Core Algorithm Worker

• Create test file with comprehensive test cases
• Tests must fail before implementation
• Cover: correctness, edge cases, numerical accuracy, roundtrip

Phase 2: Core Algorithm Implementation

3. Implement TPQ (Temporal-Polar Quantization)

   • cartesian_to_polar(): Convert to radius + angle
   • recursive_polar_transform(): Log2(C) levels
   • adaptive_bit_allocation(): 60/40 radii/angle split
   • pack_quantized(): Bit-level tensor packing
   • Must pass VAL-TPQ-001 through VAL-TPQ-005

4. Implement SQJL (Spatial-QJL)

   • create_jl_projections(): Random projection matrices
   • spatial_jl_transform(): Apply projections preserving 2D
   • sign_quantize(): 1-bit sign encoding
   • unbiased_estimator(): Special estimator for attention
   • Must pass VAL-SQJL-001 through VAL-SQJL-004

5. Implement MAMP (Metric-Aware Mixed Precision)

   • Layer sensitivity configuration
   • Timestep-aware allocation
   • Metric preservation logic
   • Must pass VAL-MAMP-001 through VAL-MAMP-005

Phase 3: CPU Optimization

6. Implement CPU-optimized kernels
   • Use NumPy vectorization
   • Consider Numba for JIT compilation
   • AVX2/SIMD where applicable (via PyTorch)
   • Profile and optimize hot paths

Phase 4: Verification

7. Run tests to verify (GREEN)

   • All tests must pass
   • Roundtrip accuracy > 99%
   • No NaN/Inf in outputs
   • Performance benchmarks

8. Manual verification

   • Test on synthetic video tensors
   • Verify numerical stability
   • Check memory efficiency
   • Profile with cProfile

Phase 5: Integration Prep

9. Export clean interfaces
   • Document tensor shapes expected
   • Provide configuration schemas
   • Add type hints
   • Create examples

Example Handoff

{
  "salientSummary": "Implemented TPQ with polar transform achieving 99.2% roundtrip accuracy and 2.7x effective compression. SQJL sign-bit quantization verified with zero memory overhead. MAMP layer profiles configured for DiT attention types.",
  "whatWasImplemented": "Three-stage quantization pipeline: (1) TPQ with recursive polar compression, (2) SQJL with JL projection and 1-bit sign encoding, (3) MAMP with layer-specific precision allocation. CPU-optimized kernels using Numba JIT.",
  "whatWasLeftUndone": "Integration with Diffusers pipeline pending. CUDA kernels not implemented (no GPU).",
  "verification": {
    "commandsRun": [
      {"command": "python -m pytest tests/test_tpq.py -v", "exitCode": 0, "observation": "5 tests passed, roundtrip accuracy 99.2%"},
      {"command": "python -m pytest tests/test_sqjl.py -v", "exitCode": 0, "observation": "4 tests passed, distance preservation 0.98"},
      {"command": "python -m pytest tests/test_mamp.py -v", "exitCode": 0, "observation": "5 tests passed"},
      {"command": "python benchmarks/compression_ratio.py", "exitCode": 0, "observation": "2.7x compression achieved at 3.5-bit average"}
    ],
    "interactiveChecks": []
  },
  "tests": {
    "added": [
      {"file": "tests/test_tpq.py", "cases": [
        {"name": "test_polar_transform_accuracy", "verifies": "VAL-TPQ-001"},
        {"name": "test_recursive_compression", "verifies": "VAL-TPQ-002"},
        {"name": "test_bit_allocation", "verifies": "VAL-TPQ-003"},
        {"name": "test_temporal_redundancy", "verifies": "VAL-TPQ-004"},
        {"name": "test_roundtrip_accuracy", "verifies": "VAL-TPQ-005"}
      ]},
      {"file": "tests/test_sqjl.py", "cases": [
        {"name": "test_jl_distance_preservation", "verifies": "VAL-SQJL-001"},
        {"name": "test_sign_bit_overhead", "verifies": "VAL-SQJL-002"},
        {"name": "test_unbiased_estimator", "verifies": "VAL-SQJL-003"},
        {"name": "test_spatial_preservation", "verifies": "VAL-SQJL-004"}
      ]},
      {"file": "tests/test_mamp.py", "cases": [
        {"name": "test_layer_precision", "verifies": "VAL-MAMP-001"},
        {"name": "test_timestep_allocation", "verifies": "VAL-MAMP-002"}
      ]}
    ]
  },
  "discoveredIssues": [
    {
      "severity": "low",
      "description": "Recursive polar transform O(N log N) - could optimize with iterative approach",
      "suggestedFix": "Consider GPU/CUDA for production"
    }
  ]
}

When to Return to Orchestrator

• Numerical instability discovered (diverging errors, NaN)
• Algorithm correctness cannot be verified
• Performance is unacceptable (>10x slower than target)
• Tensor shapes don't match DiT requirements