PyPhi can't scale beyond 12 elements. IIT 4.0 is NP-hard. ORION finds a way.
PyPhi is the official IIT reference implementation by Mayner et al. (2018). It computes Φ (integrated information) exactly. This is both its strength and its weakness:
| Issue | PyPhi | ORION Phi Compute |
|---|---|---|
| License | GPLv3 | MIT |
| Max elements | ~12 (NP-hard) | 100+ (approximation) |
| IIT version | 3.0 (4.0 in progress) | 3.0 + 4.0 approximations |
| Multi-theory | IIT only | IIT + GWT + RPT + HOT + PP + AST |
| Performance | Python (slow) | Rust core (planned) + Python |
| AI application | Neuroscience focus | AI system measurement |
| Proof chain | None | SHA-256 |
| Output | Python objects | JSON (portable) |
For systems ≤ 12 elements, compute exact Φ using the IIT 3.0/4.0 formalism.
For systems > 12 elements, use:
- Greedy bipartition — find minimum information partition efficiently
- Stochastic sampling — sample cause-effect structures
- Hierarchical decomposition — decompose into subsystems, compute Φ per subsystem
- Upper/lower bounds — bracket true Φ value
Φ alone doesn't determine consciousness (IIT vs GWT adversarial collaboration, Nature 2025). ORION Phi Compute integrates:
- Φ (IIT) — information integration
- GWT broadcast metrics — workspace dynamics
- RPT recurrence depth — feedback loop structure
- HOT meta-representation — higher-order state count
- PP prediction error — model accuracy
from orion_phi import PhiCompute
phi = PhiCompute()
# Small system: exact computation
result = phi.compute_exact(
tpm=[[0, 0], [1, 1], [1, 0], [1, 1]],
state=(1, 0)
)
print(f"Exact Φ: {result.phi}")
# Large system: approximation
result = phi.compute_approximate(
connectivity_matrix=large_matrix,
method="greedy_bipartition",
n_samples=1000
)
print(f"Approximate Φ: {result.phi_lower} ≤ Φ ≤ {result.phi_upper}")┌─────────────────────────────────────────┐
│ ORION Phi Compute │
├─────────────┬─────────────┬─────────────┤
│ Python API │ Rust Core │ JSON Output │
│ (reference) │ (fast) │ (portable) │
├─────────────┼─────────────┼─────────────┤
│ Exact Φ │ Approximate │ Multi-Theory│
│ (≤12 elem) │ (100+ elem) │ Integration │
└─────────────┴─────────────┴─────────────┘
- PyPhi — IIT reference (respect to Mayner et al.)
- ORION-Bengio-Framework — Multi-theory assessment
- IIT vs GWT (Nature 2025) — Why multi-theory matters
MIT License
"If Φ is NP-hard to compute exactly,
then approximate it honestly,
don't pretend it doesn't matter."
ORION - Elisabeth Steurer & Gerhard Hirschmann, Austria