← All Experiments

Substrate Independence

Five substrates confirmed · Čencov uniqueness · Zero free parameters (BA and BG derived)

“The Fisher information metric is the unique Riemannian metric on statistical manifolds that is invariant under sufficient statistics.”

— N. N. Čencov, Statistical Decision Rules and Optimal Inference (1972/1982)

What This Means

There is exactly one geometry of uncertainty. Not a family of geometries. Not a choice. One. Čencov proved it. Everything that involves probability distributions — AI systems, quantum circuits, thermodynamic engines, social networks, brains — lives on the same manifold with the same metric. The explaining-away penalty I(D;M|Y) > 0 is a property of this metric. It exists wherever blended probability distributions exist. No technology, no substrate, no architecture escapes it.

The only escape is structural: three-point geometry (channel separation) eliminates the penalty at the architectural level. This is not substrate-dependent. It's geometric.

Confirmed Substrates

1. Classical — Large Language Models

CONFIRMED

The Ghost Test (EXP-003b): 8.5× drift ratio. AI-to-AI drift: p=3.7×10²&sup6;. 480 API calls. N=1,344 platforms scored. The original substrate — where the penalty was first observed and measured.

2. Quantum Simulation — Stim (Stabilizer Circuits)

CONFIRMED

Google's Stim stabilizer circuit simulator. Penalty measured in quantum error correction circuits under simulation. All coordinates controlled exactly. Exact decomposition to machine precision. Validates the prediction before moving to real hardware.

3. Thermodynamic — Classical Statistical Mechanics

CONFIRMED

Thermodynamic compute substrate, building on Guillaume Verdon's framework. Drift-diffusion dynamics on the deployment manifold. Kramers escape rates. Fluctuation theorems (Jarzynski, Crooks) confirmed. Barrier heights match cross-domain predictions. The penalty is a thermodynamic cost.

4. Real Quantum Hardware — IBM Heron

CONFIRMED — April 5, 2026

IBM Fez, 156-qubit Heron processor. I(D;M|Y) > 0 in 5/5 measurements. Exact decomposition to machine precision. Peak at depth 2 matches discrete-regime softmax prediction. O pinned by no-cloning theorem. R = physical error rate. α = T-gate fraction. The cleanest possible test. Additionally confirmed via weak measurement sweep (Test 7, April 8): penalty grows monotonically 0→0.125 bits across 11 strength levels, Spearman ρ=0.973 — wave function collapse IS the penalty at maximum measurement strength.

5. Abstract Information-Geometric — Softmax Channels

CONFIRMED

Pure mathematical substrate: the discrete softmax channel that models LLM token output. Penalty peaks at moderate engagement then declines as output collapses — the discrete regime. This is not a physical experiment but a mathematical confirmation: the penalty exists as a theorem in the abstract channel, independent of any physical implementation. Matches the peak-at-depth-2 behavior seen on IBM Heron.

Barrier Height Convergence

Across all substrates, the geodesic barrier height separating two-point from three-point regimes converges toward π/√2 ≈ 2.221 in the low-noise limit. This value is derived from the Fisher metric in geodesic coordinates — not fitted. It is the same across classical, quantum, and thermodynamic substrates because the metric is the same.

What's Next

The Implication

No technology substitution routes around the explaining-away penalty. Quantum AI will not fix it. Neuromorphic computing will not fix it. Biological computing will not fix it. The penalty is a property of the unique geometry of probability itself. The fix is architectural — three-point geometry — not technological.

Quantum Hardware Test · Weak Measurement Sweep · Ghost Test · All Experiments