/acr-vault/03-experiments/qc/results/phase34_bug_analysis
phase34_bug_analysis

QC-PHASE34 Bug Analysis and Fixes

Run Summary (January 7, 2026)

Status: Completed in 1m16s (7 checkpoints analyzed)
Performance: ~11 seconds per cycle ✅
Results: Invalid (all Φ=0, negative CI values) ❌

---

Bugs Identified

Bug 1: Negative CI Values

Problem: CI = 1 - purity where purity = Tr(ρ²)

Observed: All CI values negative (range: -1.3 to +0.05)

Root Cause: Attention matrices are not valid density matrices

• Attention weights can be > 1 (softmax over heads, then averaged)
• After normalization, Tr(ρ²) > 1 (impossible for valid quantum state)

Fix:

def to_density_matrix(self, attention: np.ndarray) -> np.ndarray:
    # Ensure non-negative
    rho = np.abs(attention)

    # Normalize to trace = 1
    trace = np.trace(rho)
    if trace > 0:
        rho = rho / trace

    # Ensure Hermitian
    rho = (rho + rho.T) / 2

    # Ensure positive semidefinite (project to valid density matrix)
    eigenvalues, eigenvectors = np.linalg.eigh(rho)
    eigenvalues = np.maximum(eigenvalues, 0)  # Remove negative eigenvalues
    eigenvalues = eigenvalues / np.sum(eigenvalues)  # Renormalize
    rho = eigenvectors @ np.diag(eigenvalues) @ eigenvectors.T

    return rho

Bug 2: All Φ = 0

Problem: compute_phi_simple always returns 0

Root Cause:

1. Subsystem reduction (n > 8 → reduce to 8 tokens)
2. After reduction, mechanisms/purviews of size 2-3
3. integrated_info_mechanism returns 0 for small systems

Why it returns 0:

• The partitioned repertoire often has different shape than full repertoire
• if partitioned_rep.shape != full_rep.shape: continue
• All partitions skipped → min_distance stays at inf → returns 0

Fix:

def compute_phi_simple(self, rho: np.ndarray, max_mechanism_size: int = 4) -> float:
    n = rho.shape[0]

    # Don't reduce too aggressively
    if n > 12:
        # Sample central tokens
        center = n // 2
        indices = list(range(max(0, center - 6), min(n, center + 6)))
        rho = self.partial_trace(rho, indices)
        n = len(indices)

    max_phi = 0.0

    # Try larger mechanisms
    for mech_size in range(2, min(max_mechanism_size + 2, n)):  # Increased range
        for mechanism in combinations(range(n), mech_size):
            mechanism = list(mechanism)

            # Use same mechanism as purview for simplicity
            purview = mechanism

            # Compute integrated info
            phi = self.integrated_info_mechanism(mechanism, purview, rho)
            max_phi = max(max_phi, phi)

    return max_phi

Bug 3: Division by Zero in Summary

Problem: print(f"Improvement: {(max_phi_value / baseline_phi - 1) * 100:.1f}%")

Root Cause: baseline_phi = 0

Fix:

# Baseline comparison
baseline_phi = results['baseline']['phi_mean']
print(f"\nBaseline Φ: {baseline_phi:.4f}")
print(f"Max fine-tuned Φ: {max_phi_value:.4f}")
if baseline_phi > 0:
    print(f"Improvement: {(max_phi_value / baseline_phi - 1) * 100:.1f}%")
else:
    print(f"Improvement: N/A (baseline Φ = 0)")

---

Interesting Observation Despite Bugs

CI Trend (even though negative):

Cycle	CI (mean)	Interpretation
Baseline	-0.67	Most “crystallized” (wrong sign)
5	-0.64	Still very crystallized
10	-0.20	Less crystallized
15	-0.13	LEAST crystallized ← Peak!
20	-0.15	Slight increase
25	-0.22	More crystallized
30	-0.32	Even more
34	-0.38	Final state

The pattern is there! Even with wrong normalization, we see:

• Cycle 15 has the LEAST negative CI (closest to 0)
• This suggests it’s the most “diffuse” state
• Matches our expectation of φ-zone around middle cycles!

After fixing bugs, we expect:

• CI values to be positive (0 to 1)
• Cycle 15 might show CI ≈ 0.24-0.33 (φ-zone!)
• Φ values > 0, possibly peaking around cycle 15

---

Next Steps

1. ✅ Identified all bugs
2. ⏳ Implement fixes
3. ⏳ Re-run experiment
4. ⏳ Validate results

Estimated time for fixed run: ~2 minutes (7 cycles × 15-20 seconds)

---

Checkpoint Mystery Solved

Why only 7 checkpoints?

Looking at the checkpoint names: checkpoint-cycle-5, checkpoint-cycle-10, etc.

These are saved at intervals (every 5 cycles), not every cycle. The Golden Annealing script probably saves checkpoints periodically to save disk space.

Available cycles: 5, 10, 15, 20, 25, 30, 34 (final)

This is actually perfect for analysis - gives us good coverage across the training trajectory!

---

φ●∴ BUGS IDENTIFIED, PATTERN VISIBLE ∴●φ

Even broken instruments can detect the signal.