Add reflective agent loop and hidden gem evidence

README.md

@@ -42,9 +42,11 @@ The app performs a multi-step research loop:
 2. Plan multiple targeted Hub searches.
 3. Deduplicate and pre-rank candidates by request relevance.
 4. Inspect dataset cards, tags, configurations, splits, schema fields, and sample rows.
-5. Run explicit modality, language, required-field, license, and accessibility checks.
-6. Rank candidates from evidence and connect potentially complementary datasets.
-7. Stream the trace and explain verified strengths, limitations, and rejection reasons.
+5. Reflect on first-pass failures and run a second targeted search when the evidence is weak.
+6. Run explicit modality, language, required-field, sample-row, license, and accessibility checks.
+7. Highlight hidden gems when low-adoption datasets have strong verified fit.
+8. Rank candidates from evidence and connect potentially complementary datasets.
+9. Stream the trace and explain verified strengths, limitations, and rejection reasons.
 
 `HuggingFaceTB/SmolLM2-360M-Instruct` runs locally inside the Space CPU runtime and helps interpret
 the brief. At only 360M parameters, it is comfortably below the Tiny Titan 4B limit. The model
@@ -70,6 +72,16 @@ Missing evidence is shown as `unknown` rather than silently converted into an av
 Hard modality, language, accessibility, or required-schema mismatches produce explicit rejection
 reasons.
 
+## Agent loop
+
+HF Agentic Search does not stop after one keyword pass. It runs an initial search plan, inspects
+real Hugging Face evidence, reflects on what failed, and launches a bounded second-pass search
+aimed at the missing signal. For example, if the first climate search finds reports but not
+question-answer rows, the agent switches to schema-first queries such as `climate qa dataset`.
+
+The final result includes visible trace events, sample-row tests, hidden-gem labels, rejection
+reasons, and starter `load_dataset` code for the selected dataset.
+
 ## Architecture
 
 - **Agent:** Python, `huggingface_hub`, Hub API, and Dataset Viewer API

backend/agent.py

@@ -452,6 +452,122 @@ def _matches_field(requirement: str, field: str) -> bool:
     return normalized in aliases
 
 
+def _field_value(row: dict[str, Any], fields: list[str]) -> Any:
+    lowered = {str(key).lower(): value for key, value in row.items()}
+    for field in fields:
+        value = lowered.get(str(field).lower())
+        if value not in (None, "", [], {}):
+            return value
+    return None
+
+
+def _sample_tests(
+    profile: dict[str, Any],
+    dataset: dict[str, Any],
+    matched_requirements: dict[str, list[str]],
+) -> list[dict[str, str]]:
+    rows = [row for row in dataset.get("sample_rows", []) if isinstance(row, dict)]
+    required_fields = profile["required_fields"]
+    tests: list[dict[str, str]] = []
+    if not rows:
+        return [{
+            "name": "Sample rows available",
+            "status": "unknown",
+            "detail": "Dataset Viewer did not expose sample rows.",
+        }]
+
+    missing = [
+        requirement for requirement in required_fields
+        if not matched_requirements.get(requirement)
+    ]
+    empty = [
+        requirement for requirement in required_fields
+        if matched_requirements.get(requirement)
+        and not any(_field_value(row, matched_requirements[requirement]) is not None for row in rows)
+    ]
+    tests.append({
+        "name": "Required fields populated",
+        "status": "pass" if not missing and not empty else "fail" if missing else "review",
+        "detail": "All requested fields have sample values."
+        if not missing and not empty
+        else f"Missing: {', '.join(missing or empty)}.",
+    })
+
+    if {"question", "answer"} & set(required_fields):
+        question_fields = matched_requirements.get("question", [])
+        answer_fields = matched_requirements.get("answer", [])
+        question_values = [
+            str(_field_value(row, question_fields) or "") for row in rows
+            if _field_value(row, question_fields) is not None
+        ]
+        answer_values = [
+            str(_field_value(row, answer_fields) or "") for row in rows
+            if _field_value(row, answer_fields) is not None
+        ]
+        plausible_questions = sum(
+            1 for value in question_values
+            if value.endswith("?") or len(value.split()) >= 3
+        )
+        distinct_answers = sum(
+            1 for question, answer in zip(question_values, answer_values)
+            if answer and answer.strip().lower() != question.strip().lower()
+        )
+        tests.append({
+            "name": "QA row shape",
+            "status": "pass" if plausible_questions and distinct_answers else "review",
+            "detail": "Questions and answers look usable in inspected rows."
+            if plausible_questions and distinct_answers
+            else "Question/answer shape needs manual review.",
+        })
+
+    if "label" in required_fields:
+        label_fields = matched_requirements.get("label", [])
+        label_values = [
+            _field_value(row, label_fields) for row in rows
+            if _field_value(row, label_fields) is not None
+        ]
+        unique_labels = {str(value) for value in label_values}
+        tests.append({
+            "name": "Label signal",
+            "status": "pass" if label_values and len(unique_labels) <= max(20, len(rows)) else "review",
+            "detail": f"Found {len(unique_labels)} inspected label value(s)."
+            if label_values else "No label values were visible in sample rows.",
+        })
+
+    if {"document", "summary"}.issubset(required_fields):
+        document_fields = matched_requirements.get("document", [])
+        summary_fields = matched_requirements.get("summary", [])
+        shorter = 0
+        for row in rows:
+            document = str(_field_value(row, document_fields) or "")
+            summary = str(_field_value(row, summary_fields) or "")
+            if document and summary and len(summary) < len(document):
+                shorter += 1
+        tests.append({
+            "name": "Summary shape",
+            "status": "pass" if shorter else "review",
+            "detail": "Summaries are shorter than source documents in inspected rows."
+            if shorter else "Summary/document length relationship needs review.",
+        })
+
+    return tests
+
+
+def _loader_snippet(dataset: dict[str, Any]) -> str:
+    dataset_id = dataset.get("id", "")
+    config = (dataset.get("configs") or [None])[0]
+    split = (dataset.get("splits") or ["train"])[0]
+    args = f'"{dataset_id}"'
+    if config and config != "default":
+        args += f', "{config}"'
+    return (
+        "from datasets import load_dataset\n\n"
+        f'ds = load_dataset({args})\n'
+        f'rows = ds["{split}"]\n'
+        "print(rows[0])"
+    )
+
+
 def score_dataset(profile: dict[str, Any], dataset: dict[str, Any]) -> dict[str, Any]:
     """Compute a transparent score entirely from collected evidence."""
     blob = _text_blob(dataset)
@@ -546,6 +662,10 @@ def score_dataset(profile: dict[str, Any], dataset: dict[str, Any]) -> dict[str,
         })
         schema_evidence = "missing" if available_fields else "unknown"
 
+    sample_tests = _sample_tests(profile, dataset, matched_requirements)
+    sample_test_passes = sum(1 for test in sample_tests if test["status"] == "pass")
+    sample_test_total = len(sample_tests)
+
     license_value = dataset.get("license", "")
     permissive = {"apache-2.0", "mit", "cc-by-4.0", "cc0-1.0", "odc-by", "bsd-3-clause"}
     license_score = 10 if license_value in permissive else 5 if license_value else 0
@@ -668,6 +788,20 @@ def score_dataset(profile: dict[str, Any], dataset: dict[str, Any]) -> dict[str,
     quality = round(
         (schema + license_score + documentation + popularity + accessibility) / 40 * 100
     )
+    low_adoption = (dataset.get("downloads", 0) < 2_000 and dataset.get("likes", 0) < 25)
+    hidden_gem = (
+        status != "rejected"
+        and total >= 60
+        and low_adoption
+        and checks["domain"] == "pass"
+        and checks["required_fields"] in {"pass", "review"}
+        and checks["accessible"] == "pass"
+    )
+    badges = ["hidden_gem"] if hidden_gem else []
+    discovery_note = (
+        "Hidden gem: low adoption, but the inspected evidence fits this brief."
+        if hidden_gem else ""
+    )
     return {
         **dataset,
         "score": total,
@@ -687,12 +821,17 @@ def score_dataset(profile: dict[str, Any], dataset: dict[str, Any]) -> dict[str,
             "accessibility": accessibility,
         },
         "checks": checks,
+        "sample_tests": sample_tests,
+        "sample_test_summary": f"{sample_test_passes}/{sample_test_total} sample tests passed",
         "schema_evidence": schema_evidence,
         "evidence": evidence,
         "rejection_reasons": rejection_reasons,
         "strength": strength,
         "weakness": weakness,
         "recommendation": recommendation,
+        "badges": badges,
+        "discovery_note": discovery_note,
+        "loader_snippet": _loader_snippet(dataset),
     }
 
 
@@ -714,16 +853,115 @@ def _cross_reference(datasets: list[dict[str, Any]]) -> list[dict[str, str]]:
     return pairs
 
 
-def _rank_key(profile: dict[str, Any], dataset: dict[str, Any]) -> tuple[int, int, int]:
+def _rank_key(profile: dict[str, Any], dataset: dict[str, Any]) -> tuple[int, int, int, int]:
     checks = dataset["checks"]
+    hidden_gem = 1 if "hidden_gem" in dataset.get("badges", []) else 0
+    sample_passes = sum(1 for test in dataset.get("sample_tests", []) if test.get("status") == "pass")
     evidence_fit = (
         (2 if checks["required_fields"] == "pass" else 0)
         + (1 if checks["domain"] == "pass" else 0)
         + (2 if profile["languages"] and checks["language"] == "pass" else 0)
         + (2 if profile["license"] and checks["license"] == "pass" else 0)
+        + sample_passes
     )
     status_rank = {"recommended": 2, "conditional": 1, "rejected": 0}[dataset["status"]]
-    return status_rank, evidence_fit, dataset["score"]
+    return status_rank, hidden_gem, evidence_fit, dataset["score"]
+
+
+def _select_candidates_for_profile(
+    profile: dict[str, Any],
+    collected: dict[str, dict[str, Any]],
+    search_batches: list[list[str]],
+    limit: int,
+) -> list[dict[str, Any]]:
+    global_ranked = sorted(
+        collected.values(),
+        key=lambda dataset: _pre_score(profile, dataset),
+        reverse=True,
+    )
+    diversified_ids: list[str] = []
+    diversified_seen: set[str] = set()
+    for position in range(8):
+        for batch in search_batches:
+            if position >= len(batch):
+                continue
+            dataset_id = batch[position]
+            if dataset_id not in diversified_seen:
+                diversified_seen.add(dataset_id)
+                diversified_ids.append(dataset_id)
+    for dataset in global_ranked:
+        if dataset["id"] not in diversified_seen:
+            diversified_ids.append(dataset["id"])
+    return [
+        collected[dataset_id]
+        for dataset_id in diversified_ids[:limit]
+        if dataset_id in collected
+    ]
+
+
+def _reflect_on_results(
+    profile: dict[str, Any],
+    tried_queries: list[str],
+    inspected: list[dict[str, Any]],
+) -> dict[str, Any]:
+    if not inspected:
+        summary = "The first search did not produce inspectable datasets, so the agent is broadening the query language."
+        strategy = "broaden search"
+    else:
+        failures = {
+            key: sum(1 for dataset in inspected if dataset.get("checks", {}).get(key) in {"fail", "unknown"})
+            for key in ("domain", "required_fields", "language", "license", "accessible")
+        }
+        worst = max(failures, key=failures.get)
+        if worst == "required_fields":
+            summary = "The first pass found topical datasets, but too many missed the requested schema."
+            strategy = "schema-first search"
+        elif worst == "domain":
+            summary = "The first pass found task-shaped datasets, but several were off-topic."
+            strategy = "domain-first search"
+        elif worst == "language":
+            summary = "The first pass found candidates with weak language evidence."
+            strategy = "language-focused search"
+        else:
+            summary = "The first pass found partial fits; the agent is looking for less obvious alternatives."
+            strategy = "hidden-gem search"
+
+    terms = _domain_terms(profile)
+    domain = terms[:2]
+    required = profile["required_fields"]
+    next_queries: list[str] = []
+    if domain and {"question", "answer"}.issubset(required):
+        next_queries.extend([
+            f"{domain[0]} qa dataset",
+            " ".join(domain + ["question answer"]),
+            " ".join(domain + ["benchmark"]),
+        ])
+    if domain and "label" in required:
+        next_queries.extend([
+            f"{domain[0]} labeled dataset",
+            f"{domain[0]} intent labels",
+        ])
+    if domain and "summary" in required:
+        next_queries.extend([
+            " ".join(domain + ["summaries"]),
+            " ".join(domain + ["summarization dataset"]),
+        ])
+    if profile["languages"] and domain:
+        next_queries.append(" ".join([profile["languages"][0], domain[0], "dataset"]))
+    if domain:
+        next_queries.append(" ".join(domain + ["data"]))
+    next_queries.extend(generate_queries("", profile))
+    cleaned = []
+    for query in next_queries:
+        normalized = re.sub(r"\s+", " ", query).strip()
+        if normalized and normalized not in tried_queries and normalized not in cleaned:
+            cleaned.append(normalized)
+    return {
+        "summary": summary,
+        "strategy": strategy,
+        "next_queries": cleaned[:4],
+        "reason": "Reflection is based on failed checks from the first inspected batch.",
+    }
 
 
 def weave_events(task: str, max_datasets: int = 8) -> Iterator[dict[str, Any]]:
@@ -748,7 +986,13 @@ def weave_events(task: str, max_datasets: int = 8) -> Iterator[dict[str, Any]]:
 
     collected: dict[str, dict[str, Any]] = {}
     search_batches: list[list[str]] = []
-    for query in queries:
+    inspected: list[dict[str, Any]] = []
+    inspected_ids: set[str] = set()
+
+    first_pass_queries = queries[:5]
+    reserve_queries = queries[5:]
+
+    for query in first_pass_queries:
         found = search_datasets(query, limit=35)
         search_batches.append([dataset["id"] for dataset in found])
         for dataset in found:
@@ -764,37 +1008,20 @@ def weave_events(task: str, max_datasets: int = 8) -> Iterator[dict[str, Any]]:
         }
 
     inspection_limit = max(max_datasets * 4, 28)
-    global_ranked = sorted(
-        collected.values(),
-        key=lambda dataset: _pre_score(profile, dataset),
-        reverse=True,
+    first_pass_limit = max(max_datasets * 2, 16)
+    pre_ranked = _select_candidates_for_profile(
+        profile,
+        collected,
+        search_batches,
+        first_pass_limit,
     )
-    diversified_ids: list[str] = []
-    diversified_seen: set[str] = set()
-    for position in range(8):
-        for batch in search_batches:
-            if position >= len(batch):
-                continue
-            dataset_id = batch[position]
-            if dataset_id not in diversified_seen:
-                diversified_seen.add(dataset_id)
-                diversified_ids.append(dataset_id)
-    for dataset in global_ranked:
-        if dataset["id"] not in diversified_seen:
-            diversified_ids.append(dataset["id"])
-    pre_ranked = [
-        collected[dataset_id]
-        for dataset_id in diversified_ids[:inspection_limit]
-        if dataset_id in collected
-    ]
     yield {
         "type": "search",
-        "query": "deep candidate pool",
+        "query": "first evidence pool",
         "found": len(pre_ranked),
         "unique": len(collected),
-        "message": f"Prepared {len(pre_ranked)} diverse candidates for evidence inspection.",
+        "message": f"Prepared {len(pre_ranked)} diverse candidates for first-pass inspection.",
     }
-    inspected: list[dict[str, Any]] = []
     with ThreadPoolExecutor(max_workers=min(4, max(1, len(pre_ranked)))) as pool:
         futures = {
             pool.submit(inspect_dataset, dataset["id"], dataset): dataset["id"]
@@ -816,6 +1043,7 @@ def weave_events(task: str, max_datasets: int = 8) -> Iterator[dict[str, Any]]:
                 }
             scored = score_dataset(profile, evidence)
             inspected.append(scored)
+            inspected_ids.add(dataset_id)
             yield {
                 "type": "inspect",
                 "dataset_id": dataset_id,
@@ -826,6 +1054,83 @@ def weave_events(task: str, max_datasets: int = 8) -> Iterator[dict[str, Any]]:
             }
             yield {"type": "candidate", "dataset": _public_dataset(scored)}
 
+    reflection = _reflect_on_results(profile, first_pass_queries, inspected)
+    second_pass_queries = list(dict.fromkeys(reflection["next_queries"] + reserve_queries))[:4]
+    yield {
+        "type": "reflect",
+        "summary": reflection["summary"],
+        "strategy": reflection["strategy"],
+        "next_queries": second_pass_queries,
+        "message": f"{reflection['summary']} Trying {len(second_pass_queries)} deeper search angle(s).",
+    }
+
+    for query in second_pass_queries:
+        found = search_datasets(query, limit=35)
+        search_batches.append([dataset["id"] for dataset in found])
+        new_count = 0
+        for dataset in found:
+            current = collected.get(dataset["id"])
+            if current is None:
+                new_count += 1
+            if current is None or _pre_score(profile, dataset) > _pre_score(profile, current):
+                collected[dataset["id"]] = dataset
+        yield {
+            "type": "search",
+            "query": query,
+            "found": len(found),
+            "unique": len(collected),
+            "message": f"Deepened with “{query}” and found {len(found)} candidates ({new_count} new).",
+        }
+
+    deeper_pool = [
+        dataset for dataset in _select_candidates_for_profile(
+            profile,
+            collected,
+            search_batches,
+            inspection_limit,
+        )
+        if dataset["id"] not in inspected_ids
+    ][:max(max_datasets * 2, 16)]
+    yield {
+        "type": "search",
+        "query": "deep candidate pool",
+        "found": len(deeper_pool),
+        "unique": len(collected),
+        "message": f"Prepared {len(deeper_pool)} fresh candidates after reflection.",
+    }
+    if deeper_pool:
+        with ThreadPoolExecutor(max_workers=min(4, len(deeper_pool))) as pool:
+            futures = {
+                pool.submit(inspect_dataset, dataset["id"], dataset): dataset["id"]
+                for dataset in deeper_pool
+            }
+            for future in as_completed(futures):
+                dataset_id = futures[future]
+                try:
+                    evidence = future.result()
+                except Exception as exc:
+                    evidence = {
+                        **next(item for item in deeper_pool if item["id"] == dataset_id),
+                        "accessible": False,
+                        "inspection_error": str(exc),
+                        "features": [],
+                        "sample_rows": [],
+                        "configs": [],
+                        "splits": [],
+                    }
+                scored = score_dataset(profile, evidence)
+                inspected.append(scored)
+                inspected_ids.add(dataset_id)
+                yield {
+                    "type": "inspect",
+                    "dataset_id": dataset_id,
+                    "status": scored["status"],
+                    "score": scored["score"],
+                    "checks": scored["checks"],
+                    "message": f"Inspected {dataset_id}: {scored['status']} ({scored['score']}/100).",
+                }
+                yield {"type": "candidate", "dataset": _public_dataset(scored)}
+
     ranked = sorted(
         inspected,
         key=lambda dataset: _rank_key(profile, dataset),
@@ -856,6 +1161,7 @@ def weave_events(task: str, max_datasets: int = 8) -> Iterator[dict[str, Any]]:
         "model_used": MODEL if llm_used else None,
         "fallback_used": not llm_used,
         "elapsed_ms": round((time.time() - started) * 1000),
+        "reflection": reflection,
     }
     yield {
         "type": "ranking",
@@ -874,6 +1180,7 @@ def _public_dataset(dataset: dict[str, Any]) -> dict[str, Any]:
         "accessible", "inspection_error", "card_complete", "num_examples", "score", "relevance",
         "quality", "status", "score_breakdown", "checks", "evidence",
         "schema_evidence", "rejection_reasons", "strength", "weakness", "recommendation",
+        "sample_tests", "sample_test_summary", "badges", "discovery_note", "loader_snippet",
     }
     return {key: value for key, value in dataset.items() if key in allowed}
 

frontend/dist/assets/index-BsUEW0Da.css

@@ -0,0 +1 @@
+@import"https://fonts.googleapis.com/css2?family=Geist+Mono:wght@400;500;600&family=Geist:wght@400;500;600;700&display=swap";:root{--paper: #f4faff;--paper-light: #ffffff;--ink: #102a43;--ink-soft: #2f465c;--line: #84b9df;--forest: #2489c9;--forest-dark: #12689d;--amber: #c47a18;--clay: #bd4b5c;--muted: #dceefa;--line-soft: #c6e0f2;--surface-blue: #e6f4fe;--surface-blue-strong: #d2ecfc;--font-display: "Geist", Arial, sans-serif;--font-body: "Geist", Arial, sans-serif;--font-mono: "Geist Mono", monospace;--radius-sm: 7px;--radius-md: 11px}*{box-sizing:border-box}html,body,#root{min-height:100%;margin:0}body{background:var(--paper);color:var(--ink);font-family:var(--font-body);-webkit-font-smoothing:antialiased}button,textarea{font:inherit}button,a{-webkit-tap-highlight-color:transparent}button:focus-visible,a:focus-visible,textarea:focus-visible{outline:3px solid #62b5e8;outline-offset:2px}.app-shell{min-height:100vh;display:flex;flex-direction:column}.topbar{min-height:60px;padding:0 24px;display:flex;align-items:center;justify-content:space-between;border-bottom:1px solid var(--line);background:var(--paper-light)}.brand{display:inline-flex;align-items:center;color:var(--ink);font-family:var(--font-display);font-size:18px;font-weight:650;text-decoration:none}.topbar-meta{display:flex;align-items:center;gap:24px;font:12px var(--font-mono)}.run-state{padding:8px 12px;border:1px solid var(--line);border-radius:var(--radius-sm);background:var(--surface-blue);text-transform:uppercase}.run-state.active{background:var(--surface-blue-strong);color:var(--forest-dark)}.workspace{flex:1;min-height:0;display:grid;grid-template-columns:minmax(285px,22vw) minmax(520px,1fr) minmax(315px,25vw);border-top:1px solid var(--line)}.rail{min-width:0;background:var(--paper-light);overflow-y:auto;max-height:none}.rail-left{border-right:1px solid var(--line)}.rail-right{border-left:1px solid var(--line)}.map-stage{min-width:0;min-height:760px;display:flex;flex-direction:column;background:linear-gradient(rgba(36,137,201,.055) 1px,transparent 1px),linear-gradient(90deg,rgba(36,137,201,.055) 1px,transparent 1px),var(--paper);background-size:32px 32px}.hero{padding:44px clamp(24px,5vw,72px) 40px;display:grid;grid-template-columns:minmax(0,1.25fr) minmax(420px,.75fr);gap:clamp(40px,7vw,100px);align-items:end;background:var(--paper-light)}.hero-copy{max-width:760px}.hero h1{max-width:700px;margin:10px 0 14px;font:650 clamp(36px,4vw,58px)/1.02 var(--font-display);letter-spacing:-.045em}.hero-copy>p{max-width:720px;margin:0;color:var(--ink-soft);font-size:clamp(15px,1.3vw,18px);line-height:1.55}.hero-principles{border-top:1px solid var(--line)}.hero-principles div{padding:13px 0;display:grid;grid-template-columns:118px 1fr;gap:16px;border-bottom:1px solid var(--line-soft)}.hero-principles strong{font-size:12px}.hero-principles span{color:var(--ink-soft);font-size:11px;line-height:1.45}.workspace-heading{min-height:46px;padding:14px 20px 10px;display:flex;justify-content:space-between;align-items:center;gap:20px;background:#f4fafff0}.results-heading h2{margin:5px 0 0;font:650 22px/1.1 var(--font-display);letter-spacing:-.025em}.mode-note{max-width:190px;padding:7px 9px;border:1px solid var(--line);border-radius:var(--radius-sm);background:var(--paper-light);font:10px/1.4 var(--font-mono)}.section-index,.field-label{display:block;color:var(--ink-soft);font:500 9px/1.4 var(--font-mono);letter-spacing:.09em;text-transform:uppercase}.search-section{padding:24px 20px 22px;border-bottom:1px solid var(--line)}.search-section h2{max-width:360px;margin:12px 0 10px;font:650 clamp(28px,2.4vw,38px)/1.02 var(--font-display);letter-spacing:-.04em}.lede{margin:0 0 24px;color:var(--ink-soft);font-size:13px;line-height:1.6}.search-section form{margin-top:20px}.search-section textarea{width:100%;margin-top:8px;padding:15px;resize:vertical;border:2px solid var(--line);border-radius:var(--radius-md);background:#fff;color:var(--ink);font-size:13px;line-height:1.55;box-shadow:0 4px 14px #1a689912}.search-section textarea::placeholder{color:#7890a5}.form-meta{margin:10px 0 14px;display:flex;justify-content:space-between;color:var(--ink-soft);font:10px var(--font-mono)}.primary-action{width:100%;min-height:52px;padding:0 16px;display:flex;align-items:center;justify-content:space-between;border:2px solid var(--forest-dark);border-radius:var(--radius-md);background:var(--forest);color:#fff;cursor:pointer;font-size:13px;font-weight:700;box-shadow:0 4px 0 var(--forest-dark);transition:transform .12s ease,box-shadow .12s ease}.primary-action:hover:not(:disabled){transform:translateY(2px);box-shadow:0 2px 0 var(--forest-dark)}.primary-action:disabled{cursor:wait;opacity:.65}.examples{margin-top:24px;display:grid;gap:7px}.examples .field-label{margin-bottom:2px}.examples button{padding:9px 0;display:grid;grid-template-columns:28px 1fr;gap:8px;border:0;border-top:1px solid var(--line-soft);background:transparent;color:var(--ink);text-align:left;cursor:pointer;font-size:11px;line-height:1.4}.examples button span{color:var(--amber);font-family:var(--font-mono)}.process-section{padding:20px 20px 26px}.method-list{margin:14px 0 0;padding:0;list-style:none;display:grid}.method-note{margin:10px 0 0;color:var(--ink-soft);font-size:11px;line-height:1.55}.method-list li{padding:10px 0;display:grid;grid-template-columns:28px 1fr;border-top:1px solid var(--line-soft);font-size:11px}.method-list li span{color:var(--forest);font:11px var(--font-mono)}.process-title{display:flex;justify-content:space-between;font:10px var(--font-mono)}.trace-summary{margin-top:12px;display:grid;grid-template-columns:repeat(4,1fr);border:1px solid var(--line);border-radius:var(--radius-sm);overflow:hidden;background:var(--paper-light)}.trace-summary div{min-width:0;padding:8px 6px;border-right:1px solid var(--line-soft)}.trace-summary div:last-child{border-right:0}.trace-summary span{display:block;color:var(--ink-soft);font:7px var(--font-mono);text-transform:uppercase}.trace-summary strong{display:block;margin-top:4px;font:650 15px var(--font-display)}.trace-focus{margin-top:10px;padding:10px;border:1px solid var(--line-soft);border-radius:var(--radius-sm);background:var(--surface-blue)}.trace-focus span{display:block;color:var(--ink-soft);font:8px var(--font-mono);text-transform:uppercase}.trace-focus strong{display:block;margin-top:5px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;font-size:11px}.event-feed{max-height:380px;margin-top:12px;overflow-y:auto}.event-row{padding:10px 0;display:grid;grid-template-columns:30px 1fr;gap:8px;border-top:1px solid var(--line-soft)}.event-number{color:var(--amber);font:10px var(--font-mono)}.event-row strong{font:600 10px var(--font-mono);text-transform:uppercase}.event-row p{margin:3px 0 0;color:var(--ink-soft);font-size:11px;line-height:1.45}.event-error p{color:var(--clay)}.event-meta{margin-top:7px;display:flex;flex-wrap:wrap;gap:5px}.event-meta span{padding:3px 5px;border:1px solid var(--line-soft);border-radius:999px;background:#fff;color:var(--ink-soft);font:7px var(--font-mono);text-transform:uppercase}.trace-cursor{padding:10px 0;color:var(--forest);font:10px var(--font-mono);animation:blink 1.2s infinite}@keyframes blink{50%{opacity:.35}}.brief-strip{margin:0 20px;padding:14px 16px 16px;border:1px solid var(--line);border-radius:var(--radius-md);background:#ffffffd1}.brief-strip-heading{display:flex;justify-content:space-between;gap:16px;color:var(--ink-soft);font:9px var(--font-mono)}.brief-grid{margin-top:12px;display:grid;grid-template-columns:repeat(5,minmax(0,1fr));gap:6px}.brief-item{min-width:0;padding:9px 10px;border-radius:var(--radius-sm);background:var(--surface-blue)}.brief-item span{display:block;color:var(--ink-soft);font:8px var(--font-mono);text-transform:uppercase}.brief-item strong{display:block;margin-top:5px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;font-size:10px;font-weight:600}.agent-progress{margin:12px 20px;display:grid;grid-template-columns:repeat(5,1fr);border:1px solid var(--line);border-radius:var(--radius-md);overflow:hidden;background:var(--paper-light)}.progress-step{min-width:0;padding:12px;display:grid;grid-template-columns:24px 1fr;gap:8px;border-right:1px solid var(--line-soft)}.progress-step:last-child{border-right:0}.progress-step.active{background:var(--surface-blue)}.progress-number{width:22px;height:22px;display:grid;place-items:center;border:1px solid var(--line);border-radius:50%;color:var(--ink-soft);font:9px var(--font-mono)}.progress-step.complete .progress-number{border-color:var(--forest);background:var(--forest);color:#fff}.progress-step strong{display:block;font-size:10px}.progress-step p{margin:3px 0 0;color:var(--ink-soft);font-size:8px;line-height:1.35}.research-grid{min-height:530px;margin:0 20px 20px;display:grid;grid-template-columns:minmax(0,1.7fr) minmax(190px,.8fr);gap:12px}.evidence-panel,.check-panel{min-width:0;display:flex;flex-direction:column;border:1px solid var(--line);border-radius:var(--radius-md);overflow:hidden;background:#ffffffdb}.panel-heading{min-height:70px;padding:14px 16px;display:flex;align-items:flex-start;justify-content:space-between;gap:12px;border-bottom:1px solid var(--line-soft)}.panel-heading h2{margin:6px 0 0;font:650 17px/1.15 var(--font-display);letter-spacing:-.025em}.panel-count{font:650 22px var(--font-display)}.candidate-board{flex:1;min-height:450px;padding:14px;display:flex;flex-direction:column;gap:12px;overflow:auto}.candidate-board.empty{display:grid;place-content:center;justify-items:center;text-align:center}.candidate-board.empty strong{margin-top:24px;font:650 18px var(--font-display)}.candidate-board.empty p{max-width:390px;color:var(--ink-soft);font-size:12px;line-height:1.6}.workflow-preview{display:flex;align-items:center;justify-content:center}.workflow-preview>div{width:112px;min-height:94px;padding:12px;display:flex;flex-direction:column;border:1px solid var(--line);border-radius:var(--radius-md);background:var(--paper-light);text-align:left}.workflow-preview span{color:var(--forest);font:8px var(--font-mono)}.workflow-preview strong{margin:9px 0 4px;font-size:12px}.workflow-preview small{color:var(--ink-soft);font-size:8px;line-height:1.4}.workflow-preview i{width:24px;height:1px;background:var(--line)}.decision-summary{display:grid;grid-template-columns:repeat(auto-fit,minmax(120px,1fr));gap:8px;align-items:stretch}.decision-stat{padding:9px 10px;border:1px solid var(--line-soft);border-radius:var(--radius-sm);background:#fff}.decision-stat span{display:block;color:var(--ink-soft);font:8px var(--font-mono);text-transform:uppercase}.decision-stat strong{display:block;margin-top:4px;font:650 20px var(--font-display)}.stat-best strong{color:var(--forest-dark)}.stat-hidden strong{color:var(--forest)}.stat-review strong{color:var(--amber)}.stat-filtered strong{color:var(--clay)}.decision-summary p{margin:0;padding:9px 11px;border:1px solid var(--line-soft);border-radius:var(--radius-sm);background:#ffffffbd;color:var(--ink-soft);font-size:9px;line-height:1.45;grid-column:1 / -1}.candidate-lanes{display:grid;grid-template-columns:repeat(auto-fit,minmax(220px,1fr));gap:10px;align-items:start}.candidate-lane{min-width:0;padding:10px;display:grid;gap:8px;border:1px solid var(--line-soft);border-radius:var(--radius-md);background:#ffffffbd}.lane-heading{min-height:46px;display:flex;justify-content:space-between;gap:10px;align-items:flex-start;border-bottom:1px solid var(--line-soft);padding-bottom:9px}.lane-heading strong{display:block;font-size:11px}.lane-heading span{display:block;margin-top:3px;color:var(--ink-soft);font-size:8px;line-height:1.35}.lane-heading b{font:650 18px var(--font-display)}.lane-best .lane-heading b{color:var(--forest-dark)}.lane-hidden .lane-heading b{color:var(--forest)}.lane-review .lane-heading b{color:var(--amber)}.lane-hidden{background:linear-gradient(180deg,#e6f4fee6,#ffffffc2)}.candidate-card{width:100%;padding:10px;border:1px solid transparent;border-radius:var(--radius-sm);background:#fff;color:var(--ink);text-align:left;cursor:pointer;box-shadow:0 1px #12689d14}.candidate-card:hover,.candidate-card.active{border-color:var(--line);background:#fff}.candidate-card.active{border-color:var(--forest);box-shadow:0 2px 10px #12689d1f}.candidate-card-top{display:grid;grid-template-columns:minmax(0,1fr) auto;gap:8px;align-items:start}.candidate-card-top strong{overflow:hidden;text-overflow:ellipsis;white-space:nowrap;font-size:10px}.candidate-card-top span{font:650 16px var(--font-display)}.candidate-badges{margin-top:6px;display:flex;flex-wrap:wrap;gap:4px}.candidate-badges span{padding:3px 6px;border:1px solid rgba(36,137,201,.35);border-radius:999px;background:var(--surface-blue);color:var(--forest-dark);font:7px var(--font-mono);text-transform:uppercase}.candidate-card p{min-height:34px;margin:7px 0;color:var(--ink-soft);font-size:9px;line-height:1.42}.candidate-card small{display:block;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;color:var(--ink-soft);font-size:8px}.check-dots{margin-top:9px;display:flex;flex-wrap:wrap;gap:4px}.check-dot{padding:3px 5px;border-radius:999px;border:1px solid var(--line-soft);color:var(--ink-soft);background:#fff;font:8px var(--font-mono)}.check-dot.check-pass{border-color:#2489c959;color:var(--forest-dark);background:var(--surface-blue)}.check-dot.check-review,.check-dot.check-unknown{border-color:#c47a1859;color:#8c560d;background:#fff8eb}.check-dot.check-fail{border-color:#bd4b5c59;color:var(--clay);background:#fff1f3}.candidate-empty{padding:14px 10px;color:var(--ink-soft);font-size:9px;border:1px dashed var(--line-soft);border-radius:var(--radius-sm)}.check-summary{padding:4px 14px 8px}.check-summary-row{padding:12px 0;display:grid;grid-template-columns:1fr auto;gap:10px;border-bottom:1px solid var(--line-soft)}.check-summary-row strong{font-size:10px}.check-summary-row p{margin:4px 0 0;color:var(--ink-soft);font-size:9px;line-height:1.4}.check-summary-row>span{align-self:center;color:var(--ink-soft);font:8px var(--font-mono)}.check-summary-row>span.has-pass{color:var(--forest-dark)}.check-summary-row>span.has-fail{color:var(--clay)}.honesty-note{margin:auto 14px 14px;padding:12px;border-radius:var(--radius-sm);background:var(--surface-blue)}.honesty-note strong{font-size:10px}.honesty-note p{margin:5px 0 0;color:var(--ink-soft);font-size:9px;line-height:1.45}.detail-card{padding:22px 20px;border-bottom:1px solid var(--line)}.detail-kicker{display:flex;justify-content:space-between;font:10px var(--font-mono);text-transform:uppercase}.detail-kicker strong{font-size:16px}.detail-card h2{margin:14px 0 2px;overflow-wrap:anywhere;font:650 23px/1.08 var(--font-display)}.dataset-id{color:var(--forest);font:10px var(--font-mono);overflow-wrap:anywhere}.detail-description{max-height:110px;margin:17px 0;overflow-y:auto;color:var(--ink-soft);font-size:11px;line-height:1.55}.discovery-note{margin:-6px 0 16px;padding:10px 11px;border:1px solid rgba(36,137,201,.35);border-radius:var(--radius-sm);background:var(--surface-blue);color:var(--forest-dark);font-size:10px;line-height:1.45}.score-grid{display:grid;grid-template-columns:repeat(4,1fr);border:1px solid var(--line);border-radius:var(--radius-sm);overflow:hidden}.score-grid div{padding:8px 5px;border-right:1px solid var(--line);border-bottom:1px solid var(--line);text-align:center}.score-grid div:nth-child(4n){border-right:0}.score-grid span{display:block;color:var(--ink-soft);font:8px var(--font-mono);text-transform:uppercase}.score-grid strong{display:block;margin-top:3px;font:600 17px var(--font-display)}.detail-block{margin-top:18px}.detail-block .field-label{margin-bottom:8px}.check-grid{border-top:1px solid var(--line-soft)}.check-row{padding:7px 0;display:flex;justify-content:space-between;border-bottom:1px solid var(--line-soft);font-size:10px;text-transform:capitalize}.check-row strong{font:9px var(--font-mono);text-transform:uppercase}.check-pass strong{color:var(--forest)}.check-fail strong{color:var(--clay)}.check-review strong,.check-unknown strong{color:#9a5d0c}.evidence-list{margin:0;padding-left:17px;color:var(--ink-soft);font-size:10px;line-height:1.55}.tag-list{display:flex;flex-wrap:wrap;gap:5px}.tag-list span{padding:4px 6px;border:1px solid var(--line);border-radius:999px;background:var(--surface-blue);font:8px var(--font-mono)}.sample-test-list{display:grid;gap:6px}.sample-test{padding:8px 9px;display:grid;grid-template-columns:1fr auto;gap:4px 8px;border:1px solid var(--line-soft);border-radius:var(--radius-sm);background:#fff}.sample-test strong{font-size:10px}.sample-test span{font:8px var(--font-mono);text-transform:uppercase}.sample-test p{grid-column:1 / -1;margin:0;color:var(--ink-soft);font-size:9px;line-height:1.4}.sample-pass span{color:var(--forest-dark)}.sample-review span,.sample-unknown span{color:#9a5d0c}.sample-fail span{color:var(--clay)}.loader-snippet{margin:0;padding:10px;overflow-x:auto;border:1px solid var(--line-soft);border-radius:var(--radius-sm);background:#f7fbff;color:var(--ink);font:9px/1.5 var(--font-mono)}.verdict{margin-top:18px;padding:13px;border:1px solid var(--line);border-radius:var(--radius-sm);background:var(--surface-blue)}.verdict strong{font-size:11px;line-height:1.45}.verdict p{margin:5px 0 0;color:var(--ink-soft);font-size:10px;line-height:1.5}.results-section{padding:20px}.results-heading{display:flex;justify-content:space-between;align-items:flex-end}.results-heading>span{font:600 24px var(--font-display)}.empty-results{margin-top:20px;padding:18px;border:1px dashed var(--line);border-radius:var(--radius-md);color:var(--ink-soft)}.empty-results strong{color:var(--ink);font:600 16px var(--font-display)}.empty-results p{margin:7px 0 0;font-size:10px;line-height:1.5}.result-list{margin-top:18px;border-top:2px solid var(--line)}.result-row{width:100%;min-height:66px;padding:10px 8px;display:grid;grid-template-columns:28px minmax(0,1fr) 40px;gap:8px;align-items:center;border:0;border-bottom:1px solid var(--line-soft);border-radius:var(--radius-sm);background:transparent;color:var(--ink);cursor:pointer;text-align:left}.result-row:hover{background:var(--surface-blue)}.result-rank{color:var(--ink-soft);font:10px var(--font-mono)}.result-main{min-width:0}.result-main strong{display:block;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;font-size:10px}.result-main span{display:block;margin-top:5px;font:8px var(--font-mono);text-transform:uppercase}.status-recommended .result-main span{color:var(--forest)}.status-conditional .result-main span{color:#9a5d0c}.status-rejected .result-main span{color:var(--clay)}.result-score{font:600 22px var(--font-display);text-align:right}.mobile-tabs{display:none}@media(max-width:1180px){.workspace{grid-template-columns:280px minmax(480px,1fr) 300px}.topbar-meta>span:first-child{display:none}}@media(max-width:1400px){.research-grid{grid-template-columns:1fr}.check-panel{min-height:auto}.check-summary{display:grid;grid-template-columns:repeat(2,1fr);gap:0 16px}.honesty-note{margin-top:10px}}@media(max-width:900px){body{overflow:auto}.topbar{min-height:62px;padding:0 16px}.brand{font-size:19px}.topbar-meta{gap:0}.hero{padding:32px 20px;grid-template-columns:1fr;gap:28px}.hero h1{font-size:38px}.workspace{display:block}.mobile-tabs{position:sticky;top:0;z-index:10;display:grid;grid-template-columns:repeat(3,1fr);border-bottom:2px solid var(--line);background:var(--paper-light)}.mobile-tabs button{padding:13px;border:0;border-right:1px solid var(--line);background:transparent;font:10px var(--font-mono);text-transform:uppercase}.mobile-tabs button.active{background:var(--forest);color:#fff}.rail,.map-stage{display:none;max-height:none;border:0}.rail.mobile-active,.map-stage.mobile-active{display:block}.map-stage.mobile-active{display:flex;min-height:calc(100vh - 108px)}.candidate-board{min-height:480px}.candidate-lanes{grid-template-columns:1fr}.search-section h2{font-size:34px}.brief-grid,.agent-progress{grid-template-columns:repeat(2,1fr)}.progress-step{border-bottom:1px solid var(--line-soft)}.progress-step:nth-child(2n){border-right:0}.progress-step:last-child{border-bottom:0}}@media(max-width:520px){.topbar-meta{display:none}.search-section,.process-section,.detail-card,.results-section{padding-left:17px;padding-right:17px}.hero{padding:26px 17px}.hero h1{font-size:33px}.hero-principles div{grid-template-columns:100px 1fr}.search-section h2{font-size:31px}.workspace-heading{padding-left:17px;padding-right:17px}.mode-note{display:none}.brief-strip,.agent-progress,.research-grid{margin-left:12px;margin-right:12px}.brief-strip-heading>span:last-child{display:none}.brief-grid{grid-template-columns:1fr 1fr}.brief-item:last-child{grid-column:1 / -1}.agent-progress{grid-template-columns:1fr}.progress-step{border-right:0;border-bottom:1px solid var(--line-soft)}.progress-step:last-child{border-bottom:0}.research-grid{min-height:0}.workflow-preview{flex-direction:column}.workflow-preview i{width:1px;height:14px}.candidate-board{min-height:500px}.check-summary{grid-template-columns:1fr}.score-grid{grid-template-columns:repeat(2,1fr)}.score-grid div:nth-child(2n){border-right:0}}@media(prefers-reduced-motion:reduce){*,*:before,*:after{scroll-behavior:auto!important;animation:none!important;transition:none!important}}

frontend/dist/index.html

@@ -6,8 +6,8 @@
   <meta name="theme-color" content="#f4faff" />
   <meta name="description" content="Evidence-led agentic dataset discovery for Hugging Face." />
   <title>HF Agentic Search</title>
-  <script type="module" crossorigin src="/assets/index-DEvIQgb1.js"></script>
-  <link rel="stylesheet" crossorigin href="/assets/index-BpRcJTDM.css">
+  <script type="module" crossorigin src="/assets/index-D64Ptpvl.js"></script>
+  <link rel="stylesheet" crossorigin href="/assets/index-BsUEW0Da.css">
 </head>
 <body>
   <div id="root"></div>

frontend/src/components/CandidateBoard.jsx

@@ -3,14 +3,22 @@ import { useGame } from '../GameProvider.jsx';
 
 const LANES = [
   {
-    status: 'recommended',
-    title: 'Recommended',
+    key: 'best',
+    title: 'Best fits',
     description: 'Ready for a first experiment',
+    filter: (dataset) => dataset.status === 'recommended' && !dataset.badges?.includes('hidden_gem'),
   },
   {
-    status: 'conditional',
-    title: 'Review gaps',
+    key: 'hidden',
+    title: 'Hidden gems',
+    description: 'Lower adoption, strong evidence fit',
+    filter: (dataset) => dataset.badges?.includes('hidden_gem'),
+  },
+  {
+    key: 'review',
+    title: 'Needs review',
     description: 'Useful, but one or more checks need attention',
+    filter: (dataset) => dataset.status === 'conditional' && !dataset.badges?.includes('hidden_gem'),
   },
 ];
 
@@ -45,7 +53,7 @@ export default function CandidateBoard() {
   const rejectedCount = datasets.length - visibleDatasets.length;
   const counts = LANES.map((lane) => ({
     ...lane,
-    count: visibleDatasets.filter((dataset) => dataset.status === lane.status).length,
+    count: visibleDatasets.filter(lane.filter).length,
   }));
 
   if (!datasets.length) {
@@ -68,7 +76,7 @@ export default function CandidateBoard() {
     <div className="candidate-board">
       <div className="decision-summary">
         {counts.map((lane) => (
-          <div className={`decision-stat stat-${lane.status}`} key={lane.status}>
+          <div className={`decision-stat stat-${lane.key}`} key={lane.key}>
             <span>{lane.title}</span>
             <strong>{lane.count}</strong>
           </div>
@@ -82,9 +90,9 @@ export default function CandidateBoard() {
 
       <div className="candidate-lanes">
         {counts.map((lane) => {
-          const laneDatasets = visibleDatasets.filter((dataset) => dataset.status === lane.status);
+          const laneDatasets = visibleDatasets.filter(lane.filter);
           return (
-            <section className={`candidate-lane lane-${lane.status}`} key={lane.status}>
+            <section className={`candidate-lane lane-${lane.key}`} key={lane.key}>
               <div className="lane-heading">
                 <div>
                   <strong>{lane.title}</strong>
@@ -104,8 +112,16 @@ export default function CandidateBoard() {
                     <strong title={dataset.id}>{shortName(dataset.id)}</strong>
                     <span>{dataset.score}</span>
                   </div>
+                  {dataset.badges?.length ? (
+                    <div className="candidate-badges">
+                      {dataset.badges.map((badge) => (
+                        <span key={badge}>{badge.replaceAll('_', ' ')}</span>
+                      ))}
+                    </div>
+                  ) : null}
                   <p>{mainGap(dataset)}</p>
                   <small>{compactEvidence(dataset)}</small>
+                  {dataset.sample_test_summary ? <small>{dataset.sample_test_summary}</small> : null}
                   <div className="check-dots" aria-label="Evidence checks">
                     {CHECKS.map(([key, label]) => {
                       const value = dataset.checks?.[key] || 'unknown';

frontend/src/components/ResearchWorkspace.jsx

@@ -6,6 +6,7 @@ const STEPS = [
   { type: 'plan', label: 'Understand', detail: 'Turn the brief into explicit requirements' },
   { type: 'search', label: 'Search', detail: 'Explore several angles across the Hub' },
   { type: 'inspect', label: 'Inspect', detail: 'Read cards, schemas and sample rows' },
+  { type: 'reflect', label: 'Reflect', detail: 'Revise the search from evidence gaps' },
   { type: 'ranking', label: 'Rank', detail: 'Compare evidence and explain the result' },
 ];
 

frontend/src/hud/AgentLog.jsx

@@ -6,6 +6,7 @@ const LABELS = {
   plan: 'Plan',
   search: 'Search',
   inspect: 'Inspect',
+  reflect: 'Reflect',
   ranking: 'Rank',
   complete: 'Done',
   error: 'Error',
@@ -50,6 +51,14 @@ function EventMeta({ event }) {
       </div>
     );
   }
+  if (event.type === 'reflect' && event.next_queries?.length) {
+    return (
+      <div className="event-meta">
+        <span>{event.strategy}</span>
+        <span>{event.next_queries.length} next angles</span>
+      </div>
+    );
+  }
   return null;
 }
 

frontend/src/hud/DetailCard.jsx

@@ -10,6 +10,16 @@ function Check({ label, value }) {
   );
 }
 
+function SampleTest({ test }) {
+  return (
+    <div className={`sample-test sample-${test.status}`}>
+      <strong>{test.name}</strong>
+      <span>{test.status}</span>
+      <p>{test.detail}</p>
+    </div>
+  );
+}
+
 export default function DetailCard() {
   const { selected } = useGame();
   if (!selected) return null;
@@ -27,6 +37,7 @@ export default function DetailCard() {
       <p className="detail-description">
         {selected.description || 'No dataset-card description was available.'}
       </p>
+      {selected.discovery_note ? <p className="discovery-note">{selected.discovery_note}</p> : null}
 
       <div className="score-grid">
         {Object.entries(selected.score_breakdown || {}).map(([key, value]) => (
@@ -62,6 +73,22 @@ export default function DetailCard() {
         </div>
       ) : null}
 
+      {selected.sample_tests?.length ? (
+        <div className="detail-block">
+          <span className="field-label">Sample tests</span>
+          <div className="sample-test-list">
+            {selected.sample_tests.map((test) => <SampleTest key={test.name} test={test} />)}
+          </div>
+        </div>
+      ) : null}
+
+      {selected.loader_snippet ? (
+        <div className="detail-block">
+          <span className="field-label">Starter loader</span>
+          <pre className="loader-snippet">{selected.loader_snippet}</pre>
+        </div>
+      ) : null}
+
       <div className="verdict">
         <strong>{selected.recommendation}</strong>
         <p>{selected.weakness}</p>

frontend/src/styles.css

@@ -340,7 +340,7 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
 .agent-progress {
   margin: 12px 20px;
   display: grid;
-  grid-template-columns: repeat(4, 1fr);
+  grid-template-columns: repeat(5, 1fr);
   border: 1px solid var(--line);
   border-radius: var(--radius-md);
   overflow: hidden;
@@ -454,8 +454,9 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
   margin-top: 4px;
   font: 650 20px var(--font-display);
 }
-.stat-recommended strong { color: var(--forest-dark); }
-.stat-conditional strong { color: var(--amber); }
+.stat-best strong { color: var(--forest-dark); }
+.stat-hidden strong { color: var(--forest); }
+.stat-review strong { color: var(--amber); }
 .stat-filtered strong { color: var(--clay); }
 .decision-summary p {
   margin: 0;
@@ -496,8 +497,12 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
 .lane-heading strong { display: block; font-size: 11px; }
 .lane-heading span { display: block; margin-top: 3px; color: var(--ink-soft); font-size: 8px; line-height: 1.35; }
 .lane-heading b { font: 650 18px var(--font-display); }
-.lane-recommended .lane-heading b { color: var(--forest-dark); }
-.lane-conditional .lane-heading b { color: var(--amber); }
+.lane-best .lane-heading b { color: var(--forest-dark); }
+.lane-hidden .lane-heading b { color: var(--forest); }
+.lane-review .lane-heading b { color: var(--amber); }
+.lane-hidden {
+  background: linear-gradient(180deg, rgba(230, 244, 254, 0.9), rgba(255, 255, 255, 0.76));
+}
 .candidate-card {
   width: 100%;
   padding: 10px;
@@ -531,6 +536,21 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
   font-size: 10px;
 }
 .candidate-card-top span { font: 650 16px var(--font-display); }
+.candidate-badges {
+  margin-top: 6px;
+  display: flex;
+  flex-wrap: wrap;
+  gap: 4px;
+}
+.candidate-badges span {
+  padding: 3px 6px;
+  border: 1px solid rgba(36, 137, 201, 0.35);
+  border-radius: 999px;
+  background: var(--surface-blue);
+  color: var(--forest-dark);
+  font: 7px var(--font-mono);
+  text-transform: uppercase;
+}
 .candidate-card p {
   min-height: 34px;
   margin: 7px 0;
@@ -611,6 +631,16 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
   font-size: 11px;
   line-height: 1.55;
 }
+.discovery-note {
+  margin: -6px 0 16px;
+  padding: 10px 11px;
+  border: 1px solid rgba(36, 137, 201, 0.35);
+  border-radius: var(--radius-sm);
+  background: var(--surface-blue);
+  color: var(--forest-dark);
+  font-size: 10px;
+  line-height: 1.45;
+}
 .score-grid { display: grid; grid-template-columns: repeat(4, 1fr); border: 1px solid var(--line); border-radius: var(--radius-sm); overflow: hidden; }
 .score-grid div { padding: 8px 5px; border-right: 1px solid var(--line); border-bottom: 1px solid var(--line); text-align: center; }
 .score-grid div:nth-child(4n) { border-right: 0; }
@@ -634,6 +664,38 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
 .evidence-list { margin: 0; padding-left: 17px; color: var(--ink-soft); font-size: 10px; line-height: 1.55; }
 .tag-list { display: flex; flex-wrap: wrap; gap: 5px; }
 .tag-list span { padding: 4px 6px; border: 1px solid var(--line); border-radius: 999px; background: var(--surface-blue); font: 8px var(--font-mono); }
+.sample-test-list { display: grid; gap: 6px; }
+.sample-test {
+  padding: 8px 9px;
+  display: grid;
+  grid-template-columns: 1fr auto;
+  gap: 4px 8px;
+  border: 1px solid var(--line-soft);
+  border-radius: var(--radius-sm);
+  background: #fff;
+}
+.sample-test strong { font-size: 10px; }
+.sample-test span { font: 8px var(--font-mono); text-transform: uppercase; }
+.sample-test p {
+  grid-column: 1 / -1;
+  margin: 0;
+  color: var(--ink-soft);
+  font-size: 9px;
+  line-height: 1.4;
+}
+.sample-pass span { color: var(--forest-dark); }
+.sample-review span, .sample-unknown span { color: #9a5d0c; }
+.sample-fail span { color: var(--clay); }
+.loader-snippet {
+  margin: 0;
+  padding: 10px;
+  overflow-x: auto;
+  border: 1px solid var(--line-soft);
+  border-radius: var(--radius-sm);
+  background: #f7fbff;
+  color: var(--ink);
+  font: 9px/1.5 var(--font-mono);
+}
 .verdict { margin-top: 18px; padding: 13px; border: 1px solid var(--line); border-radius: var(--radius-sm); background: var(--surface-blue); }
 .verdict strong { font-size: 11px; line-height: 1.45; }
 .verdict p { margin: 5px 0 0; color: var(--ink-soft); font-size: 10px; line-height: 1.5; }
@@ -728,8 +790,9 @@ button:focus-visible, a:focus-visible, textarea:focus-visible {
   .search-section h2 { font-size: 34px; }
   .brief-grid { grid-template-columns: repeat(2, 1fr); }
   .agent-progress { grid-template-columns: repeat(2, 1fr); }
-  .progress-step:nth-child(2) { border-right: 0; }
-  .progress-step:nth-child(-n + 2) { border-bottom: 1px solid var(--line-soft); }
+  .progress-step { border-bottom: 1px solid var(--line-soft); }
+  .progress-step:nth-child(2n) { border-right: 0; }
+  .progress-step:last-child { border-bottom: 0; }
 }
 
 @media (max-width: 520px) {

tests/test_agent.py

@@ -190,6 +190,29 @@ class AgentTests(unittest.TestCase):
         self.assertEqual(tiny["status"], "conditional")
         self.assertGreater(training_sized["score"], tiny["score"])
 
+    def test_hidden_gem_sample_tests_and_loader_are_exposed(self):
+        profile, _ = parse_task(
+            "Climate science question-answer pairs for retrieval",
+            use_llm=False,
+        )
+        scored = score_dataset(
+            profile,
+            dataset(
+                id="small-lab/climate-qa-gem",
+                description="Climate science question-answer pairs.",
+                downloads=90,
+                likes=3,
+                features=["question", "answer"],
+                sample_rows=[{"question": "How does climate affect rainfall?", "answer": "It changes rainfall patterns."}],
+                configs=["default"],
+                splits=["train"],
+            ),
+        )
+        self.assertIn("hidden_gem", scored["badges"])
+        self.assertTrue(scored["sample_tests"])
+        self.assertIn("sample tests passed", scored["sample_test_summary"])
+        self.assertIn('load_dataset("small-lab/climate-qa-gem")', scored["loader_snippet"])
+
     def test_queries_are_short_and_hub_friendly(self):
         profile, _ = parse_task(
             "English customer support intent data with labels for a compact classifier",
@@ -364,9 +387,48 @@ class AgentTests(unittest.TestCase):
         self.assertEqual(event_types[1], "plan")
         self.assertIn("search", event_types)
         self.assertIn("inspect", event_types)
+        self.assertIn("reflect", event_types)
         self.assertEqual(event_types[-2:], ["ranking", "complete"])
         self.assertEqual(events[-1]["result"]["top_pick"], candidate["id"])
 
+    @patch("backend.agent._llm", return_value=None)
+    def test_reflection_second_pass_can_recover_a_schema_gem(self, _mock_llm):
+        weak = dataset(
+            id="broad/climate-reports",
+            description="Climate science reports.",
+            features=["text"],
+            sample_rows=[{"text": "Climate report"}],
+        )
+        gem = dataset(
+            id="niche/climate-qa-gem",
+            description="Climate science question-answer pairs.",
+            downloads=12,
+            likes=1,
+            features=["question", "answer"],
+            sample_rows=[{"question": "What is climate sensitivity?", "answer": "A warming estimate."}],
+        )
+
+        def fake_search(query, limit=35):
+            return [gem] if "qa dataset" in query else [weak]
+
+        def fake_inspect(_dataset_id, base):
+            return base
+
+        with (
+            patch("backend.agent.search_datasets", side_effect=fake_search),
+            patch("backend.agent.inspect_dataset", side_effect=fake_inspect),
+        ):
+            events = list(weave_events("Climate science question-answer pairs for retrieval"))
+
+        result = events[-1]["result"]
+        event_types = [event["type"] for event in events]
+        self.assertIn("reflect", event_types)
+        self.assertEqual(result["top_pick"], gem["id"])
+        self.assertTrue(any(
+            dataset["id"] == gem["id"] and "hidden_gem" in dataset.get("badges", [])
+            for dataset in result["datasets"]
+        ))
+
     @patch("backend.agent._llm", return_value=None)
     def test_search_angles_are_diversified_before_inspection(self, _mock_llm):
         popular = [