guided candidate feature discovery**: rank features for a selected controlled concept using prompt-wide target-minus-other mean maximum activation, with selectivity and activation-rate diagnostics.
+- Added a **completion-cue sensitivity scan** that appends controlled suffixes to a prompt stem and measures the selected feature at the resulting final token.
+- Added a one-click action to reuse a discovered candidate across single-feature, dose-response, contrastive, and evidence feature selectors.
+- Candidate discovery and cue scans are explicitly exploratory; neither creates semantic labels or overwrites `Offline concept hint`.
+
+### Validation
+- Expanded toy-runtime coverage to **40 tests**, including candidate-feature discovery and completion-cue scans.
+- Retained the actual Gradio `launch()` smoke gate, compile gate, release checker, batched-null regression tests, and final-release adversarial-test deferral.
+
 ## v0.5.0
 
 ### Causal specificity

README.md

@@ -13,7 +13,7 @@ license: mit
 
 # FeatureLens — Causal Interpretability Workbench
 
-> **v0.5:** contrastive continuation preference tests, feature-token activation traces, SAE decoder-geometry diagnostics, prompt-wide concept contrast, a wider centered research UI, copy confirmation, and dynamic-height reflow safeguards—while retaining v0.4's batched null references and 8-direction random ensembles.
+> **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.
 
 **Research question:**
 
@@ -28,14 +28,16 @@ FeatureLens is independent of thesis code and thesis datasets.
 1. **Reconstruction** — does the SAE represent the residual stream reasonably well?
 2. **Prediction** — does a feature distinguish a controlled concept on held-out paraphrase groups?
 3. **Robustness** — does the representation survive rewording locally and prompt-wide?
-4. **Controlled contrast** — does a selected feature prefer one concept group in a balanced live mini-batch?
-5. **Single-feature causality** — does ablation/scaling change downstream probability?
-6. **Dose-response** — does effect size vary coherently with feature coefficient?
-7. **Feature-set causality** — do small sparse subspaces matter more than one feature?
-8. **Non-additivity** — does a joint edit differ from the sum of individual effects?
-9. **Specificity** — are targeted effects larger than norm-matched random residual perturbations?
-10. **Contrastive specificity** — does an edit change preference between a desired and comparison continuation?
-11. **Geometry** — are selected SAE decoder directions aligned, orthogonal, or cancelling before downstream model non-linearity?
+4. **Candidate discovery** — which features are more active for a chosen controlled concept than for the other groups?
+5. **Controlled contrast** — does a selected feature prefer one concept group in a balanced live mini-batch?
+6. **Cue sensitivity** — is a feature responding to a lexical/completion cue rather than the underlying concept?
+7. **Single-feature causality** — does ablation/scaling change downstream probability?
+8. **Dose-response** — does effect size vary coherently with feature coefficient?
+9. **Feature-set causality** — do small sparse subspaces matter more than one feature?
+10. **Non-additivity** — does a joint edit differ from the sum of individual effects?
+11. **Specificity** — are targeted effects larger than norm-matched random residual perturbations?
+12. **Contrastive specificity** — does an edit change preference between a desired and comparison continuation?
+13. **Geometry** — are selected SAE decoder directions aligned, orthogonal, or cancelling before downstream model non-linearity?
 
 A high AUROC, high paraphrase overlap, or a large activation is still correlational evidence. Causal evidence comes from downstream change under controlled intervention.
 
@@ -60,7 +62,7 @@ The UI also reports single-forward → batched-null execution drift as a diagnos
 
 ## Random-control ensemble
 
-A single random residual direction can be unusually weak or unusually disruptive by chance. FeatureLens uses with **8 deterministic norm-matched random directions**.
+A single random residual direction can be unusually weak or unusually disruptive by chance. FeatureLens uses **8 deterministic norm-matched random directions**.
 
 For a targeted residual delta `Δh`, each control satisfies:
 
@@ -125,10 +127,12 @@ The difference is reported as an **interaction excess**. A non-zero excess means
 
 ### Feature evidence
 
-The **Feature evidence** tab now provides two complementary live views:
+The **Feature evidence** tab now supports four complementary live questions:
 
-1. **Feature-token activation trace** — shows where the selected feature enters the SAE TopK support across every token of the current Workbench prompt.
-2. **Prompt-wide controlled concept contrast** — evaluates the selected feature on a small balanced batch drawn from the repository's seven controlled concept groups:
+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**.
+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:
 
 - code;
 - mathematics;
@@ -138,9 +142,9 @@ The **Feature evidence** tab now provides two complementary live views:
 - factual entities;
 - uncertainty.
 
-One wording per paraphrase pair is used to reduce near-duplicate inflation. Each prompt is summarized by the feature's **maximum activation across non-padding prompt tokens**. The live scan reports mean/median prompt-wide maxima, prompt activation rate, mean activation when active, and maximum activation by concept. If the feature is inactive in every sampled prompt, the UI reports that explicitly and does not fabricate a leading concept.
+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.
 
-This is an exploratory live feature-evidence tool, **not an automatic semantic label**. The held-out offline AUROC/F1 benchmark remains the stronger claim.
+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
 
@@ -169,11 +173,20 @@ 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.
 
-## v0.5 interface and embedded-Space behavior
+## v0.6 interface and navigation
 
-The app uses `gr.Blocks(fill_width=True)` together with a centered 1600 px research canvas so wide result tables make better use of desktop screens. Result tables use bounded heights to reduce runaway page growth in the Hugging Face iframe. A lightweight browser-side resize/mutation observer requests a layout reflow when dynamic outputs change height, addressing the intermittent case where the embedded Space stopped scrolling until the user switched tabs.
+v0.6 is designed so a reviewer does not need prior mechanistic-interpretability vocabulary to find the controls:
 
-All explicit action/copy buttons share the same muted teal treatment. Copy actions now briefly change their own label to **✓ Copied with headers** before returning to the normal label. Typography and control sizes were normalized around a serif research-document style.
+- 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;
+- experiments that reuse that state explicitly say **Uses current Workbench context**;
+- 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;
+- 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.
 
 ## Copy tables with headers
 
@@ -291,7 +304,7 @@ FEATURELENS_CONTRAST_PROMPTS_PER_CONCEPT=4
 
 ## UI design
 
-v0.5 keeps the interface deliberately closer to a conventional research instrument than a rounded dashboard:
+v0.6 keeps the interface deliberately 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;
@@ -343,12 +356,13 @@ python scripts/ui_smoke.py
 python scripts/release_check.py
 ```
 
-See [`docs/VALIDATION.md`](docs/VALIDATION.md). The v0.5 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 v0.6 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.
 
 ## Limitations
 
 - SAE features are sparse directions, not guaranteed monosemantic concepts.
-- The live concept contrast is a small controlled diagnostic, not a universal ontology.
+- Live candidate discovery and concept contrast are small controlled diagnostics, not a universal ontology or semantic labeling system.
+- Completion-cue sensitivity diagnoses lexical/structural dependence but does not establish a feature's full semantics.
 - Prompt-wide max pooling measures feature presence anywhere in the prompt and discards token order.
 - Joint residual edits can be non-additive downstream without implying direct feature-feature interaction.
 - Eight live random controls give only a coarse empirical tail probability.
@@ -358,7 +372,7 @@ See [`docs/VALIDATION.md`](docs/VALIDATION.md). The v0.5 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, token/prompt-wide 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 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.
 
 ## Acknowledgements
 

app.py

@@ -38,12 +38,14 @@ CSS = r"""
 .gradio-container input,
 .gradio-container textarea,
 .gradio-container select { font-size: 15.5px !important; }
-.gradio-container table { font-size: 14.5px !important; }
+.gradio-container table { font-size: 15.5px !important; }
+.gradio-container table thead th, .gradio-container [role="columnheader"] { font-size: 16px !important; font-weight: 700 !important; }
+.gradio-container table tbody td, .gradio-container [role="gridcell"] { font-size: 15.5px !important; }
 .gradio-container label, .gradio-container .label-wrap { font-size: 15px !important; }
 .gradio-container p, .gradio-container li { font-size: 16px; }
-.gradio-container h2 { font-size: 1.55rem; }
-.gradio-container h3 { font-size: 1.30rem; }
-.gradio-container h4 { font-size: 1.08rem; }
+.gradio-container h2 { font-size: 1.72rem; line-height: 1.25; }
+.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;
   border-bottom: 1px solid var(--border-color-primary);
@@ -71,8 +73,9 @@ CSS = r"""
   border-top: 1px solid var(--border-color-primary);
   font-variant: small-caps;
   letter-spacing: .045em;
-  font-size: .98rem;
-  opacity: .78;
+  font-size: 1.12rem;
+  font-weight: 700;
+  opacity: .88;
 }
 .token-wrap { display: flex; flex-wrap: wrap; gap: 6px; padding: 8px 2px 14px; line-height: 1.9; }
 .token {
@@ -115,6 +118,25 @@ CSS = r"""
   min-height: 110px;
 }
 .tabs, .tabitem { padding-bottom: 26px !important; }
+.context-card {
+  border: 1px solid var(--border-color-primary);
+  border-left: 4px solid #708B86;
+  background: var(--background-fill-secondary);
+  padding: 10px 14px;
+  margin: 8px 0 16px;
+  border-radius: 2px;
+}
+.context-card p { margin: 0 !important; }
+.start-card {
+  border: 1px solid var(--border-color-primary);
+  background: var(--background-fill-secondary);
+  padding: 14px 16px;
+  min-height: 150px;
+  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; }
 @media (max-width: 900px) {
   .gradio-container { width: 100% !important; padding-left: 12px !important; padding-right: 12px !important; }
 }
@@ -409,6 +431,42 @@ def _contrastive_metrics_markdown(result) -> str:
     )
 
 
+
+
+def _discovery_metrics_markdown(result) -> str:
+    if not result.candidate_ids:
+        return (
+            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."
+        )
+    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."
+    )
+
+
+def _cue_metrics_markdown(result) -> str:
+    return (
+        f"Feature **{result.feature_id}**, layer **{result.layer}** · active for "
+        f"**{result.active_cue_count}/{result.cue_count}** tested completion cues at the final token.  \n"
+        "This helps distinguish a concept-associated feature from a lexical/structural completion-cue feature. "
+        "It is a controlled diagnostic, not a feature label."
+    )
+
+
+def _global_context_markdown(prompt: str, layer: int, result) -> str:
+    token = result.tokens[result.token_index] if result.tokens else ""
+    short = prompt[:120] + ("…" if len(prompt) > 120 else "")
+    return (
+        "**Current Workbench context** — "
+        f"Prompt: `{short}` · layer **{int(layer)}** · token **{result.token_index}** ({token!r}).  \n"
+        "Panels marked **Uses current Workbench context** inherit this prompt/layer/token. "
+        "Feature selectors remain editable inside each experiment."
+    )
+
 @gpu(duration=30)
 def analyze_prompt(prompt: str, layer: int, token_index: int, top_n: int):
     try:
@@ -437,12 +495,15 @@ def analyze_prompt(prompt: str, layer: int, token_index: int, top_n: int):
             table,
             chart_df,
             feature_update,
+            gr.update(choices=choices, value=choices[0] if choices else None),
+            gr.update(choices=choices, value=choices[0] if choices else None),
             feature_set_update,
             contrast_update,
             gr.update(value=int(layer)),
             _analysis_metrics_markdown(result),
             location,
             location,
+            _global_context_markdown(prompt, int(layer), result),
             _tsv(table),
         )
     except Exception as exc:
@@ -853,6 +914,65 @@ def run_contrastive_causal(
         _raise_ui_error(exc)
 
 
+
+@gpu(duration=35)
+def run_concept_feature_discovery(concept: str, layer: int, prompts_per_concept: int, top_n: int):
+    try:
+        result = RUNTIME.concept_feature_discovery(
+            concept=concept,
+            layer=int(layer),
+            prompts_per_concept=int(prompts_per_concept),
+            top_n=int(top_n),
+        )
+        columns = [
+            "Rank",
+            "Feature id",
+            "Target mean max",
+            "Other mean max",
+            "Mean difference",
+            "Selectivity score",
+            "Target activation rate",
+            "Other activation rate",
+            "Target max activation",
+        ]
+        table = pd.DataFrame(result.rows, columns=columns)
+        chart = pd.DataFrame(result.chart_rows, columns=["Feature", "Target − other mean"])
+        chart["Series"] = "Concept contrast"
+        choices = [str(fid) for fid in result.candidate_ids]
+        candidate_update = gr.update(choices=choices, value=choices[0] if choices else None)
+        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:
+        if feature_id is None or str(feature_id).strip() == "":
+            raise ValueError("Choose a feature id first.")
+        cues = [line for line in str(cue_text).splitlines() if line.strip()]
+        result = RUNTIME.feature_cue_scan(
+            feature_id=int(float(feature_id)),
+            layer=int(layer),
+            prompt_stem=prompt_stem,
+            cues=cues,
+        )
+        columns = ["Cue", "Full prompt", "Final token", "Activation", "Active in TopK"]
+        table = pd.DataFrame(result.rows, columns=columns)
+        chart = pd.DataFrame(result.chart_rows, columns=["Cue", "Activation"])
+        chart["Series"] = "Cue response"
+        return _cue_metrics_markdown(result), table, chart, _tsv(table)
+    except Exception as exc:
+        _raise_ui_error(exc)
+
+
+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
+
+
 def mode_help(mode: str):
     if mode == "ablate":
         return gr.update(value=0.0, interactive=False, label="Coefficient (unused for ablation)")
@@ -870,7 +990,7 @@ 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">v0.5</span></h1>'
+        '<h1>FeatureLens <span style="font-size:.48em;opacity:.58;font-weight:400">v0.6</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>'
@@ -879,6 +999,45 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
         'also causally influence the model’s behaviour?</div>'
     )
 
+    global_context = gr.Markdown(
+        "**Current Workbench context:** none yet. Go to **Workbench → I. Inspect a prompt location**, "
+        "set the prompt/layer/token, then click **Inspect sparse features**.",
+        elem_classes=["context-card"],
+    )
+
+    with gr.Tab("Start here"):
+        gr.Markdown(
+            "## Start here\n"
+            "FeatureLens is organized as a short evidence workflow. You do **not** need to know a feature id before "
+            "starting: inspect a prompt first, or use **Feature evidence → Concept-guided candidate discovery**."
+        )
+        with gr.Row(equal_height=True):
+            gr.HTML(
+                '<div class="start-card"><h3>1. Choose a location</h3>'
+                '<p>In <b>Workbench</b>, enter a prompt, choose a residual layer and token, then click '
+                '<b>Inspect sparse features</b>. This establishes the shared Workbench context.</p></div>'
+            )
+            gr.HTML(
+                '<div class="start-card"><h3>2. Choose evidence</h3>'
+                '<p>Pick a feature in the experiment you want to run. Feature selectors are populated from the latest '
+                'inspection but remain editable. Feature sets use 2–5 features.</p></div>'
+            )
+            gr.HTML(
+                '<div class="start-card"><h3>3. Ask a causal question</h3>'
+                '<p>Ablate/scale features, compare against norm-matched random controls, test dose response, '
+                'preference shifts, robustness, geometry, or concept association.</p></div>'
+            )
+        gr.Markdown(
+            "### Plain-language glossary\n"
+            "- **Token:** a chunk of text the model processes.\n"
+            "- **Residual layer:** a checkpoint in the model's internal computation.\n"
+            "- **SAE feature:** one sparse direction used by the autoencoder to represent a residual activation.\n"
+            "- **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."
+        )
+
     with gr.Tab("Workbench"):
         gr.HTML('<div class="section-rule">I. Inspect a prompt location</div>')
         with gr.Row(equal_height=False):
@@ -942,6 +1101,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     x_title="Feature id",
                     y_title="Activation",
                     x_label_angle=-35,
+                    buttons=["fullscreen", "export"],
                     height=330,
                 )
 
@@ -956,6 +1116,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     choices=[],
                     allow_custom_value=True,
                     label="Single feature id",
+                    info="Populated from the latest inspection. You can choose another active feature or enter any valid id.",
                 )
                 mode = gr.Dropdown(
                     choices=["ablate", "scale", "inject"],
@@ -1005,7 +1166,15 @@ 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):
+        with gr.Accordion("Scale dose-response", open=False, elem_classes=["experiment-accordion"]):
+            gr.Markdown("### Scale dose-response")
+            gr.Markdown("**Uses current Workbench prompt / layer / token.** Choose the feature for this experiment below.")
+            dose_feature_id = gr.Dropdown(
+                choices=[],
+                allow_custom_value=True,
+                label="Dose-response feature id",
+                info="Populated from the latest Workbench inspection; editable here.",
+            )
             gr.Markdown(
                 "Always a **scale** experiment: 0× = ablation, 1× = numerical no-edit reference, 2× = double "
                 "the native coefficient. The six conditions are evaluated together, and all deltas are measured "
@@ -1033,11 +1202,20 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         title="Scale dose-response",
                         x_title="Feature multiplier",
                         y_title="Δ mean log p/token",
+                        buttons=["fullscreen", "export"],
                         height=330,
                     )
 
         gr.HTML('<div class="section-rule">IV. Contrastive continuation preference</div>')
-        with gr.Accordion("Contrastive causal preference test", open=False):
+        with gr.Accordion("Contrastive causal preference test", open=False, elem_classes=["experiment-accordion"]):
+            gr.Markdown("### Contrastive causal preference test")
+            gr.Markdown("**Uses current Workbench prompt / layer / token.** Select the feature explicitly below.")
+            contrastive_feature_id = gr.Dropdown(
+                choices=[],
+                allow_custom_value=True,
+                label="Contrastive feature id",
+                info="Populated from the latest Workbench inspection; editable here.",
+            )
             gr.Markdown(
                 "Absolute target probability can move because an intervention broadly perturbs the distribution. "
                 "This test asks the stricter question: **does the edit shift preference between two exact continuations?** "
@@ -1083,14 +1261,15 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         title="Preference between exact continuations",
                         x_title="Execution condition",
                         y_title="Sequence log-odds A−B",
+                        buttons=["fullscreen", "export"],
                         height=320,
                     )
 
     with gr.Tab("Feature sets"):
         gr.Markdown(
             "### Distributed sparse representations\n"
-            "Run **Inspect sparse features** in the Workbench first. The prompt, layer, token index, and active "
-            "feature selector below reuse that Workbench state."
+            "Run **Inspect sparse features** in the Workbench first. **Uses current Workbench context:** prompt, layer, "
+            "and token are inherited from that inspection. The feature set below is always editable."
         )
         feature_set_location = gr.Markdown("No Workbench inspection has been run yet.")
         feature_set_ids = gr.Dropdown(
@@ -1182,6 +1361,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     title="Effect vs feature-set size",
                     x_title="Number of jointly ablated features",
                     y_title="Δ mean log p/token",
+                    buttons=["fullscreen", "export"],
                     height=330,
                 )
 
@@ -1215,12 +1395,14 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     x_title="Intervention condition",
                     y_title="Δ mean log p/token",
                     x_label_angle=-25,
+                    buttons=["fullscreen", "export"],
                     height=330,
                 )
 
 
         gr.HTML('<div class="section-rule">Decoder geometry</div>')
-        with gr.Accordion("Feature-set decoder geometry", open=False):
+        with gr.Accordion("Feature-set decoder geometry", open=False, elem_classes=["experiment-accordion"]):
+            gr.Markdown("### Feature-set decoder geometry")
             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 "
@@ -1251,6 +1433,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                         x_title="Feature pair",
                         y_title="Cosine similarity",
                         x_label_angle=-30,
+                        buttons=["fullscreen", "export"],
                         height=320,
                     )
 
@@ -1261,7 +1444,48 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
             "across FeatureLens's seven controlled concept groups. The concept scan uses the feature's **maximum "
             "activation across non-padding tokens in each prompt**, avoiding arbitrary final-token comparisons."
         )
-        contrast_location = gr.Markdown("No Workbench inspection has been run yet.")
+        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."
+        )
+        with gr.Row(equal_height=True):
+            discovery_concept = gr.Dropdown(
+                choices=[
+                    "code", "mathematics", "positive_sentiment", "negative_sentiment",
+                    "french_language", "factual_entities", "uncertainty"
+                ],
+                value="mathematics",
+                label="Target concept",
+            )
+            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")
+        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
+                )
+                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
+                )
+        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)
+
+        gr.HTML('<div class="section-rule">B. Inspect one feature</div>')
+        contrast_location = gr.Markdown(
+            "**Activation trace uses the current Workbench prompt.** The controlled concept scan below uses its own balanced prompt set."
+        )
         with gr.Row(equal_height=True):
             contrast_feature_id = gr.Dropdown(
                 choices=[],
@@ -1284,7 +1508,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                 scale=2,
             )
 
-        gr.Markdown("#### Activation trace across the current Workbench prompt")
+        gr.Markdown("### Activation trace across the current Workbench prompt")
         gr.Markdown(
             "Uses the prompt currently entered in **Workbench** and shows exactly which prompt tokens include the "
             "selected feature in the SAE TopK support."
@@ -1312,10 +1536,35 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     x_title="Prompt token",
                     y_title="Activation",
                     x_label_angle=-35,
+                    buttons=["fullscreen", "export"],
                     height=320,
                 )
 
-        gr.Markdown("#### Prompt-wide controlled concept contrast")
+        gr.HTML('<div class="section-rule">C. Completion-cue sensitivity</div>')
+        gr.Markdown(
+            "Feature 22632 in the current math example fires only on the final `is` token. This scan tests whether a "
+            "selected feature responds to **completion cues themselves** rather than the underlying concept. Enter a stem "
+            "and one cue per line; FeatureLens appends each cue and measures the feature at the final token."
+        )
+        with gr.Row(equal_height=True):
+            cue_stem = gr.Textbox(label="Prompt stem", value="The derivative of x squared", lines=2, scale=3)
+            cue_text = gr.Textbox(label="Completion cues (one per line)", value="is\n=\n:\nequals\ntherefore", lines=5, scale=2)
+        cue_btn = gr.Button("Run completion-cue scan", variant="primary", elem_classes=["action-btn"])
+        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_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
+                )
+
+        gr.HTML('<div class="section-rule">D. 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()
         with gr.Row(equal_height=False):
@@ -1339,6 +1588,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     x_title="Concept",
                     y_title="Mean max activation",
                     x_label_angle=-25,
+                    buttons=["fullscreen", "export"],
                     height=330,
                 )
 
@@ -1346,7 +1596,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
         gr.Markdown(
             "### Local and prompt-wide robustness\n"
             "The selected-token comparison is strict but can be misleading if the two chosen tokens play different "
-            "semantic roles. v0.5 reports a **prompt-wide max-pooled feature profile**: for every SAE "
+            "semantic roles. FeatureLens reports a **prompt-wide max-pooled feature profile**: for every SAE "
             "feature, take its maximum activation across all prompt tokens, then compare the two sparse profiles."
         )
         with gr.Row():
@@ -1399,6 +1649,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     x_title="Feature id",
                     y_title="Activation",
                     x_label_angle=-35,
+                    buttons=["fullscreen", "export"],
                     height=330,
                 )
 
@@ -1448,6 +1699,7 @@ with gr.Blocks(title="FeatureLens — Causal Interpretability Workbench", fill_w
                     title="Representation trajectory",
                     x_title="Layer",
                     y_title="Normalized value",
+                    buttons=["fullscreen", "export"],
                     height=330,
                 )
 
@@ -1456,7 +1708,7 @@ 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. "
-            "v0.5 strengthens the live workbench with batch-context null references and random-control ensembles; "
+            "v0.6 strengthens the live workbench with batch-context null references and random-control ensembles; "
             "the saved report should be generated only from real experiment artifacts."
         )
 
@@ -1477,7 +1729,7 @@ $$h' = h + \sum_{i \in S}\Delta z_i d_i.$$
 
 FeatureLens patches the delta into the **original residual**; it never replaces the residual with the complete SAE reconstruction.
 
-### v0.5 control discipline
+### v0.6 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.
 
@@ -1485,13 +1737,14 @@ Batched experiments include an explicit **zero-edit row**. Causal effects are me
 
 1. SAE reconstruction quality.
 2. Held-out feature/concept prediction.
-3. Token-local activation traces and prompt-wide feature evidence.
-4. Local and prompt-wide paraphrase robustness.
-5. Single-feature causal intervention and dose-response.
-6. Contrastive continuation preference under intervention.
-7. Joint feature-set intervention and set-size sensitivity.
-8. Decoder-direction geometry and individual-vs-joint non-additivity.
-9. Specificity relative to norm-matched random controls.
+3. Concept-guided candidate discovery and selected-feature concept contrast.
+4. Token-local activation traces and completion-cue sensitivity.
+5. Local and prompt-wide paraphrase robustness.
+6. Single-feature causal intervention and dose-response.
+7. Contrastive continuation preference under intervention.
+8. Joint feature-set intervention and set-size sensitivity.
+9. Decoder-direction geometry and individual-vs-joint non-additivity.
+10. Specificity relative to norm-matched random controls.
 
 Association, robustness, geometry, and intervention evidence remain separate claims.
 """
@@ -1514,12 +1767,15 @@ Association, robustness, geometry, and intervention evidence remain separate cla
             feature_table,
             feature_plot,
             feature_id,
+            dose_feature_id,
+            contrastive_feature_id,
             feature_set_ids,
             contrast_feature_id,
             contrast_layer,
             analysis_metrics,
             feature_set_location,
             contrast_location,
+            global_context,
             feature_tsv,
         ],
     )
@@ -1539,13 +1795,13 @@ Association, robustness, geometry, and intervention evidence remain separate cla
     )
     dose_btn.click(
         run_dose_response,
-        inputs=[prompt, layer, token_index, feature_id, target_text],
+        inputs=[prompt, layer, token_index, dose_feature_id, target_text],
         outputs=[dose_table, dose_plot, dose_metrics, dose_tsv],
     )
     contrastive_mode.change(mode_help, inputs=[contrastive_mode], outputs=[contrastive_coefficient])
     contrastive_btn.click(
         run_contrastive_causal,
-        inputs=[prompt, layer, token_index, feature_id, contrastive_mode, contrastive_coefficient, contrastive_a, contrastive_b],
+        inputs=[prompt, layer, token_index, contrastive_feature_id, contrastive_mode, contrastive_coefficient, contrastive_a, contrastive_b],
         outputs=[contrastive_metrics, contrastive_table, contrastive_plot, contrastive_tsv],
     )
     set_mode.change(set_mode_help, inputs=[set_mode], outputs=[set_coefficient])
@@ -1579,6 +1835,22 @@ Association, robustness, geometry, and intervention evidence remain separate cla
         inputs=[contrast_feature_id, contrast_layer, contrast_n],
         outputs=[contrast_metrics, contrast_table, contrast_plot, contrast_tsv],
     )
+    discovery_btn.click(
+        run_concept_feature_discovery,
+        inputs=[discovery_concept, discovery_layer, discovery_n, discovery_top_n],
+        outputs=[discovery_metrics, discovery_table, discovery_plot, discovery_candidate, discovery_tsv],
+    )
+    use_candidate_btn.click(
+        use_candidate_feature,
+        inputs=[discovery_candidate],
+        outputs=[feature_id, dose_feature_id, contrastive_feature_id, contrast_feature_id],
+        queue=False,
+    )
+    cue_btn.click(
+        run_feature_cue_scan,
+        inputs=[contrast_feature_id, contrast_layer, cue_stem, cue_text],
+        outputs=[cue_metrics, cue_table, cue_plot, cue_tsv],
+    )
     para_btn.click(
         run_paraphrase_compare,
         inputs=[para_a, para_b, para_layer, para_idx_a, para_idx_b, para_top_n],
@@ -1603,6 +1875,8 @@ Association, robustness, geometry, and intervention evidence remain separate cla
         (geometry_copy, geometry_tsv),
         (trace_copy, trace_tsv),
         (contrast_copy, contrast_tsv),
+        (discovery_copy, discovery_tsv),
+        (cue_copy, cue_tsv),
         (para_copy, para_tsv),
         (trajectory_copy, trajectory_tsv),
     ]:

docs/HF_DEPLOY.md

@@ -15,8 +15,11 @@ The live app does not need every SAE layer from the full repository.
 
 ## GPU-decorated actions
 
-Current v0.5 live actions include:
+Current v0.6 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;
 - **Inspect sparse features**;
 - **Run single-feature causal test**;
 - **Run scale dose-response**;

docs/METHODOLOGY.md

@@ -264,7 +264,7 @@ The selected-token and prompt-wide metrics answer different questions and are in
 
 ## Feature-token activation trace
 
-For a selected SAE feature and layer, v0.5 can encode every prompt-token residual and report the feature's TopK activation token by token. This answers a basic localization question that a single selected-token view cannot: is the feature concentrated at one syntactic/semantic position, or does it recur across the prompt?
+For a selected SAE feature and layer, FeatureLens can encode every prompt-token residual and report the feature's TopK activation token by token. This answers a basic localization question that a single selected-token view cannot: is the feature concentrated at one syntactic/semantic position, or does it recur across the prompt?
 
 The trace reports activation at each token, active-token count, and the peak token. A zero entry means that feature is not present in that token's TopK SAE support.
 
@@ -274,7 +274,7 @@ The trace reports activation at each token, active-token count, and the peak tok
 
 The live **Feature evidence** tab also provides an exploratory concept contrast for one selected feature.
 
-It samples a small balanced batch from the same seven controlled concept groups, using only one wording from each paraphrase pair. Earlier versions sampled only each prompt's final token, which can miss a feature that is active elsewhere in the prompt. v0.5 instead defines each prompt-level feature score as:
+It samples a small balanced batch from the same seven controlled concept groups, using only one wording from each paraphrase pair. Earlier versions sampled only each prompt's final token, which can miss a feature that is active elsewhere in the prompt. the current live scan instead defines each prompt-level feature score as:
 
 ```text
 score(feature, prompt) = max over non-padding prompt tokens of z_feature
@@ -346,3 +346,32 @@ It then uses:
 A large point-estimate ratio alone is not sufficient for a strong causal-specificity narrative if paired uncertainty remains weak.
 
 FeatureLens keeps the raw rows even when the resulting conclusion is null, mixed, or contrary to the original hypothesis.
+
+
+## v0.6 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?**
+
+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:
+
+```text
+target_mean(f) = mean prompt-wide max over prompts in c
+other_mean(f)  = mean prompt-wide max over prompts outside c
+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.
+
+## v0.6 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.
+
+The cue scan is intended to distinguish hypotheses such as:
+
+- concept-linked activation;
+- lexical activation tied to a particular token such as `is`;
+- structural activation at a completion boundary;
+- broad activation across several continuation cues.
+
+It is a controlled diagnostic only. A cue response cannot establish the feature's complete semantics.

docs/VALIDATION.md

@@ -1,8 +1,10 @@
-# FeatureLens v0.5.0 Validation
+# FeatureLens v0.6 validation
 
-This validation pass is intentionally focused on the **new v0.5 behavior and the core v0.4 regressions it could affect**. The full adversarial/edge-case battery is deferred to the final release, as requested.
+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.
 
-## A. Local release gate
+Use the exact UI labels below.
+
+## A. Local release gates
 
 From the repository root:
 
@@ -16,50 +18,33 @@ python3 scripts/release_check.py
 Expected:
 
 ```text
-38 passed
-FeatureLens UI launch smoke: PASS (...)
+40 passed
+FeatureLens UI launch smoke: PASS
 FeatureLens release check: PASS
   discovery prompts: 224
   causal tasks: 28
   layers: [4, 14, 26]
   feature-set sizes: [1, 3, 5]
   random controls: 8
-  release: v0.5.0
+  release: v0.6.0
 ```
 
-The UI smoke is important: it executes the real Gradio `launch()` path, which is where the v0.4 theme-font error occurred.
+## B. Navigation / typography / plot controls
 
----
+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.
 
-## B. Hugging Face startup and layout
+Pass condition: no heading level appears smaller than the body text it introduces, tables remain readable, and every native plot has fullscreen/export controls.
 
-After pushing, inspect the container log.
+## C. Establish the shared Workbench context
 
-Expected startup:
-
-```text
-* Running on local URL:  http://0.0.0.0:7860
-```
-
-There should be no SSR line, un-awaited coroutine warning, or theme/font traceback.
-
-### Layout acceptance
-
-On a desktop-width browser:
-
-- The application should be **centered**, not pinned to the left edge.
-- The usable canvas should expand substantially farther across a wide monitor than v0.4.
-- Explicit action buttons and **Copy table with headers** buttons should use the same muted teal treatment.
-- Body copy, labels, controls, and headings should use a consistent serif-oriented typography system.
-- Tables should use the available width before horizontal scrolling is needed.
-
-The widest research tables may still need horizontal scrolling; that is preferable to wrapping numerical headings into unreadable multi-line columns.
-
----
-
-## C. Core Workbench inspection
-
-In **Workbench** use:
+In **Workbench → I. Inspect a prompt location** set:
 
 ```text
 Prompt: The derivative of x squared is
@@ -70,239 +55,170 @@ Displayed active features: 12
 
 Click **Inspect sparse features**.
 
-Expected regression values should remain approximately:
+Expected regression values are approximately:
 
 ```text
-prompt token: 5
-Reconstruction cosine: 0.9015
+selected token: 5 (' is')
+feature 22632 activation: 32.90625
+reconstruction cosine: 0.9015
 NMSE: 0.1874
 Top-5 activation mass: 26.7%
-Top feature: 22632
-Top activation: 32.90625
 ```
 
-The **Strongest active SAE features** table and **Activation profile** should populate.
-
-### Copy feedback
+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.
 
-Click **Copy table with headers** below the feature table.
+## D. Explicit experiment feature selectors
 
-Pass condition:
+### D1. Scale dose-response
 
-1. Clipboard content starts with the header row, e.g.
-   ```text
-   Rank<TAB>Feature id<TAB>Activation<TAB>Offline concept hint
-   ```
-2. The clicked button briefly changes to:
-   ```text
-   ✓ Copied with headers
-   ```
-   and then returns to **Copy table with headers**.
+Open **Workbench → III. Single-feature scale dose-response → Scale dose-response**.
 
----
-
-## D. Dynamic-height / embedded scrolling regression
-
-Still in **Workbench**, select feature `22632` and run:
+Confirm there is a visible field named:
 
 ```text
-Single-feature intervention: ablate
-Target continuation (optional): 2x
-Greedy generation length: 8
+Dose-response feature id
 ```
 
-Click **Run single-feature causal test**.
-
-After the result expands, **do not switch tabs**.
-
-Immediately try to scroll to:
+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.
 
-- **III. Single-feature scale dose-response**
-- **IV. Contrastive continuation preference**
-- and eventually the bottom of the Workbench tab.
+The `1.0` multiplier row must remain the exact numerical null:
 
-Pass condition: the page continues scrolling after the dynamic result appears. A tab switch should not be required to recover the missing lower portion of the page.
-
-v0.5 installs a browser-side resize/mutation observer specifically to request reflow after dynamic output-height changes. If the page still gets stuck, record whether the failure occurs on the Hugging Face Space page only or also when opening the direct `*.hf.space` app URL.
-
----
+```text
+Δ feature coefficient = 0
+Perturbation L2 = 0
+Δ mean log p/token = 0
+Δ sequence log p = 0
+Next-token JS = 0
+```
 
-## E. Scale dose-response numerical-null regression
+### D2. Contrastive continuation preference
 
-Open **Scale dose-response** and click **Run scale dose-response** using target continuation `2x`.
+Open **Workbench → IV. Contrastive continuation preference → Contrastive causal preference test**.
 
-The `1.0` row must remain the exact batched numerical reference:
+Confirm there is now a visible field named:
 
 ```text
-Multiplier: 1.0
-Δ feature coefficient: 0.0
-Perturbation L2: 0.0
-Δ mean log p/token: 0.0
-Δ sequence log p: 0.0
-Next-token JS: 0.0
+Contrastive feature id
 ```
 
-This is a hard regression requirement.
-
----
-
-## F. Contrastive continuation preference — new v0.5 test
-
-In **Workbench → IV. Contrastive continuation preference**, use:
+Set:
 
 ```text
+Contrastive feature id: 22632
 Continuation A (preferred): 2x
 Continuation B (comparison): x
 Contrastive intervention: ablate
 ```
 
-Use feature `22632`, layer `14`, prompt token `5` inherited from the Workbench.
-
 Click **Run contrastive preference test**.
 
-Expected outputs:
-
-- **Contrastive continuation scores** with two rows: `A (preferred)` and `B (comparison)`.
-- Exact sequence log-probability and mean log-probability/token for both continuations.
-- Baseline A−B sequence log-odds.
-- SAE-edited A−B sequence log-odds.
-- Causal A−B log-odds shift.
-- 8-control random magnitude mean/std, specificity ratio, and empirical tail p.
-- **Preference between exact continuations** chart.
+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.
 
-There is no predetermined scientific sign. The test passes if the quantities are finite and internally coherent.
+## E. Concept-guided candidate feature discovery
 
-Interpretation rule: a strong JS shift with a weak contrastive preference shift supports **general distributional influence** more than selective preference control.
+Open **Feature evidence → A. Concept-guided candidate discovery**.
 
----
-
-## G. Feature-token activation trace — new v0.5 test
-
-Open **Feature evidence**. The Workbench inspection should have populated:
+Set:
 
 ```text
-Feature id: 22632
+Target concept: mathematics
 Residual layer: 14
+Prompts per concept: 4
+Candidate features: 12
 ```
 
-Click **Trace feature across prompt tokens**.
-
-For the canonical prompt, the output should contain one row per prompt token. The final selected token (`is`, token index `5`) should reproduce the known feature activation approximately:
-
-```text
-Feature 22632 at token 5 ≈ 32.90625
-```
-
-Expected outputs:
+Click **Discover concept-associated candidates**.
 
-- active token count;
-- peak token position and activation;
-- **Feature activation by prompt token** table;
-- **Feature activation across prompt tokens** chart.
+Verify:
 
-This test directly reveals whether a feature is localized to one token or distributed across the prompt.
+- 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.
 
----
+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.
 
-## H. Prompt-wide controlled concept contrast — corrected v0.5 test
+## F. Reuse a discovered candidate
 
-In **Feature evidence** use:
+After Test E, choose any value in **Discovered candidate feature id** and click:
 
 ```text
-Feature id: 22632
-Residual layer: 14
-Prompts per concept: 4
+Use candidate in feature experiments
 ```
 
-Click **Run controlled concept contrast**.
+Verify that the same id appears in:
 
-v0.5 no longer evaluates only the final token. For each sampled prompt it takes the feature's **maximum activation across all non-padding prompt tokens**.
+- **Workbench → Single feature id**;
+- **Workbench → Dose-response feature id**;
+- **Workbench → Contrastive feature id**;
+- **Feature evidence → Feature id**.
 
-The table columns should be:
+This is a UI/state-routing test only; it should not invoke the GPU.
 
-```text
-Concept
-Prompts
-Mean prompt-wide max
-Median prompt-wide max
-Prompt activation rate
-Mean when active
-Max activation
-```
-
-Two outcomes are valid:
-
-### If at least one sampled prompt activates the feature
-The UI may report a highest-mean concept and a runner-up ratio.
+## G. Completion-cue sensitivity
 
-### If all sampled activations remain zero
-The UI must say that the feature was **inactive in every sampled prompt**. It must **not** report `code` (or any other concept) as a winner and must not print an undefined winner ratio.
+Open **Feature evidence → C. Completion-cue sensitivity**.
 
----
-
-## I. Feature-set decoder geometry — new v0.5 test
-
-Open **Feature sets** after the canonical Workbench inspection. Select the top three features:
+Use:
 
 ```text
-22632
-25695
-16369
+Feature id: 22632
+Residual layer: 14
+Prompt stem: The derivative of x squared
+Completion cues:
+is
+=
+:
+equals
+therefore
 ```
 
-Open **Feature-set decoder geometry** and click **Inspect selected-feature geometry**.
+Click **Run completion-cue scan**.
 
-Expected:
+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`.
 
-- exactly three pair rows (`3 choose 2`);
-- decoder cosine for each pair;
-- mean absolute decoder cosine;
-- maximum absolute decoder cosine;
-- activation-weighted joint-ablation L2;
-- independent-direction root-sum-square L2 reference;
-- geometry ratio.
+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.
 
-Interpretation:
+## H. Prompt-wide selected-feature contrast regression
 
-```text
-ratio < 1  -> net geometric cancellation
-ratio ≈ 1  -> approximately independent/orthogonal net norm
-ratio > 1  -> net geometric alignment
-```
+Still in **Feature evidence**, use feature `22632`, layer `14`, 4 prompts/concept and click **Run controlled concept contrast**.
 
-This does not replace the individual-vs-joint causal test; it helps separate decoder geometry from downstream non-linearity.
+Reference v0.5 behavior:
 
----
+```text
+active in 7/28 prompts
+uncertainty mean ≈ 7.55
+mathematics mean ≈ 5.42
+french_language mean = 0
+```
 
-## J. Existing feature-set and paraphrase regressions
+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?*
 
-A full rerun is not necessary for every v0.4 panel. Do these quick checks:
+## I. Quick existing regressions
 
-1. **Run 1/3/5-feature ablation sweep** still produces rows for `k = 1, 3, 5` and 8-control statistics.
-2. **Run individual-vs-joint decomposition** still produces three individual rows plus one joint row for the selected top three features.
-3. In **Paraphrase robustness**, identical prompt vs identical prompt should still return:
-   ```text
-   selected-token Jaccard = 1.000
-   selected-token cosine = 1.000
-   prompt-wide Jaccard = 1.000
-   prompt-wide cosine = 1.000
-   ```
-4. **Layer trajectory → Compare layers** still returns layers `4`, `14`, and `26`.
+You do not need to 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.
 
-## Deferred final-release tests
+## J. Deferred final hardening
 
-The comprehensive adversarial suite remains intentionally deferred until the project is feature-complete. The final release should include, at minimum:
+Do not spend ZeroGPU quota on the complete adversarial suite yet. Keep these for the final release:
 
-- empty prompts/targets;
-- invalid token indices;
-- feature IDs `0`, `32767`, and invalid `32768`;
-- inactive feature ablation/scale/injection semantics;
-- whitespace-sensitive continuations (`2x` vs ` 2x`);
-- unrelated and negated paraphrases;
-- mixed-language prompts;
+- empty prompts;
+- out-of-range token indices;
+- feature ids `0`, `32767`, invalid `32768`;
+- inactive feature ablation vs injection;
+- whitespace-sensitive targets such as `2x` vs ` 2x`;
+- missing required target continuations;
+- unrelated paraphrases;
+- negation and mixed-language prompts;
 - long-prompt truncation;
-- rapid repeated clicks / ZeroGPU queue behavior;
-- narrow/mobile layout;
-- direct Space URL vs Hugging Face embedded-page scrolling.
+- rapid repeated clicks / queue behavior;
+- narrow/mobile-width layout;
+- final table-copy and plot fullscreen/export sweep.

featurelens/runtime.py

@@ -227,6 +227,28 @@ class ContrastiveCausalResult:
     random_control_count: int
 
 
+@dataclass
+class ConceptFeatureDiscoveryResult:
+    concept: str
+    layer: int
+    prompts_per_concept: int
+    top_n: int
+    rows: list[list[object]]
+    chart_rows: list[list[object]]
+    candidate_ids: list[int]
+
+
+@dataclass
+class FeatureCueScanResult:
+    feature_id: int
+    layer: int
+    prompt_stem: str
+    rows: list[list[object]]
+    chart_rows: list[list[object]]
+    active_cue_count: int
+    cue_count: int
+
+
 @dataclass
 class ParaphraseResult:
     tokens_a: list[str]
@@ -1749,6 +1771,199 @@ class FeatureLensRuntime:
             total_prompt_count=len(values),
         )
 
+    @torch.inference_mode()
+    def concept_feature_discovery(
+        self,
+        concept: str,
+        layer: int,
+        prompts_per_concept: int | None = None,
+        top_n: int = 12,
+    ) -> ConceptFeatureDiscoveryResult:
+        """Find SAE features whose prompt-wide activation is higher for one controlled concept."""
+        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()
+        if not concept:
+            raise ValueError('Choose a target concept.')
+        n = int(prompts_per_concept or self.settings.contrast_prompts_per_concept)
+        if n < 1 or n > 8:
+            raise ValueError('Discovery prompts per concept must be between 1 and 8.')
+        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.')
+        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}.')
+
+        texts = [str(row['text']) for row in rows]
+        batch = self.tokenizer(
+            texts,
+            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()
+        valid = attention.unsqueeze(-1).expand_as(encoding.values)
+        values = torch.where(valid, encoding.values, torch.zeros_like(encoding.values)).float()
+
+        dense = torch.zeros(
+            (values.shape[0], self.settings.sae_width),
+            device=values.device,
+            dtype=torch.float32,
+        )
+        dense.scatter_reduce_(
+            1,
+            encoding.indices.reshape(values.shape[0], -1),
+            values.reshape(values.shape[0], -1),
+            reduce='amax',
+            include_self=True,
+        )
+        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_mean = target.mean(dim=0)
+        other_mean = other.mean(dim=0)
+        target_rate = (target > 0).float().mean(dim=0)
+        other_rate = (other > 0).float().mean(dim=0)
+        mean_diff = target_mean - other_mean
+        selectivity = mean_diff / (target_mean + other_mean + 1e-8)
+
+        eligible = (target_mean > 0) & (mean_diff > 0)
+        candidate_idx = torch.nonzero(eligible, as_tuple=False).flatten()
+        if candidate_idx.numel() == 0:
+            return ConceptFeatureDiscoveryResult(
+                concept=concept,
+                layer=int(layer),
+                prompts_per_concept=n,
+                top_n=top_n,
+                rows=[],
+                chart_rows=[],
+                candidate_ids=[],
+            )
+        order = torch.argsort(mean_diff[candidate_idx], descending=True)
+        candidate_idx = candidate_idx[order[:top_n]]
+
+        table_rows: list[list[object]] = []
+        chart_rows: list[list[object]] = []
+        for rank, feature_tensor in enumerate(candidate_idx, start=1):
+            fid = int(feature_tensor.item())
+            row = [
+                rank,
+                fid,
+                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()),
+            ]
+            table_rows.append(row)
+            chart_rows.append([str(fid), float(mean_diff[fid].item())])
+        return ConceptFeatureDiscoveryResult(
+            concept=concept,
+            layer=int(layer),
+            prompts_per_concept=n,
+            top_n=top_n,
+            rows=table_rows,
+            chart_rows=chart_rows,
+            candidate_ids=[int(x.item()) for x in candidate_idx],
+        )
+
+    @staticmethod
+    def _cue_prompt(stem: str, cue: str) -> str:
+        stem = stem.rstrip()
+        cue = cue.strip()
+        if not cue:
+            return stem
+        if cue[0] in ':;,.!?=)]}':
+            return stem + cue
+        return stem + ' ' + cue
+
+    @torch.inference_mode()
+    def feature_cue_scan(
+        self,
+        feature_id: int,
+        layer: int,
+        prompt_stem: str,
+        cues: Sequence[str],
+    ) -> FeatureCueScanResult:
+        """Measure one feature at the final token after appending controlled completion cues."""
+        if not prompt_stem.strip():
+            raise ValueError('Enter a prompt stem.')
+        cue_list: list[str] = []
+        seen: set[str] = set()
+        for raw in cues:
+            cue = str(raw).strip()
+            if not cue or cue in seen:
+                continue
+            seen.add(cue)
+            cue_list.append(cue)
+        if not cue_list:
+            raise ValueError('Enter at least one cue.')
+        if len(cue_list) > 12:
+            raise ValueError('Cue scan supports at most 12 cues 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
+        prompts = [self._cue_prompt(prompt_stem, cue) for cue in cue_list]
+        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
+        for row_idx, (cue, full_prompt) in enumerate(zip(cue_list, prompts, strict=True)):
+            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)
+            token_id = int(batch['input_ids'][row_idx, final_pos].item())
+            final_token = self.tokenizer.decode([token_id])
+            rows.append([cue, full_prompt, repr(final_token), activation, active])
+            chart_rows.append([cue, activation])
+
+        return FeatureCueScanResult(
+            feature_id=int(feature_id),
+            layer=int(layer),
+            prompt_stem=prompt_stem,
+            rows=rows,
+            chart_rows=chart_rows,
+            active_cue_count=active_count,
+            cue_count=len(cue_list),
+        )
+
 
 RUNTIME = FeatureLensRuntime()
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "featurelens"
-version = "0.5.0"
+version = "0.6.0"
 description = "Causal sparse-feature interpretability workbench for Qwen3 and Qwen-Scope SAEs"
 requires-python = ">=3.10"
 

research_config.json

@@ -89,5 +89,16 @@
     "feature_token_activation_trace",
     "contrastive_continuation_preference_test",
     "feature_decoder_geometry"
+  ],
+  "concept_candidate_discovery_metric": "prompt-wide target-minus-other mean max activation; exploratory only",
+  "completion_cue_scan": "final-token feature activation after controlled suffix/cue substitution",
+  "live_features_v0_6": [
+    "start_here_plain_language_onboarding",
+    "persistent_workbench_context_banner",
+    "explicit_per_experiment_feature_selectors",
+    "plot_fullscreen_and_export_controls",
+    "consistent_heading_and_table_typography",
+    "concept_guided_candidate_feature_discovery",
+    "completion_cue_sensitivity_scan"
   ]
 }

scripts/release_check.py

@@ -120,6 +120,22 @@ def check_config(config: dict) -> None:
             f'{sorted(actual_live_v05)}'
         )
 
+    required_live_v06 = {
+        'start_here_plain_language_onboarding',
+        'persistent_workbench_context_banner',
+        'explicit_per_experiment_feature_selectors',
+        'plot_fullscreen_and_export_controls',
+        'consistent_heading_and_table_typography',
+        'concept_guided_candidate_feature_discovery',
+        'completion_cue_sensitivity_scan',
+    }
+    actual_live_v06 = set(config.get('live_features_v0_6', []))
+    if actual_live_v06 != required_live_v06:
+        raise SystemExit(
+            'research_config.json live_features_v0_6 mismatch: '
+            f'{sorted(actual_live_v06)}'
+        )
+
 
 def check_datasets(config: dict) -> tuple[list[dict], list[dict]]:
     prompts = load_jsonl(ROOT / 'data' / 'prompts.jsonl')
@@ -176,23 +192,26 @@ def check_readme() -> None:
         'full-continuation',
         'feature-set',
         'paraphrase',
-        'random ensemble',
+        'random-control ensemble',
         'batched zero-edit',
         'concept contrast',
         'non-additivity',
         'contrastive',
         'decoder geometry',
         'token activation',
+        'concept-guided candidate',
+        'completion-cue',
+        '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 v0.5 content: {missing}')
+        raise SystemExit(f'README.md is missing required v0.6 content: {missing}')
 
 
 def check_pyproject() -> None:
     text = (ROOT / 'pyproject.toml').read_text(encoding='utf-8')
-    if 'version = "0.5.0"' not in text:
-        raise SystemExit('pyproject.toml must declare version 0.5.0.')
+    if 'version = "0.6.0"' not in text:
+        raise SystemExit('pyproject.toml must declare version 0.6.0.')
 
 
 def main() -> None:
@@ -210,7 +229,7 @@ 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: v0.5.0')
+    print('  release: v0.6.0')
 
 
 if __name__ == '__main__':

tests/test_live_runtime_helpers.py

@@ -214,3 +214,28 @@ def test_concept_contrast_promptwide_scan_runs_on_toy_runtime() -> None:
     assert len(result.rows) == 7
     assert all(len(row) == 7 for row in result.rows)
     assert 0 <= result.active_prompt_count <= result.total_prompt_count
+
+
+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
+    )
+    assert result.concept == 'mathematics'
+    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)
+
+
+def test_feature_cue_scan_runs_on_toy_runtime() -> None:
+    runtime = make_runtime()
+    result = runtime.feature_cue_scan(
+        feature_id=0,
+        layer=0,
+        prompt_stem='abc',
+        cues=['is', '=', ':'],
+    )
+    assert result.cue_count == 3
+    assert len(result.rows) == 3
+    assert all(len(row) == 5 for row in result.rows)
+    assert 0 <= result.active_cue_count <= result.cue_count