refactor: anonymize LLM labels, merge fragmented tracks, and tune defaults

Replace GPT-4o/Claude/Gemini branding with generic Analyst 1/2/3 labels
in both frontend and backend. Add track merging to deduplicate fragmented
IDs for the same physical object. Filter non-mission-relevant objects from
timeline events. Bump frame store budget to 24 GiB, increase default
segmentation step to 12, auto-detect API base URL, and fix video playback
reliability (blob MIME type, canplay wait, timeout increase).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.gitignore

@@ -10,4 +10,5 @@ checkpoints/
 *.md
 .agent
 .claude/
-.superpowers/
+.superpowers/
+videos/

app.py

@@ -409,7 +409,7 @@ async def detect_async_endpoint(
     depth_estimator: str = Form("depth"),
     depth_scale: float = Form(25.0),
     enable_depth: bool = Form(False),
-    step: int = Form(7),
+    step: int = Form(12),
     mission: str = Form(None),
 ):
     _ttfs_t0 = time.perf_counter()

demo/js/api.js

@@ -1,8 +1,10 @@
 window.ISR = window.ISR || {};
 
 // ── API Base URL ──────────────────────────────────────────────────
-// Set to your HuggingFace Space URL for GPU access, or '' for same-origin
-const API_BASE = 'https://biaslab2025-isr.hf.space';
+// Auto-detect: same-origin in dev, HuggingFace Space in production demo
+const API_BASE = window.location.hostname === 'biaslab2025-isr.hf.space'
+  ? ''
+  : (localStorage.getItem('ISR_API_BASE') || '');
 
 // ── Async Data Functions (real backend API) ───────────────────────
 

demo/js/explain.js

@@ -44,7 +44,7 @@ async function loadExplainability(jobId, trackId) {
   if (_explainAbort) _explainAbort.abort();
   _explainAbort = new AbortController();
 
-  panel.innerHTML = '<div class="explain-loading"><div class="explain-spinner"></div><span>Analyzing with GPT-4o, Claude, and Gemini...</span></div>';
+  panel.innerHTML = '<div class="explain-loading"><div class="explain-spinner"></div><span>Running multi-analyst consensus pipeline...</span></div>';
 
   try {
     const resp = await fetch(`${ISR.API_BASE}/inspect/explain/${jobId}/${encodeURIComponent(trackId)}`, { signal: _explainAbort.signal });
@@ -97,9 +97,9 @@ function renderExplainGraph(data, container) {
     + '</div></div>';
 
   var llmHtml = '<div class="ex-llm-strip">'
-    + '<span class="ex-llm-tag active">GPT-4o</span>'
-    + '<span class="ex-llm-tag' + (data.consensus_bar && data.consensus_bar.validators_available >= 1 ? ' active' : '') + '">Claude</span>'
-    + '<span class="ex-llm-tag' + (data.consensus_bar && data.consensus_bar.validators_available >= 2 ? ' active' : '') + '">Gemini</span>'
+    + '<span class="ex-llm-tag active">Analyst 1</span>'
+    + '<span class="ex-llm-tag' + (data.consensus_bar && data.consensus_bar.validators_available >= 1 ? ' active' : '') + '">Analyst 2</span>'
+    + '<span class="ex-llm-tag' + (data.consensus_bar && data.consensus_bar.validators_available >= 2 ? ' active' : '') + '">Analyst 3</span>'
     + '</div>';
 
   // Trunk line
@@ -239,11 +239,11 @@ function render3DGraph(data, wrap) {
   cats.forEach(function(c) { allFeatures = allFeatures.concat(c.features || []); });
 
   // Aggregate validator counts for tooltips
-  var claudeAgreed = 0, claudeTotal = 0, geminiAgreed = 0, geminiTotal = 0;
+  var analyst2Agreed = 0, analyst2Total = 0, analyst3Agreed = 0, analyst3Total = 0;
   allFeatures.forEach(function(f) {
     var vals = f.validators || {};
-    if (vals.claude) { claudeTotal++; if (vals.claude.agree) claudeAgreed++; }
-    if (vals.gemini) { geminiTotal++; if (vals.gemini.agree) geminiAgreed++; }
+    if (vals.analyst_2) { analyst2Total++; if (vals.analyst_2.agree) analyst2Agreed++; }
+    if (vals.analyst_3) { analyst3Total++; if (vals.analyst_3.agree) analyst3Agreed++; }
   });
   var validatorsAvailable = (data.consensus_bar && data.consensus_bar.validators_available) || 0;
 
@@ -348,20 +348,21 @@ function render3DGraph(data, wrap) {
       html += '<strong style="color:' + node.color + ';font-size:12px">' + ISR.escHtml(node.label) + '</strong>';
       if (node.nodeType === 'feature' && node.nodeData.feat) {
         var f = node.nodeData.feat;
-        if (f.reasoning) html += '<div style="margin-top:5px;padding-top:5px;border-top:1px solid rgba(255,255,255,0.1);color:#cbd5e1"><span style="color:#10b981;font-weight:600">GPT-4o:</span> ' + ISR.escHtml(f.reasoning) + '</div>';
+        if (f.reasoning) html += '<div style="margin-top:5px;padding-top:5px;border-top:1px solid rgba(255,255,255,0.1);color:#cbd5e1"><span style="color:#10b981;font-weight:600">Analyst 1:</span> ' + ISR.escHtml(f.reasoning) + '</div>';
         var vals = f.validators || {};
         Object.keys(vals).forEach(function(name) {
           var v = vals[name];
           var ic = v.agree ? '#4ade80' : '#f87171';
           var icon = v.agree ? '\u2713' : '\u2717';
-          html += '<div style="margin-top:3px;color:' + ic + '"><span style="font-weight:600">' + icon + ' ' + ISR.escHtml(name) + ':</span> ' + ISR.escHtml(v.note || '') + '</div>';
+          var displayName = {analyst_2: 'Analyst 2', analyst_3: 'Analyst 3'}[name] || name;
+          html += '<div style="margin-top:3px;color:' + ic + '"><span style="font-weight:600">' + icon + ' ' + ISR.escHtml(displayName) + ':</span> ' + ISR.escHtml(v.note || '') + '</div>';
         });
       } else if (node.nodeType === 'category' && node.nodeData.features) {
         html += '<div style="margin-top:4px;color:#94a3b8">' + node.nodeData.features.length + ' features analyzed</div>';
-        if (validatorsAvailable > 0) html += '<div style="color:#94a3b8">Validated by Claude + Gemini</div>';
+        if (validatorsAvailable > 0) html += '<div style="color:#94a3b8">Validated by Analyst 2 + Analyst 3</div>';
       } else if (node.nodeType === 'verdict') {
         if (node.nodeData.reasoning_summary) html += '<div style="margin-top:4px;color:#cbd5e1">' + ISR.escHtml(node.nodeData.reasoning_summary) + '</div>';
-        if (validatorsAvailable > 0) html += '<div style="margin-top:3px;color:#94a3b8">Claude: ' + claudeAgreed + '/' + claudeTotal + ' \u00b7 Gemini: ' + geminiAgreed + '/' + geminiTotal + '</div>';
+        if (validatorsAvailable > 0) html += '<div style="margin-top:3px;color:#94a3b8">Analyst 2: ' + analyst2Agreed + '/' + analyst2Total + ' \u00b7 Analyst 3: ' + analyst3Agreed + '/' + analyst3Total + '</div>';
       } else if (node.nodeType === 'input') {
         if (data.reasoning_summary) html += '<div style="margin-top:4px;color:#cbd5e1">' + ISR.escHtml(data.reasoning_summary) + '</div>';
         html += '<div style="margin-top:3px;color:#94a3b8">Mission: ' + ISR.escHtml((ISR.STATE.mission || 'N/A').slice(0, 50)) + '</div>';
@@ -420,18 +421,27 @@ function render3DGraph(data, wrap) {
   graph.d3Force('charge').strength(-180);
   graph.d3Force('link').distance(75);
 
-  // Auto-fit camera: fire multiple times to catch layout as it stabilizes
-  // The graph simulation runs warmupTicks synchronously then cooldownTicks async,
-  // so we need progressive fits to ensure the final settled layout is fully visible.
-  [300, 800, 1500, 2500].forEach(function(ms) {
+  // Set initial camera to look down at the top-down DAG from a centered position.
+  // The DAG grows downward (td mode), so offset Y slightly negative to center vertically.
+  var nLevels = 4; // object, categories, features, verdict
+  var dagSpanY = (nLevels - 1) * 100; // dagLevelDistance=100
+  var centerY = dagSpanY / 2;
+  graph.cameraPosition({ x: 0, y: -centerY, z: dagSpanY * 1.2 }, { x: 0, y: -centerY, z: 0 });
+
+  // Immediate fit after warmup ticks (run synchronously), then progressive fits
+  // as the async cooldown ticks settle the layout.
+  requestAnimationFrame(function() {
+    graph.zoomToFit(0, 40);
+  });
+  [400, 1000, 2000].forEach(function(ms) {
     setTimeout(function() {
-      graph.zoomToFit(600, 80);
+      graph.zoomToFit(400, 40);
     }, ms);
   });
 
-  // Also fit when simulation finishes (cooldown complete)
+  // Final fit when simulation finishes (cooldown complete)
   graph.onEngineStop(function() {
-    graph.zoomToFit(600, 80);
+    graph.zoomToFit(400, 40);
   });
 
   // Handle container resize

demo/js/inspect.js

@@ -156,8 +156,8 @@ function enterInspectState(trackId) {
   if (hasRealData && ISR.STATE.jobId) {
     const tid = realTrack ? realTrack.track_id : trackId;
     ISR.loadExplainability(ISR.STATE.jobId, tid);
-    // Auto-switch to EXPLAIN tab after a brief delay so user sees it loading
-    setTimeout(() => ISR.switchDrawerTab('explain'), 300);
+    // Auto-switch to REASONING tab after a brief delay so user sees it loading
+    setTimeout(() => ISR.switchDrawerTab('reasoning'), 300);
   }
 
   // Wire back button

demo/js/real-backend.js

@@ -277,7 +277,8 @@ async function enterRealAnalysis(jobId, status) {
       if (res.status === 410) { ISR.showToast('Job was cancelled.', 6000); resetToReady(); return; }
       if (res.status === 202) { await new Promise(r => setTimeout(r, 1000)); continue; }
       if (res.ok) {
-        const blob = await res.blob();
+        const rawBlob = await res.blob();
+        const blob = new Blob([rawBlob], { type: 'video/mp4' });
         videoBlobUrl = URL.createObjectURL(blob);
         ISR.STATE._videoBlobUrl = videoBlobUrl;
         break;
@@ -300,7 +301,7 @@ async function enterRealAnalysis(jobId, status) {
     processedVideo.classList.remove('hidden');
     processedVideo.style.display = 'block';
     await new Promise(resolve => {
-      const timeout = setTimeout(resolve, 5000); // safety timeout
+      const timeout = setTimeout(resolve, 30000); // safety timeout — large videos need time
       processedVideo.addEventListener('canplay', () => { clearTimeout(timeout); resolve(); }, { once: true });
       processedVideo.load();
     });
@@ -693,6 +694,52 @@ function _deduplicateTracksByBbox(tracks) {
   return kept;
 }
 
+/**
+ * Merge fragmented track IDs that represent the same physical object.
+ *
+ * When the tracker loses and re-acquires an object (common in aerial/SAR footage
+ * with low-confidence detections), multiple track IDs get created for the same
+ * physical entity. This groups tracks by label + satisfies verdict and keeps
+ * only the best representative per group, merging subsidiary track IDs so
+ * inspect/selection still works.
+ *
+ * Grouping key: (label, satisfies). Within each group, keep the track with the
+ * highest score as the representative. Attach merged track IDs so the SVG overlay
+ * and click handlers can still resolve any constituent ID to the representative.
+ */
+function _mergeFragmentedTracks(tracks) {
+  if (tracks.length <= 1) return tracks;
+
+  // Group by (label, satisfies) — only merge assessed tracks with same verdict
+  const groups = new Map(); // key -> [tracks]
+  const ungrouped = [];
+
+  for (const t of tracks) {
+    // Only merge tracks that have been assessed with a definite verdict
+    if (t.assessment_status === 'ASSESSED' && t.satisfies !== undefined && t.satisfies !== null) {
+      const key = `${(t.label || '').toLowerCase()}::${t.satisfies}`;
+      if (!groups.has(key)) groups.set(key, []);
+      groups.get(key).push(t);
+    } else {
+      ungrouped.push(t);
+    }
+  }
+
+  const merged = [];
+  for (const [, group] of groups) {
+    // Sort by score descending — best representative first
+    group.sort((a, b) => (b.score || 0) - (a.score || 0));
+    const best = { ...group[0] };
+    // Collect all subsidiary track IDs
+    if (group.length > 1) {
+      best._merged_track_ids = group.map(t => t.track_id);
+    }
+    merged.push(best);
+  }
+
+  return merged.concat(ungrouped);
+}
+
 function renderTrackListFromData(tracks) {
   const panel = document.getElementById('tracksPanel');
   panel.innerHTML = '';
@@ -707,7 +754,8 @@ function renderTrackListFromData(tracks) {
   const relevant = hasVerdicts
     ? enriched.filter(t => t.satisfies === true)
     : enriched;
-  const deduped = _deduplicateTracksByBbox(relevant);
+  const spatialDeduped = _deduplicateTracksByBbox(relevant);
+  const deduped = _mergeFragmentedTracks(spatialDeduped);
   const filtered = ISR.STATE.trackFilter ? deduped.filter(t => t.type === ISR.STATE.trackFilter) : deduped;
 
   if (filtered.length === 0) {
@@ -753,7 +801,8 @@ function renderTrackListFromData(tracks) {
 
     const speedStr = t.speed_kph ? `${t.speed_kph.toFixed(1)} kph` : '--';
     const dirStr = t.direction_clock ? `${t.direction_clock}h` : '';
-    const idStr = `ID:${t.track_id}`;
+    const mergedIds = t._merged_track_ids;
+    const idStr = mergedIds ? `IDs:${mergedIds.join(',')}` : `ID:${t.track_id}`;
 
     // Status badge
     const badge = ISR.getStatusBadgeHTML(t);
@@ -1146,17 +1195,24 @@ async function deriveTimelineEvents(jobId, summary, frameCounts) {
         const label = t.label_name || t.label || 'object';
         const time = ISR.formatTime(frameIdx / fps);
 
-        // First detection of this track ID (emit once)
+        // Skip non-mission-relevant tracks for event generation.
+        // Objects like "dog", "cat", "mouse" clutter the timeline when they
+        // have no bearing on the mission objective.
+        const isMissionRelevant = t.mission_relevant !== false;
+
+        // First detection of this track ID (emit once, only for mission-relevant)
         if (!seenIds.has(tid)) {
           seenIds.add(tid);
-          events.push({
-            frame: frameIdx, time, type: 'detection',
-            label: `New ${label}`, description: `Track ${tid} first detected.`,
-            priority: 'normal',
-          });
+          if (isMissionRelevant) {
+            events.push({
+              frame: frameIdx, time, type: 'detection',
+              label: `New ${label}`, description: `Track ${tid} first detected.`,
+              priority: 'normal',
+            });
+          }
         }
 
-        // Track reappeared after being lost
+        // Track reappeared after being lost (only mission-relevant tracks are in lostTracks)
         if (lostTracks.has(tid)) {
           lostTracks.delete(tid);
           events.push({
@@ -1177,8 +1233,8 @@ async function deriveTimelineEvents(jobId, summary, frameCounts) {
           });
         }
 
-        // Assessed as not relevant — emit ONCE per track
-        if (t.satisfies === false && !assessedIds.has(tid)) {
+        // Assessed as mission-relevant but does not satisfy — emit ONCE per track
+        if (t.satisfies === false && isMissionRelevant && !assessedIds.has(tid)) {
           assessedIds.add(tid);
           events.push({
             frame: frameIdx, time, type: 'system',

demo/js/ui.js

@@ -17,10 +17,26 @@ function togglePlayPause() {
     const processedVideo = document.getElementById('processedVideo');
     if (processedVideo) {
       if (ISR.STATE.isPlaying) {
-        // Check video is ready
-        if (processedVideo.readyState < 2) {
-          console.warn('[ISR] Video not ready (readyState:', processedVideo.readyState, ')');
-          ISR.showToast('Video still loading, please wait...', 3000);
+        // If video has a source but isn't ready yet, wait for it
+        if (processedVideo.readyState < 2 && processedVideo.src) {
+          ISR.showToast('Video loading...', 3000);
+          processedVideo.addEventListener('canplay', function onReady() {
+            processedVideo.removeEventListener('canplay', onReady);
+            if (!ISR.STATE.isPlaying) return; // user may have toggled off while waiting
+            processedVideo.playbackRate = ISR.playbackSpeed;
+            if (processedVideo.currentTime >= processedVideo.duration - 0.1) {
+              processedVideo.currentTime = 0;
+            }
+            processedVideo.play().catch(err => {
+              console.error('[ISR] Video play failed:', err);
+              ISR.showToast('Video playback failed: ' + err.message, 5000);
+              ISR.STATE.isPlaying = false;
+              btn.innerHTML = '&#9654;';
+            });
+          });
+          return;
+        } else if (processedVideo.readyState < 2) {
+          console.warn('[ISR] Video not ready and no source set');
           ISR.STATE.isPlaying = false;
           btn.innerHTML = '&#9654;';
           return;

models/isr/explainer.py

@@ -250,9 +250,9 @@ class ISRExplainer:
                 model="gemini-2.0-flash",
                 contents=[
                     types.Content(role="user", parts=[
-                        types.Part.from_text(_VALIDATOR_SYSTEM_PROMPT + "\n\n" + user_text),
+                        types.Part.from_text(text=_VALIDATOR_SYSTEM_PROMPT + "\n\n" + user_text),
                         types.Part.from_bytes(data=crop_bytes, mime_type="image/jpeg"),
-                        types.Part.from_text("\n[Full frame context]:"),
+                        types.Part.from_text(text="\n[Full frame context]:"),
                         types.Part.from_bytes(data=frame_bytes, mime_type="image/jpeg"),
                     ]),
                 ],
@@ -309,14 +309,14 @@ class ISRExplainer:
                 if claude and "feature_validations" in claude:
                     cv = self._find_validation(claude["feature_validations"], feat_key, feat["name"])
                     if cv:
-                        validators["claude"] = cv
+                        validators["analyst_2"] = cv
                         if cv.get("agree"):
                             feat_agreed += 1
 
                 if gemini and "feature_validations" in gemini:
                     gv = self._find_validation(gemini["feature_validations"], feat_key, feat["name"])
                     if gv:
-                        validators["gemini"] = gv
+                        validators["analyst_3"] = gv
                         if gv.get("agree"):
                             feat_agreed += 1
 

utils/frame_store.py

@@ -35,7 +35,7 @@ class SharedFrameStore:
     the budget ceiling, giving callers a chance to fall back to JPEG path.
     """
 
-    MAX_BUDGET_BYTES = 12 * 1024**3  # 12 GiB ceiling
+    MAX_BUDGET_BYTES = 24 * 1024**3  # 24 GiB ceiling
 
     def __init__(self, video_path: str, max_frames: Optional[int] = None):
         cap = cv2.VideoCapture(video_path)