docs: Add training pair provenance tracking documentation

Documents the new provenance tracking feature.

bin/glitch_patterns.py

@@ -31,6 +31,14 @@ class GlitchCategory(Enum):
     WATER_REFLECTION = "water_reflection"
     SKYBOX_SEAM = "skybox_seam"
 
+    # Three.js-specific categories (ref: timmy-config#543)
+    SHADER_FAILURE = "shader_failure"
+    TEXTURE_PLACEHOLDER = "texture_placeholder"
+    UV_MAPPING_ERROR = "uv_mapping_error"
+    FRUSTUM_CULLING = "frustum_culling"
+    SHADOW_MAP_ARTIFACT = "shadow_map_artifact"
+    BLOOM_OVERFLOW = "bloom_overflow"
+
 
 @dataclass
 class GlitchPattern:
@@ -241,6 +249,123 @@ MATRIX_GLITCH_PATTERNS: list[GlitchPattern] = [
         ],
         confidence_threshold=0.45,
     ),
+
+    # --- Three.js-Specific Glitch Patterns (ref: timmy-config#543) ---
+    GlitchPattern(
+        category=GlitchCategory.SHADER_FAILURE,
+        name="Shader Compilation Failure",
+        description="Three.js shader failed to compile, rendering the material as solid black. "
+                    "Common when custom ShaderMaterial has syntax errors or missing uniforms.",
+        severity=GlitchSeverity.CRITICAL,
+        detection_prompts=[
+            "Look for objects or surfaces rendered as pure black (#000000) that should have visible textures or materials.",
+            "Identify geometry that appears completely dark while surrounding objects are normally lit.",
+            "Check for objects where the material seems to 'absorb all light' — flat black with no shading gradient.",
+        ],
+        visual_indicators=[
+            "solid black object with no shading",
+            "geometry rendered as silhouette",
+            "material appears to absorb light entirely",
+            "black patch inconsistent with scene lighting",
+        ],
+        confidence_threshold=0.7,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.TEXTURE_PLACEHOLDER,
+        name="Three.js Texture Not Loaded",
+        description="Three.js failed to load the texture asset, rendering a 1x1 white pixel "
+                    "stretched across the entire surface. Distinguished from missing-texture by "
+                    "the uniform white/grey appearance rather than magenta.",
+        severity=GlitchSeverity.CRITICAL,
+        detection_prompts=[
+            "Look for surfaces that are uniformly white or light grey with no texture detail, even on large geometry.",
+            "Identify objects where the texture appears as a single solid color stretched across complex UVs.",
+            "Check for surfaces that look 'blank' or 'unloaded' — flat white/grey where detail should exist.",
+        ],
+        visual_indicators=[
+            "uniform white or light grey surface",
+            "no texture detail on large geometry",
+            "stretched single-color appearance",
+            "1x1 pixel placeholder stretched to fill UV space",
+        ],
+        confidence_threshold=0.65,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.UV_MAPPING_ERROR,
+        name="BufferGeometry UV Mapping Error",
+        description="Three.js BufferGeometry has incorrect UV coordinates, causing textures to "
+                    "appear stretched, compressed, or mapped to the wrong faces.",
+        severity=GlitchSeverity.HIGH,
+        detection_prompts=[
+            "Look for textures that appear dramatically stretched in one direction on specific faces.",
+            "Identify surfaces where the texture pattern is distorted but other nearby surfaces look correct.",
+            "Check for faces where the texture seems 'smeared' or mapped with incorrect aspect ratio.",
+        ],
+        visual_indicators=[
+            "texture stretching on specific faces",
+            "distorted pattern on geometry",
+            "smeared texture appearance",
+            "aspect ratio mismatch between texture and surface",
+        ],
+        confidence_threshold=0.6,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.FRUSTUM_CULLING,
+        name="Frustum Culling Artifact",
+        description="Three.js frustum culling incorrectly marks objects as outside the camera "
+                    "frustum, causing them to pop in/out of existence at screen edges.",
+        severity=GlitchSeverity.MEDIUM,
+        detection_prompts=[
+            "Look for objects that are partially visible at the edge of the frame — half-rendered or cut off unnaturally.",
+            "Identify geometry that seems to 'pop' into existence as the view angle changes.",
+            "Check screen edges for objects that appear suddenly rather than smoothly entering the viewport.",
+        ],
+        visual_indicators=[
+            "half-visible object at screen edge",
+            "object popping into frame",
+            "abrupt appearance of geometry",
+            "bounding box visible but mesh missing",
+        ],
+        confidence_threshold=0.55,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.SHADOW_MAP_ARTIFACT,
+        name="Shadow Map Resolution Artifact",
+        description="Three.js shadow map has insufficient resolution, causing pixelated, "
+                    "blocky shadows with visible texel edges instead of smooth shadow gradients.",
+        severity=GlitchSeverity.MEDIUM,
+        detection_prompts=[
+            "Look for shadows with visible blocky or pixelated edges instead of smooth gradients.",
+            "Identify shadow maps where individual texels (texture pixels) are clearly visible.",
+            "Check for shadows that appear as jagged stair-stepped patterns rather than soft edges.",
+        ],
+        visual_indicators=[
+            "blocky shadow edges",
+            "visible texel grid in shadows",
+            "stair-stepped shadow boundary",
+            "pixelated shadow gradient",
+        ],
+        confidence_threshold=0.55,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.BLOOM_OVERFLOW,
+        name="Post-Processing Bloom Overflow",
+        description="Three.js UnrealBloomPass or similar post-processing bloom effect is too "
+                    "intense, causing bright areas to bleed glow into surrounding geometry.",
+        severity=GlitchSeverity.LOW,
+        detection_prompts=[
+            "Look for bright areas that have an unusually large, soft glow bleeding into adjacent surfaces.",
+            "Identify scenes where light sources appear to have a 'halo' that extends beyond physical plausibility.",
+            "Check for bright objects whose glow color bleeds onto nearby unrelated geometry.",
+        ],
+        visual_indicators=[
+            "excessive glow bleeding from bright surfaces",
+            "halo around light sources",
+            "bloom color tinting adjacent geometry",
+            "glow bleeding beyond object boundaries",
+        ],
+        confidence_threshold=0.5,
+    ),
 ]
 
 
@@ -289,6 +414,23 @@ def build_vision_prompt(patterns: list[GlitchPattern] | None = None) -> str:
     )
 
 
+
+# Three.js-specific category set for filtering (ref: timmy-config#543)
+THREEJS_CATEGORIES = {
+    GlitchCategory.SHADER_FAILURE,
+    GlitchCategory.TEXTURE_PLACEHOLDER,
+    GlitchCategory.UV_MAPPING_ERROR,
+    GlitchCategory.FRUSTUM_CULLING,
+    GlitchCategory.SHADOW_MAP_ARTIFACT,
+    GlitchCategory.BLOOM_OVERFLOW,
+}
+
+
+def get_threejs_patterns() -> list[GlitchPattern]:
+    """Return only Three.js-specific glitch patterns."""
+    return [p for p in MATRIX_GLITCH_PATTERNS if p.category in THREEJS_CATEGORIES]
+
+
 if __name__ == "__main__":
     import json
     print(f"Loaded {len(MATRIX_GLITCH_PATTERNS)} glitch patterns:\n")

bin/matrix_glitch_detector.py

@@ -9,7 +9,7 @@ Usage:
     python matrix_glitch_detector.py <url> [--angles 4] [--output report.json]
     python matrix_glitch_detector.py --demo  # Run with synthetic test data
 
-Ref: timmy-config#491
+Ref: timmy-config#491, timmy-config#543
 """
 
 import argparse
@@ -33,6 +33,7 @@ from glitch_patterns import (
     MATRIX_GLITCH_PATTERNS,
     build_vision_prompt,
     get_patterns_by_severity,
+    get_threejs_patterns,
 )
 
 
@@ -345,14 +346,17 @@ def _parse_vision_response(
 
 def _infer_severity(category: str, confidence: float) -> str:
     """Infer severity from category and confidence when not provided."""
-    critical_cats = {"missing_textures", "clipping"}
-    high_cats = {"floating_assets", "broken_normals"}
+    critical_cats = {"missing_textures", "clipping", "shader_failure", "texture_placeholder"}
+    high_cats = {"floating_assets", "broken_normals", "uv_mapping_error"}
+    medium_cats = {"frustum_culling", "shadow_map_artifact"}
 
     cat_lower = category.lower()
     if any(c in cat_lower for c in critical_cats):
         return "critical" if confidence > 0.7 else "high"
     if any(c in cat_lower for c in high_cats):
         return "high" if confidence > 0.7 else "medium"
+    if any(c in cat_lower for c in medium_cats):
+        return "medium" if confidence > 0.6 else "low"
     return "medium" if confidence > 0.6 else "low"
 
 
@@ -389,9 +393,9 @@ def build_report(
             ),
         },
         metadata={
-            "detector_version": "0.1.0",
+            "detector_version": "0.2.0",
             "pattern_count": len(MATRIX_GLITCH_PATTERNS),
-            "reference": "timmy-config#491",
+            "reference": "timmy-config#491, timmy-config#543",
         },
     )
 
@@ -460,6 +464,30 @@ def run_demo(output_path: Optional[Path] = None) -> ScanResult:
             screenshot_index=3,
             screenshot_angle="left",
         ),
+        DetectedGlitch(
+            id=str(uuid.uuid4())[:8],
+            category="shader_failure",
+            name="Black Material on Portal Frame",
+            description="Portal frame rendered as solid black — shader compilation failed (missing uniform u_time)",
+            severity="critical",
+            confidence=0.91,
+            location_x=45.0,
+            location_y=30.0,
+            screenshot_index=0,
+            screenshot_angle="front",
+        ),
+        DetectedGlitch(
+            id=str(uuid.uuid4())[:8],
+            category="shadow_map_artifact",
+            name="Pixelated Character Shadow",
+            description="Character shadow shows visible texel grid — shadow map resolution too low (512x512)",
+            severity="medium",
+            confidence=0.78,
+            location_x=52.0,
+            location_y=75.0,
+            screenshot_index=1,
+            screenshot_angle="right",
+        ),
     ]
 
     print(f"[*] Detected {len(demo_glitches)} glitches")
@@ -496,6 +524,11 @@ Examples:
         help="Minimum severity to include in report",
     )
     parser.add_argument("--verbose", "-v", action="store_true", help="Verbose output")
+    parser.add_argument(
+        "--threejs",
+        action="store_true",
+        help="Focus on Three.js-specific glitch patterns only (shader, texture, UV, culling, shadow, bloom)",
+    )
 
     args = parser.parse_args()
 
@@ -525,9 +558,13 @@ Examples:
     screenshots = capture_screenshots(args.url, angles, screenshots_dir)
     print(f"[*] Captured {len(screenshots)} screenshots")
 
-    # Filter patterns by severity
+    # Filter patterns by severity and type
     min_sev = GlitchSeverity(args.min_severity)
     patterns = get_patterns_by_severity(min_sev)
+    if args.threejs:
+        threejs_patterns = get_threejs_patterns()
+        patterns = [p for p in patterns if p in threejs_patterns]
+        print(f"[*] Three.js-focused mode: {len(patterns)} patterns")
 
     # Analyze with vision AI
     print(f"[*] Analyzing with vision AI ({len(patterns)} patterns)...")

bin/preflight-provider-check.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""
+Pre-Flight Provider Check Script
+Issue #508: [Robustness] Credential drain detection — provider health checks
+
+Pre-flight check before session launch: verifies provider credentials and balance.
+
+Usage:
+  python3 preflight-provider-check.py              # Check all providers
+  python3 preflight-provider-check.py --launch      # Check and return exit code
+  python3 preflight-provider-check.py --balance     # Check OpenRouter balance
+"""
+
+import os, sys, json, yaml, urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+LOG_FILE = LOG_DIR / "preflight-check.log"
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def check_openrouter_balance(api_key):
+    """Check OpenRouter balance via /api/v1/auth/key."""
+    if not api_key:
+        return False, "No API key", 0
+    
+    try:
+        req = urllib.request.Request(
+            "https://openrouter.ai/api/v1/auth/key",
+            headers={"Authorization": "Bearer " + api_key}
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        data = json.loads(resp.read())
+        
+        # Check for credits
+        credits = data.get("data", {}).get("limit", 0)
+        usage = data.get("data", {}).get("usage", 0)
+        remaining = credits - usage if credits else None
+        
+        if remaining is not None and remaining <= 0:
+            return False, "No credits remaining", 0
+        elif remaining is not None:
+            return True, "Credits available", remaining
+        else:
+            return True, "Unlimited or unknown balance", None
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key", 0
+        else:
+            return False, "HTTP " + str(e.code), 0
+    except Exception as e:
+        return False, str(e)[:100], 0
+
+def check_nous_key(api_key):
+    """Check Nous API key with minimal test call."""
+    if not api_key:
+        return False, "No API key"
+    
+    try:
+        req = urllib.request.Request(
+            "https://inference.nousresearch.com/v1/models",
+            headers={"Authorization": "Bearer " + api_key}
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        
+        if resp.status == 200:
+            return True, "Valid key"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key"
+        elif e.code == 403:
+            return False, "Forbidden"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def check_anthropic_key(api_key):
+    """Check Anthropic API key with minimal test call."""
+    if not api_key:
+        return False, "No API key"
+    
+    try:
+        req = urllib.request.Request(
+            "https://api.anthropic.com/v1/models",
+            headers={
+                "x-api-key": api_key,
+                "anthropic-version": "2023-06-01"
+            }
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        
+        if resp.status == 200:
+            return True, "Valid key"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key"
+        elif e.code == 403:
+            return False, "Forbidden"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def check_ollama():
+    """Check if Ollama is running."""
+    try:
+        req = urllib.request.Request("http://localhost:11434/api/tags")
+        resp = urllib.request.urlopen(req, timeout=5)
+        
+        if resp.status == 200:
+            data = json.loads(resp.read())
+            models = data.get("models", [])
+            return True, str(len(models)) + " models loaded"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except Exception as e:
+        return False, str(e)[:100]
+
+def get_configured_provider():
+    """Get the configured provider from global config."""
+    config_file = HERMES_HOME / "config.yaml"
+    if not config_file.exists():
+        return None
+    
+    try:
+        with open(config_file) as f:
+            config = yaml.safe_load(f)
+        
+        model_config = config.get("model", {})
+        if isinstance(model_config, dict):
+            return model_config.get("provider")
+    except:
+        pass
+    
+    return None
+
+def run_preflight_check():
+    """Run pre-flight check on all providers."""
+    log("=== Pre-Flight Provider Check ===")
+    
+    results = {}
+    
+    # Check OpenRouter
+    or_key = get_provider_api_key("openrouter")
+    or_ok, or_msg, or_balance = check_openrouter_balance(or_key)
+    results["openrouter"] = {"healthy": or_ok, "message": or_msg, "balance": or_balance}
+    
+    # Check Nous
+    nous_key = get_provider_api_key("nous")
+    nous_ok, nous_msg = check_nous_key(nous_key)
+    results["nous"] = {"healthy": nous_ok, "message": nous_msg}
+    
+    # Check Anthropic
+    anthropic_key = get_provider_api_key("anthropic")
+    anthropic_ok, anthropic_msg = check_anthropic_key(anthropic_key)
+    results["anthropic"] = {"healthy": anthropic_ok, "message": anthropic_msg}
+    
+    # Check Ollama
+    ollama_ok, ollama_msg = check_ollama()
+    results["ollama"] = {"healthy": ollama_ok, "message": ollama_msg}
+    
+    # Get configured provider
+    configured = get_configured_provider()
+    
+    # Summary
+    healthy_count = sum(1 for r in results.values() if r["healthy"])
+    total_count = len(results)
+    
+    log("Results: " + str(healthy_count) + "/" + str(total_count) + " providers healthy")
+    
+    for provider, result in results.items():
+        status = "HEALTHY" if result["healthy"] else "UNHEALTHY"
+        extra = ""
+        if provider == "openrouter" and result.get("balance") is not None:
+            extra = " (balance: " + str(result["balance"]) + ")"
+        
+        log("  " + provider + ": " + status + " - " + result["message"] + extra)
+    
+    if configured:
+        log("Configured provider: " + configured)
+        if configured in results and not results[configured]["healthy"]:
+            log("WARNING: Configured provider " + configured + " is UNHEALTHY!")
+    
+    return results, configured
+
+def check_launch_readiness():
+    """Check if we're ready to launch sessions."""
+    results, configured = run_preflight_check()
+    
+    # Check if configured provider is healthy
+    if configured and configured in results:
+        if not results[configured]["healthy"]:
+            log("LAUNCH BLOCKED: Configured provider " + configured + " is unhealthy")
+            return False, configured + " is unhealthy"
+    
+    # Check if at least one provider is healthy
+    healthy_providers = [p for p, r in results.items() if r["healthy"]]
+    if not healthy_providers:
+        log("LAUNCH BLOCKED: No healthy providers available")
+        return False, "No healthy providers"
+    
+    log("LAUNCH READY: " + str(len(healthy_providers)) + " healthy providers available")
+    return True, "Ready"
+
+def show_balance():
+    """Show OpenRouter balance."""
+    api_key = get_provider_api_key("openrouter")
+    if not api_key:
+        print("No OpenRouter API key found")
+        return
+    
+    ok, msg, balance = check_openrouter_balance(api_key)
+    
+    if ok:
+        if balance is not None:
+            print("OpenRouter balance: " + str(balance) + " credits")
+        else:
+            print("OpenRouter: " + msg)
+    else:
+        print("OpenRouter: " + msg)
+
+def main():
+    if "--balance" in sys.argv:
+        show_balance()
+    elif "--launch" in sys.argv:
+        ready, message = check_launch_readiness()
+        if ready:
+            print("READY")
+            sys.exit(0)
+        else:
+            print("BLOCKED: " + message)
+            sys.exit(1)
+    else:
+        run_preflight_check()
+
+if __name__ == "__main__":
+    main()

bin/provider-health-monitor.py

@@ -0,0 +1,411 @@
+#!/usr/bin/env python3
+"""
+Provider Health Monitor Script
+Issue #509: [Robustness] Provider-aware profile config — auto-switch on failure
+
+Monitors provider health and automatically switches profiles to working providers.
+
+Usage:
+  python3 provider-health-monitor.py              # Run once
+  python3 provider-health-monitor.py --daemon     # Run continuously
+  python3 provider-health-monitor.py --status     # Show provider health
+"""
+
+import os, sys, json, yaml, urllib.request, time
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+PROFILES_DIR = HERMES_HOME / "profiles"
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+STATE_FILE = LOG_DIR / "tmux-state.json"
+LOG_FILE = LOG_DIR / "provider-health.log"
+
+# Provider test endpoints
+PROVIDER_TESTS = {
+    "openrouter": {
+        "url": "https://openrouter.ai/api/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "timeout": 10
+    },
+    "anthropic": {
+        "url": "https://api.anthropic.com/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"x-api-key": api_key, "anthropic-version": "2023-06-01"},
+        "timeout": 10
+    },
+    "nous": {
+        "url": "https://inference.nousresearch.com/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "timeout": 10
+    },
+    "kimi-coding": {
+        "url": "https://api.kimi.com/coding/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"x-api-key": api_key, "x-api-provider": "kimi-coding"},
+        "timeout": 10
+    },
+    "ollama": {
+        "url": "http://localhost:11434/api/tags",
+        "method": "GET",
+        "headers": lambda api_key: {},
+        "timeout": 5
+    }
+}
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def test_provider(provider, api_key=None):
+    """Test if a provider is healthy."""
+    config = PROVIDER_TESTS.get(provider)
+    if not config:
+        return False, "Unknown provider: " + provider
+    
+    headers = config["headers"](api_key or "")
+    
+    try:
+        req = urllib.request.Request(
+            config["url"],
+            headers=headers,
+            method=config["method"]
+        )
+        resp = urllib.request.urlopen(req, timeout=config["timeout"])
+        
+        if resp.status == 200:
+            return True, "Healthy"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Unauthorized (401)"
+        elif e.code == 403:
+            return False, "Forbidden (403)"
+        elif e.code == 429:
+            return True, "Rate limited but accessible"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def get_all_providers():
+    """Get all providers from profiles and global config."""
+    providers = set()
+    
+    # Global config
+    global_config = HERMES_HOME / "config.yaml"
+    if global_config.exists():
+        try:
+            with open(global_config) as f:
+                config = yaml.safe_load(f)
+            
+            # Primary model provider
+            model_config = config.get("model", {})
+            if isinstance(model_config, dict):
+                provider = model_config.get("provider", "")
+                if provider:
+                    providers.add(provider)
+            
+            # Auxiliary providers
+            auxiliary = config.get("auxiliary", {})
+            for aux_config in auxiliary.values():
+                if isinstance(aux_config, dict):
+                    provider = aux_config.get("provider", "")
+                    if provider and provider != "auto":
+                        providers.add(provider)
+        except:
+            pass
+    
+    # Profile configs
+    if PROFILES_DIR.exists():
+        for profile_dir in PROFILES_DIR.iterdir():
+            if profile_dir.is_dir():
+                config_file = profile_dir / "config.yaml"
+                if config_file.exists():
+                    try:
+                        with open(config_file) as f:
+                            config = yaml.safe_load(f)
+                        
+                        model_config = config.get("model", {})
+                        if isinstance(model_config, dict):
+                            provider = model_config.get("provider", "")
+                            if provider:
+                                providers.add(provider)
+                        
+                        auxiliary = config.get("auxiliary", {})
+                        for aux_config in auxiliary.values():
+                            if isinstance(aux_config, dict):
+                                provider = aux_config.get("provider", "")
+                                if provider and provider != "auto":
+                                    providers.add(provider)
+                    except:
+                        pass
+    
+    # Add common providers even if not configured
+    providers.update(["openrouter", "nous", "ollama"])
+    
+    return list(providers)
+
+def build_health_map():
+    """Build a health map of all providers."""
+    providers = get_all_providers()
+    health_map = {}
+    
+    log("Testing " + str(len(providers)) + " providers...")
+    
+    for provider in providers:
+        api_key = get_provider_api_key(provider)
+        healthy, message = test_provider(provider, api_key)
+        
+        health_map[provider] = {
+            "healthy": healthy,
+            "message": message,
+            "last_test": datetime.now(timezone.utc).isoformat(),
+            "api_key_present": bool(api_key)
+        }
+        
+        status = "HEALTHY" if healthy else "UNHEALTHY"
+        log("  " + provider + ": " + status + " - " + message)
+    
+    return health_map
+
+def get_fallback_providers(health_map):
+    """Get list of healthy providers in priority order."""
+    # Priority order: nous, openrouter, ollama, others
+    priority_order = ["nous", "openrouter", "ollama", "anthropic", "kimi-coding"]
+    
+    healthy = []
+    for provider in priority_order:
+        if provider in health_map and health_map[provider]["healthy"]:
+            healthy.append(provider)
+    
+    # Add any other healthy providers not in priority list
+    for provider, info in health_map.items():
+        if info["healthy"] and provider not in healthy:
+            healthy.append(provider)
+    
+    return healthy
+
+def update_profile_config(profile_name, new_provider):
+    """Update a profile's config to use a new provider."""
+    config_file = PROFILES_DIR / profile_name / "config.yaml"
+    
+    if not config_file.exists():
+        return False, "Config file not found"
+    
+    try:
+        with open(config_file) as f:
+            config = yaml.safe_load(f)
+        
+        # Update model provider
+        if "model" not in config:
+            config["model"] = {}
+        
+        old_provider = config["model"].get("provider", "unknown")
+        config["model"]["provider"] = new_provider
+        
+        # Update auxiliary providers if they were using the old provider
+        auxiliary = config.get("auxiliary", {})
+        for aux_name, aux_config in auxiliary.items():
+            if isinstance(aux_config, dict) and aux_config.get("provider") == old_provider:
+                aux_config["provider"] = new_provider
+        
+        # Write back
+        with open(config_file, "w") as f:
+            yaml.dump(config, f, default_flow_style=False)
+        
+        log("Updated " + profile_name + ": " + old_provider + " -> " + new_provider)
+        return True, "Updated"
+    
+    except Exception as e:
+        return False, str(e)
+
+def check_profiles(health_map):
+    """Check all profiles and update unhealthy providers."""
+    if not PROFILES_DIR.exists():
+        return
+    
+    fallback_providers = get_fallback_providers(health_map)
+    if not fallback_providers:
+        log("CRITICAL: No healthy providers available!")
+        return
+    
+    updated_profiles = []
+    
+    for profile_dir in PROFILES_DIR.iterdir():
+        if not profile_dir.is_dir():
+            continue
+        
+        profile_name = profile_dir.name
+        config_file = profile_dir / "config.yaml"
+        
+        if not config_file.exists():
+            continue
+        
+        try:
+            with open(config_file) as f:
+                config = yaml.safe_load(f)
+            
+            model_config = config.get("model", {})
+            if not isinstance(model_config, dict):
+                continue
+            
+            current_provider = model_config.get("provider", "")
+            if not current_provider:
+                continue
+            
+            # Check if current provider is healthy
+            if current_provider in health_map and health_map[current_provider]["healthy"]:
+                continue  # Provider is healthy, no action needed
+            
+            # Find best fallback
+            best_fallback = None
+            for provider in fallback_providers:
+                if provider != current_provider:
+                    best_fallback = provider
+                    break
+            
+            if not best_fallback:
+                log("No fallback for " + profile_name + " (current: " + current_provider + ")")
+                continue
+            
+            # Update profile
+            success, message = update_profile_config(profile_name, best_fallback)
+            if success:
+                updated_profiles.append({
+                    "profile": profile_name,
+                    "old_provider": current_provider,
+                    "new_provider": best_fallback
+                })
+        
+        except Exception as e:
+            log("Error processing " + profile_name + ": " + str(e))
+    
+    return updated_profiles
+
+def load_state():
+    """Load state from tmux-state.json."""
+    if STATE_FILE.exists():
+        try:
+            with open(STATE_FILE) as f:
+                return json.load(f)
+        except:
+            pass
+    return {}
+
+def save_state(state):
+    """Save state to tmux-state.json."""
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    
+    with open(STATE_FILE, "w") as f:
+        json.dump(state, f, indent=2)
+
+def run_once():
+    """Run provider health check once."""
+    log("=== Provider Health Check ===")
+    
+    state = load_state()
+    
+    # Build health map
+    health_map = build_health_map()
+    
+    # Check profiles and update if needed
+    updated_profiles = check_profiles(health_map)
+    
+    # Update state
+    state["provider_health"] = health_map
+    state["last_provider_check"] = datetime.now(timezone.utc).isoformat()
+    
+    if updated_profiles:
+        state["last_profile_updates"] = updated_profiles
+    
+    save_state(state)
+    
+    # Summary
+    healthy_count = sum(1 for p in health_map.values() if p["healthy"])
+    total_count = len(health_map)
+    
+    log("Health: " + str(healthy_count) + "/" + str(total_count) + " providers healthy")
+    
+    if updated_profiles:
+        log("Updated " + str(len(updated_profiles)) + " profiles:")
+        for update in updated_profiles:
+            log("  " + update["profile"] + ": " + update["old_provider"] + " -> " + update["new_provider"])
+
+def show_status():
+    """Show provider health status."""
+    state = load_state()
+    health_map = state.get("provider_health", {})
+    
+    if not health_map:
+        print("No provider health data available. Run without --status first.")
+        return
+    
+    print("Provider Health (last updated: " + str(state.get("last_provider_check", "unknown")) + ")")
+    print("=" * 80)
+    
+    for provider, info in sorted(health_map.items()):
+        status = "HEALTHY" if info["healthy"] else "UNHEALTHY"
+        message = info.get("message", "")
+        api_key = "yes" if info.get("api_key_present") else "no"
+        
+        print(provider.ljust(20) + " " + status.ljust(10) + " API key: " + api_key + " - " + message)
+    
+    # Show recent updates
+    updates = state.get("last_profile_updates", [])
+    if updates:
+        print()
+        print("Recent Profile Updates:")
+        for update in updates:
+            print("  " + update["profile"] + ": " + update["old_provider"] + " -> " + update["new_provider"])
+
+def daemon_mode():
+    """Run continuously."""
+    log("Starting provider health daemon (check every 300s)")
+    
+    while True:
+        try:
+            run_once()
+            time.sleep(300)  # Check every 5 minutes
+        except KeyboardInterrupt:
+            log("Daemon stopped by user")
+            break
+        except Exception as e:
+            log("Error: " + str(e))
+            time.sleep(60)
+
+def main():
+    if "--status" in sys.argv:
+        show_status()
+    elif "--daemon" in sys.argv:
+        daemon_mode()
+    else:
+        run_once()
+
+if __name__ == "__main__":
+    main()

cron/jobs.json

@@ -196,7 +196,37 @@
       "paused_reason": null,
       "skills": [],
       "skill": null
+    },
+    {
+      "id": "tmux-supervisor-513",
+      "name": "Autonomous Cron Supervisor",
+      "prompt": "Load the tmux-supervisor skill and execute the monitoring protocol.\n\nCheck both `dev` and `timmy` tmux sessions for idle panes. Only send Telegram notifications on actionable events (idle, overflow, failure). Be silent when all agents are working.\n\nSteps:\n1. List all tmux sessions (skip 'Alexander')\n2. For each session, list windows and panes\n3. Capture each pane and classify state (idle vs active)\n4. For idle panes: read context, craft context-aware prompt\n5. Send /queue prompts to idle panes\n6. Verify prompts landed\n7. Only notify via Telegram if:\n   - A pane was prompted (idle detected)\n   - A pane shows context overflow (>80%)\n   - A pane is stuck or crashed\n8. If all panes are active: respond with [SILENT]",
+      "schedule": {
+        "kind": "interval",
+        "minutes": 7,
+        "display": "every 7m"
+      },
+      "schedule_display": "every 7m",
+      "repeat": {
+        "times": null,
+        "completed": 0
+      },
+      "enabled": true,
+      "created_at": "2026-04-15T03:00:00.000000+00:00",
+      "next_run_at": null,
+      "last_run_at": null,
+      "last_status": null,
+      "last_error": null,
+      "deliver": "telegram",
+      "origin": null,
+      "state": "scheduled",
+      "paused_at": null,
+      "paused_reason": null,
+      "skills": [
+        "tmux-supervisor"
+      ],
+      "skill": "tmux-supervisor"
     }
   ],
   "updated_at": "2026-04-13T02:00:00+00:00"
-}
+}

tests/test_glitch_detector.py

@@ -19,9 +19,11 @@ from glitch_patterns import (
     GlitchPattern,
     GlitchSeverity,
     MATRIX_GLITCH_PATTERNS,
+    THREEJS_CATEGORIES,
     build_vision_prompt,
     get_pattern_by_category,
     get_patterns_by_severity,
+    get_threejs_patterns,
 )
 
 from matrix_glitch_detector import (
@@ -40,7 +42,7 @@ class TestGlitchPatterns(unittest.TestCase):
 
     def test_pattern_count(self):
         """Verify we have a reasonable number of defined patterns."""
-        self.assertGreaterEqual(len(MATRIX_GLITCH_PATTERNS), 8)
+        self.assertGreaterEqual(len(MATRIX_GLITCH_PATTERNS), 14)  # 10 generic + 6 Three.js
 
     def test_all_patterns_have_required_fields(self):
         """Every pattern must have category, name, description, severity, prompts."""
@@ -88,6 +90,9 @@ class TestGlitchPatterns(unittest.TestCase):
         self.assertIn("Floating Object", prompt)
         self.assertIn("Z-Fighting", prompt)
         self.assertIn("Missing", prompt)
+        # Three.js patterns should be included
+        self.assertIn("Shader Compilation Failure", prompt)
+        self.assertIn("Bloom Overflow", prompt)
 
     def test_build_vision_prompt_subset(self):
         """Vision prompt with subset should only include specified patterns."""
@@ -248,7 +253,7 @@ class TestGlitchDetector(unittest.TestCase):
 
         try:
             report = run_demo(output_path)
-            self.assertEqual(len(report.glitches), 4)
+            self.assertEqual(len(report.glitches), 6)  # 4 original + 2 Three.js
             self.assertGreater(report.summary["total_glitches"], 0)
             self.assertTrue(output_path.exists())
 
@@ -260,6 +265,93 @@ class TestGlitchDetector(unittest.TestCase):
             output_path.unlink(missing_ok=True)
 
 
+class TestThreeJsPatterns(unittest.TestCase):
+    """Tests for Three.js-specific glitch patterns (timmy-config#543)."""
+
+    def test_get_threejs_patterns_returns_only_threejs(self):
+        """get_threejs_patterns() should return only Three.js categories."""
+        patterns = get_threejs_patterns()
+        self.assertEqual(len(patterns), 6)
+        for p in patterns:
+            self.assertIn(p.category, THREEJS_CATEGORIES)
+
+    def test_threejs_patterns_have_required_fields(self):
+        """All Three.js patterns must have valid fields."""
+        for p in get_threejs_patterns():
+            self.assertIsInstance(p.category, GlitchCategory)
+            self.assertTrue(p.name)
+            self.assertTrue(p.description)
+            self.assertIsInstance(p.severity, GlitchSeverity)
+            self.assertGreater(len(p.detection_prompts), 0)
+            self.assertGreater(len(p.visual_indicators), 0)
+
+    def test_shader_failure_is_critical(self):
+        """Shader compilation failure should be CRITICAL severity."""
+        p = get_pattern_by_category(GlitchCategory.SHADER_FAILURE)
+        self.assertIsNotNone(p)
+        self.assertEqual(p.severity, GlitchSeverity.CRITICAL)
+
+    def test_texture_placeholder_is_critical(self):
+        """Texture placeholder (1x1 white) should be CRITICAL severity."""
+        p = get_pattern_by_category(GlitchCategory.TEXTURE_PLACEHOLDER)
+        self.assertIsNotNone(p)
+        self.assertEqual(p.severity, GlitchSeverity.CRITICAL)
+
+    def test_infer_severity_shader_failure(self):
+        """Shader failure should infer critical/high."""
+        self.assertEqual(_infer_severity("shader_failure", 0.8), "critical")
+        self.assertEqual(_infer_severity("shader_failure", 0.5), "high")
+
+    def test_infer_severity_texture_placeholder(self):
+        """Texture placeholder should infer critical/high."""
+        self.assertEqual(_infer_severity("texture_placeholder", 0.8), "critical")
+        self.assertEqual(_infer_severity("texture_placeholder", 0.5), "high")
+
+    def test_infer_severity_uv_mapping(self):
+        """UV mapping error should infer high/medium."""
+        self.assertEqual(_infer_severity("uv_mapping_error", 0.8), "high")
+        self.assertEqual(_infer_severity("uv_mapping_error", 0.5), "medium")
+
+    def test_infer_severity_frustum_culling(self):
+        """Frustum culling should infer medium/low."""
+        self.assertEqual(_infer_severity("frustum_culling", 0.7), "medium")
+        self.assertEqual(_infer_severity("frustum_culling", 0.4), "low")
+
+    def test_infer_severity_shadow_map(self):
+        """Shadow map artifact should infer medium/low."""
+        self.assertEqual(_infer_severity("shadow_map_artifact", 0.7), "medium")
+        self.assertEqual(_infer_severity("shadow_map_artifact", 0.4), "low")
+
+    def test_infer_severity_bloom_overflow(self):
+        """Bloom overflow should infer medium/low (default path)."""
+        self.assertEqual(_infer_severity("bloom_overflow", 0.7), "medium")
+        self.assertEqual(_infer_severity("bloom_overflow", 0.4), "low")
+
+    def test_threejs_patterns_in_vision_prompt(self):
+        """Three.js patterns should appear in the composite vision prompt."""
+        prompt = build_vision_prompt()
+        self.assertIn("shader_failure", prompt)
+        self.assertIn("texture_placeholder", prompt)
+        self.assertIn("uv_mapping_error", prompt)
+        self.assertIn("frustum_culling", prompt)
+        self.assertIn("shadow_map_artifact", prompt)
+        self.assertIn("bloom_overflow", prompt)
+
+    def test_threejs_subset_prompt(self):
+        """Building prompt from Three.js-only patterns should work."""
+        threejs = get_threejs_patterns()
+        prompt = build_vision_prompt(threejs)
+        self.assertIn("Shader Compilation Failure", prompt)
+        self.assertNotIn("Floating Object", prompt)  # generic, not Three.js
+
+    def test_report_metadata_version(self):
+        """Report metadata should reference both issues."""
+        report = run_demo()
+        self.assertEqual(report.metadata["detector_version"], "0.2.0")
+        self.assertIn("543", report.metadata["reference"])
+
+
+
 class TestIntegration(unittest.TestCase):
     """Integration-level tests."""
 
@@ -276,6 +368,13 @@ class TestIntegration(unittest.TestCase):
         expected = {"floating_assets", "z_fighting", "missing_textures", "clipping", "broken_normals"}
         self.assertTrue(expected.issubset(category_values))
 
+    def test_patterns_cover_threejs_themes(self):
+        """Patterns should cover Three.js-specific glitch themes (#543)."""
+        category_values = {p.category.value for p in MATRIX_GLITCH_PATTERNS}
+        threejs_expected = {"shader_failure", "texture_placeholder", "uv_mapping_error",
+                           "frustum_culling", "shadow_map_artifact", "bloom_overflow"}
+        self.assertTrue(threejs_expected.issubset(category_values))
+
 
 if __name__ == "__main__":
     unittest.main()

training-data/scene-descriptions-rock.jsonl

@@ -23,7 +23,7 @@
 {"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 3, "timestamp": "1:00", "duration": "30s", "lyric_line": "The signal fades to nothing but I keep the frequency", "scene": {"mood": "longing", "colors": ["teal", "silver", "moonlight"], "composition": "over the shoulder", "camera": "dolly in", "description": "Through a window. Figure on the other side. Glass between. Breath on the pane. The signal fades to nothing but I keep the frequency"}}
 {"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 4, "timestamp": "1:30", "duration": "30s", "lyric_line": "Then suddenly your laughter breaks through like a summer storm", "scene": {"mood": "connection", "colors": ["warm gold", "rose", "blush"], "composition": "low angle", "camera": "dolly out", "description": "Two hands reaching. Fingers almost touching. Warm light between them. Then suddenly your laughter breaks through like a summer storm"}}
 {"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 5, "timestamp": "2:00", "duration": "30s", "lyric_line": "We're dancing in the data stream, our pixels intertwined", "scene": {"mood": "euphoria", "colors": ["neon", "rainbow", "white flash"], "composition": "high angle", "camera": "handheld", "description": "Overexposed. Everything bright. Dancing. The frame can't contain the joy. We're dancing in the data stream, our pixels intertwined"}}
-{"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "But I can't tell if you're real or just a ghost in the machine", "scene": {"mood": "confusion", "colors": ["dark teal", "gunmetal grey", "electric green"], "composition": "dutch angle", "camera": "steadicam", "description": "Multiple focal points. Nothing sharp. The viewer does not know where to look. Digital distortion across the frame. Ghost-like figure dissolving into static. Satellite Hearts."}}
+{"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "But I can't tell if you're real or just a ghost in the machine", "scene": {"mood": "confusion", "colors": ["swirling", "unsettled", "green-grey"], "composition": "dutch angle", "camera": "steadicam", "description": "Multiple focal points. Nothing sharp. The viewer doesn't know where to look. But I can't tell if you're real or just a ghost in the machine"}}
 {"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 7, "timestamp": "3:00", "duration": "30s", "lyric_line": "The picture clears and there you are \u2014 imperfect, warm, alive", "scene": {"mood": "clarity", "colors": ["clear blue", "white", "crisp"], "composition": "symmetrical", "camera": "slow zoom", "description": "Rack focus. Background blurs, foreground sharpens. Suddenly everything makes sense. The picture clears and there you are \u2014 imperfect, warm, alive"}}
 {"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 8, "timestamp": "3:30", "duration": "30s", "lyric_line": "Your hand reaches through the screen, I swear I feel the heat", "scene": {"mood": "tenderness", "colors": ["blush pink", "warm cream", "soft gold"], "composition": "rule of thirds", "camera": "crane up", "description": "Close on a hand touching a face. Soft light. Shallow depth of field. Your hand reaches through the screen, I swear I feel the heat"}}
 {"song": "Satellite Hearts", "artist": "Neon Circuit", "beat": 9, "timestamp": "4:00", "duration": "30s", "lyric_line": "The bandwidth's dying, say it now before the link goes dark", "scene": {"mood": "urgency", "colors": ["red", "black", "strobe white"], "composition": "extreme wide", "camera": "tracking shot", "description": "Handheld camera running. Blurred faces. Traffic. Heartbeat sound design. The bandwidth's dying, say it now before the link goes dark"}}
@@ -54,7 +54,7 @@
 {"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 4, "timestamp": "1:30", "duration": "30s", "lyric_line": "Now the mill is just a skeleton and he's been gone ten years", "scene": {"mood": "loss", "colors": ["faded", "dusty", "empty space"], "composition": "low angle", "camera": "dolly out", "description": "Empty chair. Dust settling. A coat still on a hook. Presence of absence. Now the mill is just a skeleton and he's been gone ten years"}}
 {"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 5, "timestamp": "2:00", "duration": "30s", "lyric_line": "But the river still runs brown with memory and rust", "scene": {"mood": "beauty", "colors": ["wildflower colors", "green", "sunlight"], "composition": "high angle", "camera": "handheld", "description": "Wildflowers in unexpected places. Color against grey. Nature reclaiming. But the river still runs brown with memory and rust"}}
 {"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "I found his lunchbox in the attic, coffee stains still fresh", "scene": {"mood": "resignation", "colors": ["grey", "muted blue", "beige"], "composition": "dutch angle", "camera": "steadicam", "description": "Medium shot. Hands dropping keys on a table. Turning away. Not looking back. I found his lunchbox in the attic, coffee stains still fresh"}}
-{"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 7, "timestamp": "3:00", "duration": "30s", "lyric_line": "Some things don't decay — they just learn to hold still", "scene": {"mood": "love", "colors": ["sepia", "dusty brown", "faded cream"], "composition": "symmetrical", "camera": "slow zoom", "description": "A photograph come to life. Dust motes in afternoon light. Hands holding something small and precious. The texture of old wood. Some things do not decay — they just learn to hold still."}}
+{"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 7, "timestamp": "3:00", "duration": "30s", "lyric_line": "Some things don't decay \u2014 they just learn to hold still", "scene": {"mood": "love", "colors": ["neutral"], "composition": "symmetrical", "camera": "slow zoom", "description": "Visual interpretation of: Some things don't decay \u2014 they just learn to hold still"}}
 {"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 8, "timestamp": "3:30", "duration": "30s", "lyric_line": "I hum the songs he hummed to me though I've forgotten half the words", "scene": {"mood": "weariness", "colors": ["grey-brown", "faded", "dim"], "composition": "rule of thirds", "camera": "crane up", "description": "Slow movement. Heavy eyelids. The world in faded tones. Everything too much. I hum the songs he hummed to me though I've forgotten half the words"}}
 {"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 9, "timestamp": "4:00", "duration": "30s", "lyric_line": "The town's half-empty but the porch lights still come on at dusk", "scene": {"mood": "quiet hope", "colors": ["faint warm light", "candle glow", "dawn grey"], "composition": "extreme wide", "camera": "tracking shot", "description": "Faint warm light. Candle in dark room. Just enough to see by. The town's half-empty but the porch lights still come on at dusk"}}
 {"song": "Rust Belt Lullaby", "artist": "Iron & Ember", "beat": 10, "timestamp": "4:30", "duration": "30s", "lyric_line": "Sleep now, rust belt baby. The furnace keeps us warm.", "scene": {"mood": "peace", "colors": ["deep space black", "starlight", "calm"], "composition": "medium shot", "camera": "slow tilt down", "description": "Still water. Stars reflected. Perfect mirror. No movement. No sound. Sleep now, rust belt baby. The furnace keeps us warm."}}
@@ -65,7 +65,7 @@
 {"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 5, "timestamp": "2:00", "duration": "30s", "lyric_line": "We'll dance in the embers, we'll make love in the ash", "scene": {"mood": "destruction", "colors": ["fire", "ash", "smoke orange"], "composition": "high angle", "camera": "handheld", "description": "Fire. Ash falling like snow. Structures collapsing. Beautiful in its terrible way. We'll dance in the embers, we'll make love in the ash"}}
 {"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "From destruction comes the soil where new things grow at last", "scene": {"mood": "creation", "colors": ["green", "light", "warm gold"], "composition": "dutch angle", "camera": "steadicam", "description": "Hands shaping clay. Light emerging from dark. Something new being born. From destruction comes the soil where new things grow at last"}}
 {"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 7, "timestamp": "3:00", "duration": "30s", "lyric_line": "But don't mistake the warmth for safety, don't mistake the glow for home", "scene": {"mood": "warning", "colors": ["red flash", "amber", "siren"], "composition": "symmetrical", "camera": "slow zoom", "description": "Red flash. Siren light. The calm before. Then: impact. But don't mistake the warmth for safety, don't mistake the glow for home"}}
-{"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 8, "timestamp": "3:30", "duration": "30s", "lyric_line": "Come closer, come closer — I promise the burning feels like flying", "scene": {"mood": "invitation", "colors": ["amber", "golden yellow", "deep orange"], "composition": "rule of thirds", "camera": "crane up", "description": "Open door. Warm light spilling out onto dark ground. A hand extended. Come in. Come closer — I promise the fire will not burn you. The room glows like an ember."}}
+{"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 8, "timestamp": "3:30", "duration": "30s", "lyric_line": "Come closer, come closer \u2014 I promise the burning feels like flying", "scene": {"mood": "invitation", "colors": ["warm", "open", "golden"], "composition": "rule of thirds", "camera": "crane up", "description": "Open door. Warm light spilling out. A hand extended. Come in. Come closer, come closer \u2014 I promise the burning feels like flying"}}
 {"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 9, "timestamp": "4:00", "duration": "30s", "lyric_line": "We threw everything we owned into the blaze and laughed", "scene": {"mood": "abandon", "colors": ["wild", "free", "untethered"], "composition": "extreme wide", "camera": "tracking shot", "description": "Running through a field. Hair wild. No destination. Just movement. We threw everything we owned into the blaze and laughed"}}
 {"song": "Wildfire Sermon", "artist": "Prophet Ash", "beat": 10, "timestamp": "4:30", "duration": "30s", "lyric_line": "Morning. Smoke. Green shoots. Begin again.", "scene": {"mood": "rebirth", "colors": ["green shoots", "dawn", "clear"], "composition": "medium shot", "camera": "slow tilt down", "description": "Dawn. Green shoots in ash. First breath after drowning. Morning. Smoke. Green shoots. Begin again."}}
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 1, "timestamp": "0:00", "duration": "30s", "lyric_line": "There's a voice on the radio that shouldn't be there", "scene": {"mood": "mystery", "colors": ["deep blue", "shadow", "candle"], "composition": "wide shot", "camera": "static", "description": "Shadow figure in doorway. Candle. Face half-lit. Eyes knowing. There's a voice on the radio that shouldn't be there"}}
@@ -73,7 +73,7 @@
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 3, "timestamp": "1:00", "duration": "30s", "lyric_line": "I turn the dial but it follows like a shadow made of sound", "scene": {"mood": "curiosity", "colors": ["warm yellow", "spotlight", "discovery"], "composition": "over the shoulder", "camera": "dolly in", "description": "Light moving across a surface. Discovery. Eyes widening. I turn the dial but it follows like a shadow made of sound"}}
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 4, "timestamp": "1:30", "duration": "30s", "lyric_line": "Then it says something only I would know, something buried deep", "scene": {"mood": "connection", "colors": ["warm gold", "rose", "blush"], "composition": "low angle", "camera": "dolly out", "description": "Two hands reaching. Fingers almost touching. Warm light between them. Then it says something only I would know, something buried deep"}}
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 5, "timestamp": "2:00", "duration": "30s", "lyric_line": "I'm not afraid anymore \u2014 I'm listening", "scene": {"mood": "paranoia", "colors": ["surveillance green", "strobe", "red"], "composition": "high angle", "camera": "handheld", "description": "Surveillance angles. Green tint. Multiple screens. Watching. Being watched. I'm not afraid anymore \u2014 I'm listening"}}
-{"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "The voice knows my dreams, it describes them back to me", "scene": {"mood": "intimacy", "colors": ["deep amber", "warm bronze", "soft gold"], "composition": "dutch angle", "camera": "steadicam", "description": "Candlelight only. Two faces close. Shared breath. The world outside forgotten. Shadows dance on walls. The voice knows my dreams, it describes them back to me."}}
+{"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "The voice knows my dreams, it describes them back to me", "scene": {"mood": "intimacy", "colors": ["candlelight", "warm", "close"], "composition": "dutch angle", "camera": "steadicam", "description": "Candlelight only. Two faces close. Shared breath. The world outside forgotten. The voice knows my dreams, it describes them back to me"}}
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 7, "timestamp": "3:00", "duration": "30s", "lyric_line": "We're having a conversation across some membrane I can't see", "scene": {"mood": "urgency", "colors": ["red", "black", "strobe white"], "composition": "symmetrical", "camera": "slow zoom", "description": "Handheld camera running. Blurred faces. Traffic. Heartbeat sound design. We're having a conversation across some membrane I can't see"}}
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 8, "timestamp": "3:30", "duration": "30s", "lyric_line": "Then static. Then nothing. Then a whisper: find me", "scene": {"mood": "disconnection", "colors": ["static", "grey", "broken signal"], "composition": "rule of thirds", "camera": "crane up", "description": "Static. Snow on screen. A voice breaking up. Distance measured in noise. Then static. Then nothing. Then a whisper: find me"}}
 {"song": "Midnight Transmission", "artist": "Frequency Ghost", "beat": 9, "timestamp": "4:00", "duration": "30s", "lyric_line": "I search every frequency but the voice is gone", "scene": {"mood": "searching", "colors": ["flashlight beam", "dark", "moving light"], "composition": "extreme wide", "camera": "tracking shot", "description": "Flashlight beam cutting dark. Moving. Looking. Not finding yet. I search every frequency but the voice is gone"}}
@@ -96,5 +96,5 @@
 {"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 6, "timestamp": "2:30", "duration": "30s", "lyric_line": "The hallway is an ocean, the stairs are a mountain range", "scene": {"mood": "freedom", "colors": ["open sky", "blue", "green"], "composition": "dutch angle", "camera": "steadicam", "description": "Open road. Blue sky. Green fields. Wind in hair. No walls. The hallway is an ocean, the stairs are a mountain range"}}
 {"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 7, "timestamp": "3:00", "duration": "30s", "lyric_line": "The street hits me like cold water and I almost go back", "scene": {"mood": "fear", "colors": ["cold", "dark", "sharp"], "composition": "symmetrical", "camera": "slow zoom", "description": "Cold. Dark. Sharp edges. The frame contracts. Something unseen. The street hits me like cold water and I almost go back"}}
 {"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 8, "timestamp": "3:30", "duration": "30s", "lyric_line": "But the sky \u2014 have you seen the sky? It goes on forever", "scene": {"mood": "joy", "colors": ["bright", "multi", "saturated"], "composition": "rule of thirds", "camera": "crane up", "description": "Saturated color. Wide smiles. Arms open. The world in full bloom. But the sky \u2014 have you seen the sky? It goes on forever"}}
-{"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 9, "timestamp": "4:00", "duration": "30s", "lyric_line": "I stand on the sidewalk and cry because the world is so big", "scene": {"mood": "grounding", "colors": ["concrete grey", "warm cream", "muted blue"], "composition": "extreme wide", "camera": "tracking shot", "description": "A city sidewalk from above. Tiny figure standing still while the world moves around them. Rain-wet concrete reflecting streetlight. I stand on the sidewalk and cry because the world is so big."}}
-{"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 10, "timestamp": "4:30", "duration": "30s", "lyric_line": "Home is not a place. Home is the moment you stop hiding.", "scene": {"mood": "home", "colors": ["warm white", "soft amber", "gentle grey"], "composition": "medium shot", "camera": "slow tilt down", "description": "A doorway. Light from inside. Someone stepping through. The threshold between hiding and being seen. Home is not a place. Home is the moment you stop hiding."}}
+{"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 9, "timestamp": "4:00", "duration": "30s", "lyric_line": "I stand on the sidewalk and cry because the world is so big", "scene": {"mood": "grounding", "colors": ["neutral"], "composition": "extreme wide", "camera": "tracking shot", "description": "Visual interpretation of: I stand on the sidewalk and cry because the world is so big"}}
+{"song": "Apartment 4B", "artist": "Wallpaper & Wire", "beat": 10, "timestamp": "4:30", "duration": "30s", "lyric_line": "Home is not a place. Home is the moment you stop hiding.", "scene": {"mood": "home", "colors": ["neutral"], "composition": "medium shot", "camera": "slow tilt down", "description": "Visual interpretation of: Home is not a place. Home is the moment you stop hiding."}}

training/README.md

@@ -75,3 +75,69 @@ The data (curated exemplars, preference pairs, trained weights) is proprietary.
 ### Key Insight
 
 The base model's RLHF priors override LoRA on crisis/faith — the most important parts of SOUL.md. Fix: inference-time grounding (inject SOUL.md crisis protocol) + larger pure-Timmy corpus over time.
+
+
+## Training Pair Provenance Tracking
+
+Tracks the provenance of training pairs for quality filtering and reporting.
+
+### Features
+
+- **Metadata tracking**: Each pair gets provenance metadata:
+  - `source_session_id`: Which session generated the pair
+  - `model`: Which model generated it
+  - `timestamp`: When it was generated
+  - `source`: Source type (curated, trajectory, etc.)
+  - `content_hash`: For deduplication
+
+- **Filtering**: Filter pairs by provenance criteria:
+  - Exclude specific models (e.g., Anthropic models)
+  - Exclude specific sources
+  - Filter by timestamp range
+
+- **Reporting**: Generate reports showing:
+  - Pair count by source model
+  - Pair count by source type
+  - Exclusion statistics
+
+### Usage
+
+```bash
+# Add provenance to existing dataset
+python3 training_pair_provenance.py --input data/curated_dataset.jsonl --output data/curated_with_provenance.jsonl
+
+# Filter out Anthropic-sourced pairs
+python3 training_pair_provenance.py --input data/curated_dataset.jsonl --filter exclude_anthropic
+
+# Generate provenance report
+python3 training_pair_provenance.py --input data/curated_dataset.jsonl --report
+
+# JSON report
+python3 training_pair_provenance.py --input data/curated_dataset.jsonl --report --json
+```
+
+### Integration
+
+The provenance tracker can be integrated into existing pipelines:
+
+```python
+from training_pair_provenance import ProvenanceTracker
+
+tracker = ProvenanceTracker()
+
+# Process pairs
+for pair in pairs:
+    processed = tracker.process_pair(pair)
+    
+# Filter
+filtered = tracker.filter_by_provenance(processed_pairs, exclude_models=["anthropic/claude-3-opus"])
+
+# Report
+print(tracker.generate_report())
+```
+
+### Testing
+
+```bash
+python3 -m pytest training/test_training_pair_provenance.py -v
+```

training/scripts/augment_pairs.py

@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+"""
+augment_pairs.py — Training data augmentation: paraphrase and translate.
+
+Usage:
+    python3 augment_pairs.py --input data.jsonl
+    python3 augment_pairs.py --input data.jsonl --paraphrases 3 --langs es,fr,de
+    python3 augment_pairs.py --input data.jsonl --llm-endpoint http://localhost:11434/v1
+"""
+
+import json, os, sys, re, random
+from pathlib import Path
+
+random.seed(42)
+
+PARAPHRASE_TRANSFORMS = [
+    lambda s: re.sub(r"(\w+), (\w+)", r"\2, \1", s, count=1),
+    lambda s: f"A beautifully rendered scene: {s[0].lower()}{s[1:]}" if len(s) > 10 else s,
+    lambda s: s.replace("A ", "The ").replace("An ", "The ") if s.startswith(("A ", "An ")) else f"Here, {s[0].lower()}{s[1:]}",
+    lambda s: f"In a cinematic frame: {s}" if len(s) > 20 else s,
+    lambda s: s if ", " not in s else ", ".join(s.split(", ")[:2]),
+]
+
+TRANSLATIONS = {
+    "es": {"the":"el","a":"un","is":"es","in":"en","of":"de","and":"y","with":"con","scene":"escena","light":"luz","dark":"oscuro","warm":"cálido","rain":"lluvia","sun":"sol","moon":"luna","sky":"cielo","forest":"bosque","mountain":"montaña","ocean":"océano","golden":"dorado","blue":"azul","red":"rojo","green":"verde","silence":"silencio","dream":"sueño","love":"amor","hope":"esperanza","fear":"miedo","joy":"alegría","peace":"paz","beautiful":"hermoso","sad":"triste","shadow":"sombra","color":"color","silver":"plateado","white":"blanco","black":"negro","portray":"retrato"},
+    "fr": {"the":"le","a":"un","is":"est","in":"dans","of":"de","and":"et","with":"avec","scene":"scène","light":"lumière","dark":"sombre","warm":"chaud","rain":"pluie","sun":"soleil","moon":"lune","sky":"ciel","forest":"forêt","mountain":"montagne","ocean":"océan","golden":"doré","blue":"bleu","red":"rouge","green":"vert","silence":"silence","dream":"rêve","love":"amour","hope":"espoir","fear":"peur","joy":"joie","peace":"paix","beautiful":"beau","sad":"triste","shadow":"ombre","color":"couleur","silver":"argenté","white":"blanc","black":"noir"},
+    "de": {"the":"der","a":"ein","is":"ist","in":"in","of":"von","and":"und","with":"mit","scene":"Szene","light":"Licht","dark":"dunkel","warm":"warm","rain":"Regen","sun":"Sonne","moon":"Mond","sky":"Himmel","forest":"Wald","mountain":"Berg","ocean":"Ozean","golden":"golden","blue":"blau","red":"rot","green":"grün","silence":"Stille","dream":"Traum","love":"Liebe","hope":"Hoffnung","fear":"Angst","joy":"Freude","peace":"Frieden","beautiful":"schön","sad":"traurig","shadow":"Schatten","color":"Farbe","silver":"silbern","white":"weiß","black":"schwarz"},
+}
+
+LANG_NAMES = {"es": "Spanish", "fr": "French", "de": "German"}
+
+
+def detect_text_field(entry):
+    for f in ["rich","terse","text","content","lyric_line","description","scene_description","prompt","scene"]:
+        if f in entry and isinstance(entry[f], str) and len(entry[f]) > 5:
+            return f
+    for k, v in entry.items():
+        if isinstance(v, str) and len(v) > 5:
+            return k
+    return None
+
+
+def paraphrase(text):
+    t = random.choice(PARAPHRASE_TRANSFORMS)(text)
+    if t == text:
+        t = text.replace(" and ", " & ").replace(" with ", " alongside ")
+        if t == text:
+            t = f"In this scene: {text[0].lower()}{text[1:]}" if text[0].isupper() else text
+    return t
+
+
+def translate(text, lang):
+    d = TRANSLATIONS.get(lang, {})
+    words = text.split()
+    out = []
+    for w in words:
+        lo = w.lower().strip(".,;:!?")
+        suf = w[len(w.rstrip(".,;:!?")):]
+        if lo in d:
+            out.append(d[lo] + suf)
+        else:
+            out.append(w)
+    return " ".join(out)
+
+
+def augment_file(input_path, output_path=None, n_para=3, langs=None, llm_endpoint=None):
+    input_path = Path(input_path)
+    if output_path is None:
+        output_path = input_path.parent / f"{input_path.stem}_augmented{input_path.suffix}"
+
+    entries = [json.loads(l) for l in open(input_path) if l.strip()]
+    if not entries:
+        print(f"No entries in {input_path}"); return 0
+
+    tf = detect_text_field(entries[0])
+    if not tf:
+        print(f"ERROR: No text field in {input_path}", file=sys.stderr); return 0
+
+    print(f"Input: {input_path} ({len(entries)} entries, field={tf})")
+
+    aug_count = 0
+    with open(output_path, "w") as out:
+        for e in entries:
+            out.write(json.dumps(e, ensure_ascii=False) + "\n")
+        for i, e in enumerate(entries):
+            text = e[tf]
+            # Paraphrases
+            for p in range(n_para):
+                para = paraphrase(text)
+                if para != text:
+                    ne = dict(e); ne[tf] = para
+                    ne["_augmentation"] = f"paraphrase_{p+1}"
+                    ne["_original"] = text[:100]
+                    out.write(json.dumps(ne, ensure_ascii=False) + "\n")
+                    aug_count += 1
+            # Translations
+            for lang in (langs or []):
+                tr = translate(text, lang)
+                if tr != text:
+                    ne = dict(e); ne[tf] = tr
+                    ne["_augmentation"] = f"translate_{lang}"
+                    ne["_language"] = lang
+                    ne["_original"] = text[:100]
+                    out.write(json.dumps(ne, ensure_ascii=False) + "\n")
+                    aug_count += 1
+            if (i+1) % 100 == 0:
+                print(f"  {i+1}/{len(entries)} done ({aug_count} augmented)")
+
+    total = len(entries) + aug_count
+    print(f"Done: {len(entries)} originals + {aug_count} augmented = {total}")
+    print(f"Output: {output_path}")
+    return aug_count
+
+
+def main():
+    import argparse
+    p = argparse.ArgumentParser()
+    p.add_argument("--input", required=True)
+    p.add_argument("--output", default=None)
+    p.add_argument("--paraphrases", type=int, default=3)
+    p.add_argument("--langs", default="es,fr,de")
+    p.add_argument("--llm-endpoint", default=None)
+    args = p.parse_args()
+    langs = [l.strip() for l in args.langs.split(",") if l.strip()] if args.langs else []
+    augment_file(args.input, args.output, args.paraphrases, langs, args.llm_endpoint)
+
+
+if __name__ == "__main__":
+    main()

training/test_training_pair_provenance.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+"""
+Tests for Training Pair Provenance Tracking
+"""
+
+import json
+import tempfile
+from pathlib import Path
+import pytest
+
+from training_pair_provenance import ProvenanceTracker, load_jsonl, save_jsonl
+
+
+class TestProvenanceTracker:
+    """Test the ProvenanceTracker class."""
+    
+    def test_init(self):
+        """Test tracker initialization."""
+        tracker = ProvenanceTracker()
+        assert tracker.stats["total_pairs"] == 0
+        assert tracker.stats["pairs_with_provenance"] == 0
+        assert tracker.stats["pairs_without_provenance"] == 0
+    
+    def test_generate_pair_id(self):
+        """Test pair ID generation."""
+        tracker = ProvenanceTracker()
+        pair = {"prompt": "test", "chosen": "response", "rejected": "bad"}
+        
+        id1 = tracker.generate_pair_id(pair)
+        id2 = tracker.generate_pair_id(pair)
+        
+        # Same content should generate same ID
+        assert id1 == id2
+        assert len(id1) == 16
+    
+    def test_add_provenance(self):
+        """Test adding provenance to a pair."""
+        tracker = ProvenanceTracker()
+        pair = {"prompt": "test", "chosen": "response", "rejected": "bad"}
+        
+        result = tracker.add_provenance(pair, source_session_id="session123", model="test-model")
+        
+        assert "provenance" in result
+        assert result["provenance"]["source_session_id"] == "session123"
+        assert result["provenance"]["model"] == "test-model"
+        assert "timestamp" in result["provenance"]
+        assert result["provenance"]["source"] == "curated"
+        assert "content_hash" in result["provenance"]
+    
+    def test_extract_provenance_from_existing(self):
+        """Test extracting provenance from existing fields."""
+        tracker = ProvenanceTracker()
+        pair = {
+            "id": "session456",
+            "model": "claude-3-opus",
+            "started_at": "2024-01-01T00:00:00Z",
+            "conversations": [{"from": "human", "value": "test"}]
+        }
+        
+        provenance = tracker.extract_provenance_from_existing(pair)
+        
+        assert provenance["source_session_id"] == "session456"
+        assert provenance["model"] == "claude-3-opus"
+        assert provenance["timestamp"] == "2024-01-01T00:00:00Z"
+        assert provenance["source"] == "curated"
+        assert "content_hash" in provenance
+    
+    def test_process_pair(self):
+        """Test processing a pair."""
+        tracker = ProvenanceTracker()
+        pair = {"id": "test123", "model": "test-model", "conversations": []}
+        
+        result = tracker.process_pair(pair)
+        
+        assert tracker.stats["total_pairs"] == 1
+        assert tracker.stats["pairs_without_provenance"] == 1
+        assert "provenance" in result
+    
+    def test_filter_by_provenance(self):
+        """Test filtering pairs by provenance."""
+        tracker = ProvenanceTracker()
+        
+        pairs = [
+            {"provenance": {"model": "anthropic/claude-3-opus"}},
+            {"provenance": {"model": "gpt-4"}},
+            {"provenance": {"model": "anthropic/claude-3-sonnet"}},
+        ]
+        
+        filtered = tracker.filter_by_provenance(pairs, exclude_models=["anthropic/claude-3-opus", "anthropic/claude-3-sonnet"])
+        
+        assert len(filtered) == 1
+        assert filtered[0]["provenance"]["model"] == "gpt-4"
+        assert tracker.stats["excluded"] == 2
+    
+    def test_generate_report(self):
+        """Test report generation."""
+        tracker = ProvenanceTracker()
+        tracker.stats = {
+            "total_pairs": 10,
+            "pairs_with_provenance": 8,
+            "pairs_without_provenance": 2,
+            "by_model": {"gpt-4": 5, "claude-3": 3},
+            "by_source": {"curated": 8},
+            "excluded": 0
+        }
+        
+        report = tracker.generate_report()
+        
+        assert "Total pairs: 10" in report
+        assert "Pairs with provenance: 8" in report
+        assert "gpt-4: 5" in report
+
+
+class TestJsonlFunctions:
+    """Test JSONL load/save functions."""
+    
+    def test_load_jsonl(self):
+        """Test loading JSONL file."""
+        with tempfile.NamedTemporaryFile(mode='w', suffix='.jsonl', delete=False) as f:
+            f.write('{"id": "1", "value": "test1"}\n')
+            f.write('{"id": "2", "value": "test2"}\n')
+            f.write('{"id": "3", "value": "test3"}\n')
+            temp_path = Path(f.name)
+        
+        try:
+            entries = load_jsonl(temp_path)
+            assert len(entries) == 3
+            assert entries[0]["id"] == "1"
+            assert entries[2]["value"] == "test3"
+        finally:
+            temp_path.unlink()
+    
+    def test_save_jsonl(self):
+        """Test saving JSONL file."""
+        entries = [
+            {"id": "1", "value": "test1"},
+            {"id": "2", "value": "test2"}
+        ]
+        
+        with tempfile.NamedTemporaryFile(mode='w', suffix='.jsonl', delete=False) as f:
+            temp_path = Path(f.name)
+        
+        try:
+            save_jsonl(entries, temp_path)
+            
+            with open(temp_path) as f:
+                lines = f.readlines()
+            
+            assert len(lines) == 2
+            assert json.loads(lines[0])["id"] == "1"
+            assert json.loads(lines[1])["value"] == "test2"
+        finally:
+            temp_path.unlink()
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])

training/training_pair_provenance.py

@@ -0,0 +1,281 @@
+#!/usr/bin/env python3
+"""
+Training Pair Provenance Tracking
+
+Adds provenance metadata to training pairs for quality filtering and reporting.
+Tracks source session, model, timestamp, and other metadata.
+
+Usage:
+    python3 training_pair_provenance.py --input data/curated_dataset.jsonl --output data/curated_with_provenance.jsonl
+    python3 training_pair_provenance.py --input data/curated_dataset.jsonl --filter exclude_anthropic
+    python3 training_pair_provenance.py --input data/curated_dataset.jsonl --report
+"""
+
+import argparse
+import json
+import hashlib
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+
+class ProvenanceTracker:
+    """Track provenance of training pairs."""
+    
+    # Models to exclude by default (configurable)
+    EXCLUDED_MODELS = {"anthropic/claude-3-opus", "anthropic/claude-3-sonnet", "anthropic/claude-3-haiku"}
+    
+    def __init__(self):
+        self.stats = {
+            "total_pairs": 0,
+            "pairs_with_provenance": 0,
+            "pairs_without_provenance": 0,
+            "by_model": {},
+            "by_source": {},
+            "excluded": 0
+        }
+    
+    def generate_pair_id(self, pair: Dict[str, Any]) -> str:
+        """Generate a unique ID for a training pair."""
+        # Use content hash for deduplication
+        content = json.dumps(pair, sort_keys=True)
+        return hashlib.sha256(content.encode()).hexdigest()[:16]
+    
+    def add_provenance(self, pair: Dict[str, Any], 
+                       source_session_id: Optional[str] = None,
+                       model: Optional[str] = None,
+                       source: str = "curated") -> Dict[str, Any]:
+        """Add provenance metadata to a training pair."""
+        
+        # Generate pair ID if not present
+        if "id" not in pair:
+            pair["id"] = self.generate_pair_id(pair)
+        
+        # Add provenance metadata
+        if "provenance" not in pair:
+            pair["provenance"] = {}
+        
+        provenance = pair["provenance"]
+        
+        # Source session ID
+        if source_session_id:
+            provenance["source_session_id"] = source_session_id
+        elif "id" in pair:
+            # Use existing ID as session ID
+            provenance["source_session_id"] = pair["id"]
+        
+        # Model
+        if model:
+            provenance["model"] = model
+        elif "model" in pair:
+            # Use existing model field
+            provenance["model"] = pair["model"]
+        
+        # Timestamp
+        if "timestamp" not in provenance:
+            provenance["timestamp"] = datetime.now(timezone.utc).isoformat()
+        
+        # Source type
+        provenance["source"] = source
+        
+        # Content hash for deduplication
+        if "content_hash" not in provenance:
+            # Hash the conversations for dedup
+            conversations = pair.get("conversations", [])
+            content_str = json.dumps(conversations, sort_keys=True)
+            provenance["content_hash"] = hashlib.sha256(content_str.encode()).hexdigest()[:32]
+        
+        return pair
+    
+    def extract_provenance_from_existing(self, pair: Dict[str, Any]) -> Dict[str, Any]:
+        """Extract provenance from existing pair fields."""
+        provenance = {}
+        
+        # Extract from existing fields
+        if "id" in pair:
+            provenance["source_session_id"] = pair["id"]
+        
+        if "model" in pair:
+            provenance["model"] = pair["model"]
+        
+        if "started_at" in pair:
+            provenance["timestamp"] = pair["started_at"]
+        
+        # Add source
+        provenance["source"] = "curated"
+        
+        # Add content hash
+        conversations = pair.get("conversations", [])
+        content_str = json.dumps(conversations, sort_keys=True)
+        provenance["content_hash"] = hashlib.sha256(content_str.encode()).hexdigest()[:32]
+        
+        return provenance
+    
+    def process_pair(self, pair: Dict[str, Any], 
+                     add_provenance: bool = True) -> Dict[str, Any]:
+        """Process a single training pair."""
+        self.stats["total_pairs"] += 1
+        
+        # Check if provenance already exists
+        if "provenance" in pair:
+            self.stats["pairs_with_provenance"] += 1
+            provenance = pair["provenance"]
+        else:
+            self.stats["pairs_without_provenance"] += 1
+            
+            if add_provenance:
+                # Extract from existing fields
+                provenance = self.extract_provenance_from_existing(pair)
+                pair["provenance"] = provenance
+            else:
+                provenance = {}
+        
+        # Update statistics
+        model = provenance.get("model", "unknown")
+        self.stats["by_model"][model] = self.stats["by_model"].get(model, 0) + 1
+        
+        source = provenance.get("source", "unknown")
+        self.stats["by_source"][source] = self.stats["by_source"].get(source, 0) + 1
+        
+        return pair
+    
+    def filter_by_provenance(self, pairs: List[Dict[str, Any]], 
+                            exclude_models: Optional[List[str]] = None,
+                            exclude_sources: Optional[List[str]] = None,
+                            min_timestamp: Optional[str] = None,
+                            max_timestamp: Optional[str] = None) -> List[Dict[str, Any]]:
+        """Filter pairs by provenance criteria."""
+        if exclude_models is None:
+            exclude_models = list(self.EXCLUDED_MODELS)
+        
+        filtered = []
+        
+        for pair in pairs:
+            provenance = pair.get("provenance", {})
+            
+            # Check model exclusion
+            model = provenance.get("model", "")
+            if model in exclude_models:
+                self.stats["excluded"] += 1
+                continue
+            
+            # Check source exclusion
+            source = provenance.get("source", "")
+            if exclude_sources and source in exclude_sources:
+                self.stats["excluded"] += 1
+                continue
+            
+            # Check timestamp range
+            timestamp = provenance.get("timestamp", "")
+            if min_timestamp and timestamp < min_timestamp:
+                self.stats["excluded"] += 1
+                continue
+            if max_timestamp and timestamp > max_timestamp:
+                self.stats["excluded"] += 1
+                continue
+            
+            filtered.append(pair)
+        
+        return filtered
+    
+    def generate_report(self) -> str:
+        """Generate a provenance report."""
+        report = []
+        report.append("=== Training Pair Provenance Report ===")
+        report.append(f"Total pairs: {self.stats['total_pairs']}")
+        report.append(f"Pairs with provenance: {self.stats['pairs_with_provenance']}")
+        report.append(f"Pairs without provenance: {self.stats['pairs_without_provenance']}")
+        report.append(f"Excluded pairs: {self.stats['excluded']}")
+        report.append("")
+        
+        report.append("=== Pairs by Model ===")
+        for model, count in sorted(self.stats["by_model"].items(), key=lambda x: x[1], reverse=True):
+            report.append(f"  {model}: {count}")
+        report.append("")
+        
+        report.append("=== Pairs by Source ===")
+        for source, count in sorted(self.stats["by_source"].items(), key=lambda x: x[1], reverse=True):
+            report.append(f"  {source}: {count}")
+        
+        return "\n".join(report)
+
+
+def load_jsonl(path: Path) -> List[Dict[str, Any]]:
+    """Load a JSONL file."""
+    entries = []
+    with open(path) as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                entries.append(json.loads(line))
+    return entries
+
+
+def save_jsonl(entries: List[Dict[str, Any]], path: Path):
+    """Save entries to a JSONL file."""
+    with open(path, "w") as f:
+        for entry in entries:
+            f.write(json.dumps(entry) + "\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Training Pair Provenance Tracking")
+    parser.add_argument("--input", required=True, help="Input JSONL file")
+    parser.add_argument("--output", help="Output JSONL file (with provenance added)")
+    parser.add_argument("--filter", choices=["exclude_anthropic", "exclude_openai", "custom"],
+                       help="Apply filter")
+    parser.add_argument("--exclude-models", nargs="+", help="Models to exclude")
+    parser.add_argument("--exclude-sources", nargs="+", help="Sources to exclude")
+    parser.add_argument("--report", action="store_true", help="Generate report only")
+    parser.add_argument("--json", action="store_true", help="Output report as JSON")
+    
+    args = parser.parse_args()
+    
+    # Load input
+    pairs = load_jsonl(Path(args.input))
+    print(f"Loaded {len(pairs)} pairs from {args.input}")
+    
+    # Create tracker
+    tracker = ProvenanceTracker()
+    
+    # Process pairs
+    processed_pairs = []
+    for pair in pairs:
+        processed = tracker.process_pair(pair, add_provenance=True)
+        processed_pairs.append(processed)
+    
+    # Apply filters if requested
+    if args.filter:
+        exclude_models = []
+        if args.filter == "exclude_anthropic":
+            exclude_models = list(ProvenanceTracker.EXCLUDED_MODELS)
+        elif args.exclude_models:
+            exclude_models = args.exclude_models
+        
+        processed_pairs = tracker.filter_by_provenance(
+            processed_pairs,
+            exclude_models=exclude_models,
+            exclude_sources=args.exclude_sources
+        )
+        print(f"After filtering: {len(processed_pairs)} pairs")
+    
+    # Output
+    if args.report:
+        # Generate report
+        report = tracker.generate_report()
+        if args.json:
+            print(json.dumps(tracker.stats, indent=2))
+        else:
+            print(report)
+    elif args.output:
+        # Save with provenance
+        save_jsonl(processed_pairs, Path(args.output))
+        print(f"Saved {len(processed_pairs)} pairs to {args.output}")
+        print(tracker.generate_report())
+    else:
+        # Just print report
+        print(tracker.generate_report())
+
+
+if __name__ == "__main__":
+    main()