Computing Machinery and Understanding: The Turing, a Unit of Net Held-Out Explanatory Compression, with a First Coder-Checked Measurement and a Discovery Gap Conjecture
Description
In 1950 Turing replaced the question "can machines think?" with an operational experiment. This paper proposes an analogous substitution for a narrower question: has a machine found a compact explanation that makes previously unseen observations less surprising? I define Turing gain, T(M;B,D,G,P) = [L(B|G) + L(D|B,G)] − [L(M|G) + L(D|M,G)], the net reduction in total description length that a candidate explanation M achieves over a declared baseline B on sealed held-out observations D, given a declared shared background G and under a versioned measurement profile P. One Turing (1 T) is one bit of this net held-out gain. The quantity is a disciplined use of the minimum description length (MDL) tradition, not new mathematics; the contribution is the named reporting unit, the profile that fixes what counts, and the sealing, baseline and energy-accounting rules that make a reported number checkable. The Turing measures one operational component of understanding, the predictive and explanatory-efficiency component. It does not measure causal direction, semantic grounding, agency, value alignment, safety or consciousness.
I report three measurements on traces of the Omega program-search system, all public at the commit level in the aien-dev/omega-mv repository. (i) A first idealized result under a historical profile (TY-1+TY-2): +2,559,679.825 T_ideal over 5,834,845 held-out events, about 0.439 T per event, against an order-1 baseline, with a ladder of alternative baselines (one of them stronger than the declared baseline). (ii) A preregistered instrument-validation experiment, EXP-001, which failed at its freeze-order overlap gate before any scoring. (iii) Its preregistered successor, EXP-001R, which passed all nine criteria: an idealized gain of +2,714,280.19 T (95% crumb-bootstrap interval [2,516,825, 2,914,129]) over 6,168,907 sealed events, reproduced by an arithmetic coder (+2,713,794 T) and an ANS coder (+2,714,282 T) with every coder overhead inside the declared envelope, and with two memorizer controls scoring strongly negative. These results support the idealized-to-real coding bridge (H1) for this one regime only, with declared protocol deviations listed in the results section. No energy was joined to any result, so no Turing yield (T/J) is reported.
Finally, I state the Turing Discovery Gap: worlds whose compact explanations are worth Ω(n²) Turings but which no efficient learner can find. Assuming one-way functions exist, I prove that such a family exists (Theorem A). The unconditional version and an empirical program-reuse hypothesis are stated as conjectures; the experiment that would test the latter has not been run.
About this version. This corrected manuscript combines and supersedes three records, which stay available: Computing Machinery and Understanding: From the Imitation Game to the Turing (10.5281/zenodo.23437498, the previous version of this record), The Turing: A Unit of Verified Explanatory Gain (10.5281/zenodo.23547801) and The Turing Discovery Gap Conjecture (10.5281/zenodo.23547803). Checked facts from The Turing as a Measure Emerging from AIEN (10.5281/zenodo.23437568) are used; its narrative is not. Corrections include: EXP-001 is reported as FAIL and EXP-001R as PASS (the earlier text said "pending"); the journal-template submission text is removed; a citation misattributed to Lifschitz is corrected to Gurevich; the held-out claims and two typed-in figures from 23437568 are not carried over. The trace corpus is not public; the profiles, receipts and reports are public at the commits cited. Not peer reviewed.
Use of AI tools. Drake Stapleton directed this work and is responsible for its content. AI coding and writing assistants, including Anthropic Claude models used through the omp agent harness, wrote most of the code, ran the measurements and drafted the text under his direction. No AI tool is an author. The proof technique of Theorem A follows an independent draft by the GPT-6 Astra model, shared with the author on 2026-09-30. Every measured number is tied to a public commit or pull request listed in the paper, except where the text states that its source is not public.
Files
turing-measure.pdf
Additional details
Related works
- Is derived from
- 10.5281/zenodo.23547801 (DOI)
- 10.5281/zenodo.23547803 (DOI)
- Is supplemented by
- https://github.com/aien-dev/omega-mv (URL)
- References
- 10.5281/zenodo.23437568 (DOI)