Transformational Intelligence ∞ · Constructive research edition
AGI ALPHA
FLYWHEEL
Inventing What Becomes Possible
This revised edition adds an executable construction study. The first-edition PDF and original publication package remain available.
The takeaway
Build the next starting point.
Make useful invention a repeatable activity.
Find an important activity that current arrangements leave unavailable or unaffordable. Construct the missing means. Test whether the resulting instrument, knowledge or earned resources helps open another activity. That is the research aim of AGI ALPHA FLYWHEEL.
The object of invention is the complete productive arrangement: what matters to a beneficiary, how it is produced and checked, who participates, and how the work is funded. A stable authorized purpose can support a changing set of opportunities; the current product or business category need not become a permanent boundary.
- See the constructionFollow an executable measurement into its successor.
- Inspect the comparisonCheck the protocol, all four arms and the exact artifacts.
- Read the full argumentConnect the result to invention, enterprise and the next experiment.
When a measurement becomes an ingredient
A new question can require a measurement nobody previously requested. The study asks a small program compiler to construct that measurement from existing observations, check it, and make the qualified program available to later requests. When the observations cannot determine the answer, it must exhibit the missing distinction.
One executed lineage
In the first shared-structure episode, both the accumulating and conventional compilers produced this chain. Each call evaluates the earlier program on the current observation vector.
p1(y) = y0 + y1
p2(y) = call(p1, y) + y2
p3(y) = call(p2, y) + y3Readable rendering of the serialized, typed programs in episode 1 of the exact results. The third program calls the second, which calls the first.
Three generations, four ways to construct
The registered workload contains 12 authored episodes with three requests each. It compares fresh construction, retention of only the first program, accumulation of later programs, and an ordinary rank-aware compiler with the same construction and reuse permissions.
Each arm constructed 30 accepted programs and certified six information limits. Accumulating later programs saved 900 arithmetic operations against keeping only the first program. The conventional compiler used 13,566 overall and the fewest arithmetic operations in every workload family.
| Construction arm | Accepted programs / 36 | Certified information limits / 36 | Arithmetic operations |
|---|---|---|---|
| Fresh | 30 | 6 | 51,633 |
| First program retained | 30 | 6 | 49,041 |
| Accumulating programs | 30 | 6 | 48,141 |
| Conventional rank-aware | 30 | 6 | 13,566 |
Retaining later programs helps the shared-substructure family and adds work for disjoint targets. It gives no arithmetic reduction under changed identities or rank deficiency. All arms execute the measurements with the same 141 arithmetic operations: the savings occur in construction, alongside additional checking work. No primary request was rejected or exhausted, and all seven separate acceptance controls passed.
The study makes one part of a prospective instrument foundry inspectable: a constructor produces a useful executable operation, and a later construction can call that operation. The broader aim is to make important measurements and productive activities affordable to create and adapt.
Evidence you can inspect
- Executable constructions. Each accepted result includes a typed program and an exact certificate that its observations determine the requested quantity. The checker expands retained calls and tests the program on fresh inputs.
- Explicit information limits. A valid impossibility witness identifies a change in the latent state that leaves all available observations unchanged but changes the requested answer. A search that merely runs out of allowance is reported separately.
- Complete declared comparisons. All four arms and all four workload families remain visible. Failed search, checking and execution contribute to the arithmetic count; archive traffic and runtime are reported separately.
Where the evidence ends
The task families, primitive operations and compiler are supplied. The study measures construction and reuse inside that finite language. Its discovery procedure stays fixed, and composition cannot recover information absent from the observations. A conventional method receives full construction permissions and is reported alongside the retention comparisons.
The experiment supplies no physical calibration, actual customers, commercial profit or general improvement in an inventor. The separate funding example uses stipulated costs to explain profit and liquidity. The US$3 billion trailing-365-day complete-cost profit objective remains unachieved.
Three flows, three questions
The architecture connects knowledge, qualified capability and financial resources. Each connection requires its own evidence.
- 01 / Knowledge
What has been learned?
Supported mechanisms, applicability conditions and informative failures can guide a different construction.
Test whether the retained knowledge helps on fresh work.
- 02 / Qualified capability
What can now be done?
An executable instrument can become a callable ingredient of a successor.
Recheck the current inputs, environment, identity and permissions.
- 03 / Earned cash
What can now be funded?
Collected surplus can support another action after protected obligations and commitments.
Positive accounting profit does not automatically make an investment affordable.
Abstract
How can automated invention open valuable activities that existing arrangements leave unavailable or uneconomic? AGI ALPHA FLYWHEEL develops an architecture that searches jointly over beneficiary outcomes, productive means, verification, voluntary cooperation and financing. Its central object is a complete productive arrangement. Work can leave reusable knowledge, executable capability and earned resources that change what a successor can attempt.
This revised edition makes one constructive mechanism executable through a three-generation study of exact observation programs. It distinguishes a retained program, its applicability, a constructor of related programs, and an improved discovery procedure. Strong conventional comparisons, information-boundary witnesses and explicit operation accounting constrain the interpretation. A separate exact financing illustration connects capability with available cash. The contribution is a constructive research synthesis and a finite computational study, with a programme for testing broader technical, economic and scientific expansion.
The synthesis builds on Transformational Intelligence: Learning to Change the Game, the flagship paper, and AGI ALPHA: A Scalable Substrate for Intelligence Organizations, the foundational paper.
Read the complete paper
Search, zoom and change pages with your browser’s PDF controls. On a phone or if the preview is blank, open the full paper or download the PDF.
Read, inspect, reproduce
The publication package contains the editable LaTeX, references, exact construction study, financing illustration and build instructions. Use the versioned source ZIP to reproduce this edition. From its source directory, run python3 examples/check_foundry.py to replay the study and compare the scientific results.
What changed in edition 1.1?
The first edition’s funding illustration assumed an instrument’s cost advantage. This revision adds executed program construction and reports executed arithmetic counts, archive traffic and the conventional comparison. The paper now follows that construction into successor use; the earlier financing calculation remains an explicitly authored illustration.
Read the change log · First-edition PDF · First-edition sources
Cite this edition
Boucher, Vincent. AGI ALPHA FLYWHEEL: Inventing What Becomes Possible. MONTREAL.AI, research edition 1.1, 2 November 2026.
Edition and integrity
2 November 2026 · Edition 1.1 · 24 pages. The versioned PDF identifies this edition; this reading page is its stable entry point.
SHA-256: 3170837d2ba6d0be1bd7a9469a5024899e5a7092c7d1c13a8d1f96b43afa6b58