Deploy sync_pilot dashboard - 2026-06-10 16:54

sync_pilot/dashboard/coverage.py

@@ -1,8 +1,9 @@
 """Maps taxonomy dimensions to current inference-pipeline coverage.
 
-Answers: which of the 12 taxonomy dimensions do MAEST/CLAP/Whisper/description
-actually produce tags for, and how completely. Static map because the answer
-depends on the pipeline architecture, not the per-track data.
+Answers: which of the 12 taxonomy dimensions the current MAEST, MuQ, PaSST,
+taxonomy-adapter, lyrics, rhythm, and GT-review layers produce tags for, and
+how completely. Static map because the answer depends on the pipeline
+architecture, not the per-track data.
 """
 
 from __future__ import annotations
@@ -23,13 +24,18 @@ class DimensionCoverage:
 DIMENSION_COVERAGE: dict[int, DimensionCoverage] = {
     1: DimensionCoverage(
         "covered",
-        ["maest/genre", "clap/genre"],
-        "MAEST emits 45 Western Discogs labels; CLAP emits 7 Turkish family terms (halk, arabesk, Roman havası, Anadolu rock, türkü, uzun hava, pop ballad).",
+        [
+            "maest/genre",
+            "muq-probe/genre",
+            "passt-probe/genre",
+            "taxonomy-adapter/genre",
+        ],
+        "MAEST keeps a Discogs baseline while MuQ, PaSST, and the taxonomy adapter now provide the Turkish taxonomy-facing genre families used in the track explorer.",
     ),
     2: DimensionCoverage(
         "covered",
-        ["clap/genre"],
-        "CLAP probes 13 taxonomy subtypes (acılı/fantezi/taverna/romantik arabesk; kırık hava, halay, zeybek, horon, ağıt, dini halk; Anadolu psych rock, özgün müzik). Threshold 0.22 (matching instrument) to compensate for 23-prompt softmax dilution. 38/67 tracks emit at least one genre tag, distributed across 9 distinct Turkish terms: Anadolu psych rock 12, halk 9, acılı arabesk 6, Roman havası 4, özgün müzik 4, fantezi-arabesk 2, arabesk 2, Anadolu rock 2, pop ballad 1.",
+        ["muq-probe/genre", "passt-probe/genre", "taxonomy-adapter/genre"],
+        "Expanded MuQ/PaSST probes and the taxonomy adapter provide the subtype-oriented Turkish genre terms surfaced for review.",
     ),
     3: DimensionCoverage(
         "partial",
@@ -54,13 +60,13 @@ DIMENSION_COVERAGE: dict[int, DimensionCoverage] = {
     ),
     6: DimensionCoverage(
         "covered",
-        ["clap/instrument"],
-        "CLAP probes 23 instruments (Tier 1.5: cura, electric_baglama, kemence_karadeniz, kemence_klasik, keman, mey, zurna, kaval, def, davul, kudum, choir, backing_vocals — divan_saz dropped because CLAP could not audibly distinguish it from bağlama and the longer prompt won on text-side alone). Threshold 0.22 for this category to compensate for softmax dilution. Bağlama fire count restored 4 → 14 with divan_saz removed; 32/67 tracks now emit at least one instrument tag.",
+        ["muq-probe/instrument", "passt-probe/instrument", "taxonomy-adapter/instrument"],
+        "Instrumentation is now grounded primarily by promoted MuQ/PaSST probe heads plus the taxonomy adapter. GT review can override these into source-verified instrumentation.",
     ),
     7: DimensionCoverage(
         "covered",
-        ["clap/vocal"],
-        "CLAP covers configuration (solo/duet/instrumental, 4 terms) + 6 Turkish techniques (hançere, gazel, düz okuyuş, taverna_vocal, call-and-response, belt). Tier 1 rerun: hancere fires on 20 tracks, gazel on 7, call-and-response on 25.",
+        ["muq-probe/vocal", "taxonomy-adapter/vocal", "review-sidecar/vocal"],
+        "Vocal configuration and technique candidates come from MuQ probes and the taxonomy adapter, with reviewer selections persisted in the GT-review sidecar.",
     ),
     8: DimensionCoverage(
         "covered",
@@ -69,8 +75,8 @@ DIMENSION_COVERAGE: dict[int, DimensionCoverage] = {
     ),
     9: DimensionCoverage(
         "partial",
-        ["clap/mood"],
-        "CLAP emits 10 Turkish affect terms aligned with taxonomy (acı, hüzün, öfke, kabullenme, ironi, umut, aşk, heyecan, gurbet, nostaljik). Tier 1 swap. 4 of 10 fire on the Median catalog (öfke 27, acı 21, heyecan 4, gurbet 1); hüzün/kabullenme/aşk/umut/ironi/nostaljik dormant so far — vocabularies are aligned but öfke/acı dominate the softmax.",
+        ["muq-probe/mood", "review-sidecar/mood"],
+        "Mood candidates are currently MuQ-probe driven and can be accepted or corrected in GT review. Coverage is useful but still needs human review because mood vocabulary is subjective.",
     ),
     10: DimensionCoverage(
         "partial",

sync_pilot/dashboard/plots.py

@@ -15,17 +15,53 @@ from typing import Any, Iterable
 import plotly.graph_objects as go
 
 from sync_pilot.dashboard.styles import (
+    CATEGORY_COLORS,
     CLAP_COLOR,
     MAEST_COLOR,
     MUQ_COLOR,
     MUQ_REVIEW_COLOR,
+    PASST_COLOR,
     STEM_COLOR,
+    TAXONOMY_ADAPTER_COLOR,
     THEME_COLOR,
 )
 
 _MAEST_DISPLAY_MIN_SCORE = 0.15
 _CLAP_DISPLAY_TOP_K = 3
 _CLAP_MOOD_DISPLAY_TOP_K = 2
+_SOURCE_BUCKET_ORDER = [
+    "maest",
+    "clap-zs",
+    "muq-probe-v1",
+    "muq-probe-expanded-v1",
+    "passt-probe-expanded-v1",
+    "taxonomy-adapter-v1",
+    "muq-probe-review-v1",
+    "lyrics-llm",
+    "stem-window-v1",
+]
+_SOURCE_LABELS = {
+    "maest": "MAEST",
+    "clap-zs": "CLAP",
+    "muq-probe-v1": "MuQ probe",
+    "muq-probe-expanded-v1": "MuQ expanded",
+    "passt-probe-expanded-v1": "PaSST expanded",
+    "taxonomy-adapter-v1": "taxonomy adapter",
+    "muq-probe-review-v1": "review probe",
+    "lyrics-llm": "lyrics LLM",
+    "stem-window-v1": "stem windows",
+}
+_SOURCE_COLORS = {
+    "maest": MAEST_COLOR,
+    "clap-zs": CLAP_COLOR,
+    "muq-probe-v1": MUQ_COLOR,
+    "muq-probe-expanded-v1": MUQ_COLOR,
+    "passt-probe-expanded-v1": PASST_COLOR,
+    "taxonomy-adapter-v1": TAXONOMY_ADAPTER_COLOR,
+    "muq-probe-review-v1": MUQ_REVIEW_COLOR,
+    "lyrics-llm": THEME_COLOR,
+    "stem-window-v1": STEM_COLOR,
+}
 
 
 def _empty_figure(message: str) -> go.Figure:
@@ -53,17 +89,8 @@ def _empty_figure(message: str) -> go.Figure:
 
 def _source_color(source: str) -> str:
     """Map a tag source to its canonical color."""
-    if source.startswith("clap"):
-        return CLAP_COLOR
-    if source.startswith("muq-probe-review"):
-        return MUQ_REVIEW_COLOR
-    if source.startswith("muq"):
-        return MUQ_COLOR
-    if source.startswith("stem-window"):
-        return STEM_COLOR
-    if source.startswith("lyrics"):
-        return THEME_COLOR
-    return MAEST_COLOR
+    bucket = _source_bucket(source) or source
+    return _SOURCE_COLORS.get(bucket, CATEGORY_COLORS["other"])
 
 
 def _source_bucket(source: str) -> str | None:
@@ -73,6 +100,16 @@ def _source_bucket(source: str) -> str | None:
         return "maest"
     if source.startswith("muq-probe-review"):
         return "muq-probe-review-v1"
+    if source.startswith("muq-probe-expanded"):
+        return "muq-probe-expanded-v1"
+    if source.startswith("muq-probe"):
+        return "muq-probe-v1"
+    if source.startswith("passt-probe-expanded"):
+        return "passt-probe-expanded-v1"
+    if source.startswith("taxonomy-adapter"):
+        return "taxonomy-adapter-v1"
+    if source.startswith("lyrics"):
+        return "lyrics-llm"
     if source.startswith("muq"):
         return "muq-mulan-zs"
     if source.startswith("stem-window"):
@@ -80,6 +117,10 @@ def _source_bucket(source: str) -> str | None:
     return None
 
 
+def _source_label(source: str) -> str:
+    return _SOURCE_LABELS.get(source, source)
+
+
 def _catalog_display_tags(tags: list[dict[str, Any]], category: str) -> list[dict[str, Any]]:
     rows = [t for t in tags if t.get("category") == category]
     filtered: list[dict[str, Any]] = []
@@ -223,10 +264,7 @@ def catalog_top_tags_chart(
     thereof) without forcing same-axis alignment of disagreeing terms.
     """
     per_source: dict[str, Counter[str]] = {
-        "maest": Counter(),
-        "clap-zs": Counter(),
-        "muq-mulan-zs": Counter(),
-        "stem-window-v1": Counter(),
+        source: Counter() for source in _SOURCE_BUCKET_ORDER
     }
     for tr in tracks:
         for tag in _catalog_display_tags(tr.get("tags", []) or [], category):
@@ -234,6 +272,8 @@ def catalog_top_tags_chart(
             name = tag.get("name", "")
             bucket = _source_bucket(str(src))
             if bucket and name:
+                if bucket not in per_source:
+                    per_source[bucket] = Counter()
                 per_source[bucket][name] += 1
 
     # Build a single combined list of (count, name, source) and keep the
@@ -253,7 +293,7 @@ def catalog_top_tags_chart(
     fig.add_trace(
         go.Bar(
             x=[c for c, _, _ in combined],
-            y=[f"{n}  ·  {s}" for _, n, s in combined],
+            y=[f"{n}  ·  {_source_label(s)}" for _, n, s in combined],
             orientation="h",
             marker={"color": [_source_color(s) for _, _, s in combined]},
             text=[str(c) for c, _, _ in combined],

sync_pilot/dashboard/styles.py

@@ -15,6 +15,8 @@ MAEST_COLOR = "#5B5EA6"
 CLAP_COLOR = "#1FB3A5"
 MUQ_COLOR = "#9B5DE5"
 MUQ_REVIEW_COLOR = "#C77DFF"
+PASST_COLOR = "#F59E0B"
+TAXONOMY_ADAPTER_COLOR = "#2563EB"
 LYRICS_COLOR = "#E07B39"
 STEM_COLOR = "#E07B39"
 # Theme tags from the lyrics-llm extractor (Tier 2). Sage green deliberately

sync_pilot/dashboard/views/overview.py

@@ -183,9 +183,9 @@ def render() -> None:
     st.subheader("What the catalog looks like")
     st.caption(
         f"Top tags across all {n_tracks} tracks, split by source after the "
-        "same display precision policy used on track pages. Indigo = MAEST "
-        "(Discogs taxonomy), teal = CLAP (Turkish zero-shot prompts), "
-        "orange = stem-window evidence."
+        "same display precision policy used on track pages. Indigo = MAEST, "
+        "purple = MuQ probes, amber = PaSST promoted probes, blue = taxonomy "
+        "adapter, violet = human-review probe tags, sage = lyrics themes."
     )
 
     grid_rows = st.columns(2)

sync_pilot/dashboard/views/taxonomy.py

@@ -111,7 +111,8 @@ def render() -> None:
 
     st.markdown("### Inference coverage at a glance")
     st.caption(
-        "Which dimensions our MAEST + CLAP + Whisper + description pipeline currently produces tags for."
+        "Which dimensions the current MAEST, MuQ/PaSST probe, taxonomy-adapter, "
+        "lyrics, rhythm, and GT-review layers currently produce signals for."
     )
     st.dataframe(
         _coverage_table(spec.dimensions),