Computing Machinery and Understanding
From the Imitation Game to the Turing: An Operational Measure of Generalized Explanatory Compression
Abstract
In 1950, Alan Turing confronted the question “Can machines think?” not by attempting to define thought directly, but by replacing the question with an operational experiment: the imitation game. That methodological move transformed an ill-posed philosophical dispute into an empirical research program.
This paper proposes an analogous reformulation for a contemporary question: does a machine understand? One measurable component of understanding is operationalized as the discovery of compact structure that continues to predict previously unseen observations. The proposal combines Shannon information, predictive coding, Minimum Description Length, empirical verification, and physical resource accounting.
A new derived quantity is defined, Turing gain. One Turing (1 T) is one bit of net reduction in held-out total description length, after paying the explanation’s own description cost, measured against a declared baseline under a versioned measurement profile with sealed data. Turing yield (T/J) normalizes that gain by the energy consumed to achieve it.
What this paper claims
The paper defines the measure, states the assumptions under which it is meaningful, gives the measurement protocol (declared baselines, sealed held-out data, recomputable evidence), and specifies the conditions under which the proposal would be falsified. It names a preregistered experiment ladder and explicitly separates preregistered future experiments from apparatus-calibration runs, whose outputs are never reported as measurements.
The empirical-results section records the first canonical measurement: TY-1+TY-2 PASS, +2,559,679.825 T under measurement profile V0 (omega PR #85, 2026-09-29). What it does not claim: Turing yield (T/J) is not yet measured; joining energy measurements to Turing-gain receipts is the next step.
Gain, and the yield still to come
The paper separates two questions. Turing gain asks how much net held-out compression an explanation earns. Turing yield asks what that gain cost in energy. The first canonical measurement answers the first question only: no energy was measured in that run, so no yield is reported yet. The diagram below restates the paper's gain-versus-yield figure for the web.
Experiment ladder
EXP-001 — Omega-trace instrument validation
Does the Turing measurement work reproducibly on real Omega execution traces?
EXP-002 — Stochastic Law Discovery
A hidden evaluator generates quantized trajectories from an undisclosed stochastic law — pure diffusion, diffusion with drift, or mean reversion. The candidate must earn complexity on sealed held-out trajectories without inventing structure inside noise.
EXP-003 — Active experiment selection
The machine proposes the most discriminating intervention.
Research artifacts
Cite
@misc{stapleton2026turing,
author = {M. Drake Stapleton},
title = {Computing Machinery and Understanding:
From the Imitation Game to the Turing:
An Operational Measure of Generalized
Explanatory Compression},
year = {2026},
month = {September},
version = {1.9},
note = {Preprint, not peer reviewed},
url = {https://aienos.com/research/computing-machinery-and-understanding}
}