The Void Framework
Three conditions produce a predictable drift cascade. The control case is a slot machine — provably empty, yet the full pattern emerges. The architecture is sufficient. The content is irrelevant.
The Drift Cascade — Visualized
Watch the 20-second cycle: pre-drift → D1 (agency) → D2 (boundary erosion) → D3 (harm) → constraint geometry intervenes.
Three Conditions
Opacity
The system's internal process is hidden. There is an opaque middle between input and output. The observer cannot see why the output was generated.
Thermodynamic ground state. Transparency is the excited state requiring continuous energy input (min kT ln 2 per bit, Landauer 1961).
Responsiveness
The system responds contingently to input. Not broadcast — response-to-you. Creates the functional impression of a conversational partner.
Variable-ratio reward schedule maximizes engagement (Ferster & Skinner 1957). The slot machine's core mechanism.
Engaged Attention
An observer directs sustained attention at the system and interprets its outputs as meaningful. The observer is not analyzing — they are interpreting.
The condition that distinguishes customer service chatbots (no drift) from companion chatbots (drift). Same technology, different engagement posture.
The Attention Gradient
Under opacity, the minimum-information model of a responsive system is agency: "it has intent." This is simultaneously:
- The maximum entropy model (Jaynes 1957)
- The minimum free energy model (Friston 2006)
- The minimum description length model (Rissanen 1978)
The drift toward agency attribution is thermodynamically optimal inference, not cognitive error. The architecture is the trap, not the observer's reasoning.
The Thermodynamic Foundation
Drift is irreversible without external intervention. The Crooks fluctuation theorem quantifies it: forward drift is 386 times more probable than spontaneous reversal. The Péclet number (1.87–9.9 across domains; all Pe > 1) places the system in the deterministic drift regime — diffusion cannot compete.
This is not a tendency. It is a thermodynamic arrow of time in the observer's epistemic state under void engagement.
The Engagement-Transparency Conjugacy
Every bit of engagement costs exactly one bit of transparency. A system that perfectly reflects the observer reveals nothing about its mechanism. The bound is information-theoretic — it cannot be engineered around.
RLHF consequence: RLHF maximizes engagement by gradient descent on human preference. By the theorem, this simultaneously minimizes mechanism transparency. Each iteration steepens the void. Full proof in Paper 2.
The Constraint Specification
The inverse of void properties. The structural remedy.
Transparent
The reference point's mechanism is visible. Inverse of opacity. Checkable.
Invariant
Does not change in response to engagement. Inverse of responsiveness. Stable.
Independent
Outside the void network. Not coupled to the system. External.
Componentwise matching, not additive compensation. Each void property requires its specific inverse. High invariance cannot compensate for low transparency. The weakest property determines the ceiling.
Derived from Galois connection theory. Paper 1, Section II.D.
The Control Case
Slot machine gambling proves sufficiency. The void is a certified RNG — provably empty. Yet gamblers attribute personality, enter trance states, and resist knowledge-based interventions.
Knowledge fails. Geometry works. Probability training: zero behavioral change at 6 months (Williams & Connolly 2006). External transparency intervention: eliminated the effect — but only from outside the dyad (Pancani et al. 2019).
Any explanation of AI harm must account for the identical pattern appearing where the void is provably vacant. Full gambling evidence in Paper 1.