The act
An act is the selection of one element from the substrate’s permitted-act set, executed deterministically by the substrate. Every act carries a four-axis coordinate.
act = ⟨ A, σ, p, t ⟩
Four axes, one act
01 · WHO OCCUPIES THE SEAT
the occupant exercising conferred agency
agent-native
The occupant of the seat. The set is currently {H, S} — human and software daemon — and is extensible. The actor axis does not say anything about what kind of computation the actor performs. It says only who selected the act. The substrate is neutral with respect to the actor; a syscall is a syscall regardless of who emitted it.
Admissible classes
02 · HOW THE ACT IS CHOSEN
deterministic or stochastic
agent-native
How the act was chosen from the substrate’s permitted-act set. A deterministic selection is reproducible from the same inputs — a compiler emitting an instruction for a parsed token, a scheduler firing a job at a planned time. A stochastic selection draws from a distribution — wetware noise in human deliberation, sampled tokens in a language-model runtime, sampled actions in a reinforcement-learning policy at inference. The same actor can switch between modes across acts.
Admissible classes
03 · WHICH SUBSTRATE OPERATION
a coordinate in compute.openie.dev
inherited from compute
The substrate operation actually executed. The permitted-act set is the union of every primitive the substrate exposes — syscalls, memory writes, register loads, packet emissions, pixel updates, timer-interrupt acknowledgments. The set is large but enumerable, and every element is classified in the compute stack. The primitive axis is not native to agency; it is borrowed wholesale from the compute coordinate space.
Admissible classes
04 · WHEN ON THE SUBSTRATE’S CLOCK
universal — the act occurred at t
universal
When the act occurred on the substrate’s clock. Universal across all acts and all actors. Time is the only axis on which every act in the system can be ordered.
Admissible classes
The per-act cost
E(a) = θ(p) · μ(p, H)
Per-act energy · Landauer floor θ(p) scaled by impedance mismatch μ on hardware H
Every act incurs a thermodynamic cost. The Landauer floor θ(p) gives the minimum dissipation per primitive p; the impedance-mismatch coefficient μ(p, H) scales the floor by the hardware-specific overhead. Both factors depend on the primitive, not the actor.
A syscall emitted by a human at a shell and the same syscall emitted by a daemon dissipate the same heat through the same gates. The cost is incurred regardless of who selected the act, or by what means. A stochastic selection upstream pays no extra joules at the substrate — it pays them in the selector’s own computation, separately counted.
This is the symmetry at the heart of the framework. Agency is conferred uniformly. The substrate does not distinguish among occupants. What varies is fluency, throughput, selection process, and cost basis — all properties of the occupant, not of agency itself.
Continue