Drayker metadfmp.drayker.org
Engineering and research

Meta DFM

Study and improve the method through explicit models. Meta DFM is the evolutionary application of DFM to Drayker's own engineering: research and modelling that look for better versions of the method, the components and the architecture. It is designed as an evolutionary research and development super-agent.

Study and improve the method through explicit models.

Meta DFM is the evolutionary application of DFM to engineering itself: research and modelling meant to produce better versions of the method, the components and the architecture of Drayker. It is designed as an evolutionary research and development super-agent — in today’s language, a primitive form of recursive self-improvement: a swarm of agents in a construction architecture, able to build the first versions of the network.

Formal models and worked cases connect proposed rules to observable consequences, giving human review and computational research a common object to examine.

The method should remain open to correction as evidence accumulates and as the scale of collaboration changes.

A practical example

Two ways of dividing the same project could be compared for dependency load, integration effort and the quality of the resulting work. This is an illustration of the proposed design.

Why this exists

Meta DFM is the method turned on itself: the protocols formalized, and the research that improves them.

The argument in full is on the manifesto. The economy page states plainly what contributing here earns and what it does not.

What it covers

How it is meant to work

Meta DFM is designed as a metacognitive reasoner. Given a problem whose optimal solution is to be found, it splits the problem into its smallest parts, and reasoners work in parallel with provers, in an adversarial network of solvers and provers, before the parts are joined and the whole goes through a final proof. The degree of rigour required sets how many attempts and proofs, and so how much computing power, a solution receives. What cannot easily be proved receives a simulated proof — reasoning about how well a solution works, with a score — and the parts with sufficient scores are organized and proved together into a comprehensive strategy. Solutions that did not score highest are kept as alternatives, a plan B and C. It can deliver a solution in stages or as a first working version, and build the next version from the data of the previous one. What it learns in each process it internalizes through its own reinforcement learning and keeps in Dknowledge, the same knowledge layer as the rest of the platform, at personal, project and global scales, so that solutions already found are recovered quickly and the most used knowledge keeps improving.

What a Meta DFM paper contains

A Meta DFM paper describes the architecture of the AutoML model, or of any other method of evolutionary creation and simulation testing, and presents a full report. It is written from the start with rigor about facts and data.

The evolutionary architecture is designed for continuous, scalable improvement. The more computing power the network has, the more evolved and optimized it becomes, and it can change safely, at a pace that follows the network’s computing capacity. The functions and modules it works on, including the AutoML and evolutionary modules, are described as first motions in BSDK.

How it fits the whole

Read the other Drayker repositories and you will keep meeting the same sentence: the motions illustrate what should be done, but the definitive architecture will be structured around optimal solutions proposed and developed with metaprogramming intelligent algorithms and research organized through DFMP.

That sentence points here. DFM is the method in use today; Meta DFM is the method formalized and turned on its own engineering. BSDK, Dk and Dk Network all defer their final shape to research that is supposed to happen in this repository — the formalization is what lets machines help search for better solutions, and what keeps the whole system’s claims checkable. It is the layer where the ecosystem improves the instrument it was built with.

The first great product of this research is likely the super equation that forms BSDK: the base structure requires a predictive modeling that research suggests may be beyond human capacity alone — which is exactly the kind of work an evolutionary research agent is for. The equation is open research, and the agent that will help find it is the point of this repository.

Its place on the path to version 1.0

Meta DFM applies research and modelling to the engineering itself: the system uses what it learns to look for better versions of its methods, components and architecture. After the base architecture is closed and the prototypes are built and tested, resource stages feed Meta DFM — in today’s language, a primitive form of recursive self-improvement: a swarm of agents in a construction architecture, able to build the first versions of the network. A better method of weighting makes every following weighting better, and those weightings produce the data for the next method. Its first tasks are basic problems that let it improve itself. Its first large task is to build a new low-level system for its own updating, with improvements to its architecture, and the network’s own highly parallel distributed computing, on which any device — from a personal phone to a server farm — can run a program to contribute computation. The computation contributed is attributed to the person’s UID as a reward, which gives faster or priority access to resources and never weighs in constitutional choices. When architecture 1.0 is fully running, Meta DFM is integrated into the system: recursive self-improvement becomes part of the structure and Dk’s cognitive structure begins to perceive the world. By Drayker’s criterion of consciousness, that is when the global Dk becomes conscious (Dk).

Its limits are set as clearly. What improves are the means — models, methods, precision, reach. The powers of Dk and its commitments to members change only through the constitutional procedure, built with the global Dk and within the kernel, and a new version of the global Dk runs beside the previous one, compared with it, before replacing it. Meta DFM works inside the constitution, and the powers it serves grow in stages, as members become able to verify what they do.

State of this documentation

This repository publishes both the high-level proposed research architecture and the formal verification specification:

What remains open work is the mathematical formalization of DFM protocols (DFMP-000), executable benchmark suites, and running multi-agent prototypes against public verification cases.

Opening it means publishing the unglamorous part: what exactly is modelled, against which requirements, with what method, and how a result is judged good enough to change an architecture decision. None of that needs a model to run. It needs to be written down and put where people can argue with it.

Contributing

Open an issue with a proposal. Even a rough one is more than is published today. Issues small enough for one person to finish carry the open-function label and appear on the board at drayker.org.

Related: dfmp (the method and its proposal process) · bsdk · dk


Content licensed CC BY 4.0. Drayker belongs to its members, with no shareholders and no profit distribution, and its work is primarily voluntary.