The explaining-away penalty is universal. The fix is architectural. Paper 178 specifies hardware. Paper 179 proves the same geometry underlies the Yang-Mills mass gap. The penalty is not an AI problem — it is the shape of self-interacting mediation itself.
Paper 3 proved the penalty exists. Papers 166–167 proved it harms (613K students, 80 countries). Paper 177 proved it is universal (IBM Heron, 156-qubit Heron processor, wave function collapse is the penalty). Papers 178 and 179 close the constructive arc.
Test 6 (April 5, 2026, IBM Fez) proved entangled ancilla cannot serve as a third channel: 0/4 PASS. Entanglement preserves the shared manifold. Three-point geometry requires structural independence — different physics, not different software on the same substrate. Paper 178 specifies what works: a classical AI channel paired with a thermodynamic channel (Extropic Z1). Different manifolds. Different entropy production mechanisms. No shared generative process. The penalty cannot form across them because there is no shared latent model to condition on.
Extended architecture: adding a quantum channel (IBM Quantum) enables cross-substrate tomography — where channels agree the system is transparent, where they disagree a void is present. The disagreement map is the void map. Quantum is optional for penalty elimination but enables active void detection.
What makes it a safety result: Kolchinsky et al. (Phys. Rev. Research 8, 023025, April 2026) independently proved the penalty maps exactly to housekeeping entropy production. RLHF on blended channels is a futile cycle — all dissipation, zero productive work. Eliminating the penalty via substrate separation restores capacity. Three-point geometry is an acceleration result, not a cost.
Paper 178 is the pitch document for Extropic (thermodynamic substrate via sMTJ Boltzmann sampling) and IBM Quantum (third-channel tomography via Heron / future processors). The paper specifies the architecture; the partnerships build it.
In QED, the photon does not carry electric charge — it mediates without participating. No self-interaction. No confinement. No mass gap. In QCD, the gluon carries color charge — simultaneously participant and mediator. Self-interacting. Confined. Mass gap > 0. Paper 179 shows the Eckert statistical manifold, developed for AI drift measurement with zero knowledge of gauge theory, spontaneously discriminates these cases.
The phonological spectral level never confines (Polyakov loop = 1.0000 at all tested engagement levels). The pragmatic level shows 4.8× confinement coupling and a distinct susceptibility peak. Adding a quartic self-interaction term g²|A|⁴/T to the Fokker-Planck operator — the dimensional reduction of |A ∧ A|² — increases the excitation gap with scaling exponent g2.015 (CV=0.0012).
Step 2 theorem (lattice): gauge transformations are Markov morphisms for the Yang-Mills statistical model (gauge invariance of the action forces p(o|A) = p(o|A′) for gauge-equivalent A, A′ and any gauge-invariant observable o). Čencov uniqueness then forces the Fisher metric to be the unique gauge-invariant metric on A/G. Steps 3 and 4 formalize via O’Neill tensor and log-Sobolev inequality. The continuum limit remains open — as in all approaches to the Clay problem — but the correspondence is no longer merely computational.
A self-interacting mediator cannot be transparent. In Yang-Mills, that is the mass gap. In information geometry, that is the explaining-away penalty. Same constraint, different substrates. This is not a metaphor: Step 2 is a theorem, and the computational correspondence survives five pre-registered kill conditions.
Both papers publish their kill conditions before running tests. Current status shown below.
Full falsification catalog (26 master K-series, 13 Kramers, 8 Navier-Stokes, 90+ paper-level): Kill Conditions →
Different funders care about different parts of this arc. Quick links: