When the cost of producing software approaches the cost of writing a detailed specification, the source code is the specification.
That is the claim. Everything else follows from it.
What “source code” has been
Through the entire history of software, source code has been a programming language: a notation that compiles into machine instructions through a long chain of human-built tools. The notation imposes constraints — syntax, type systems, idioms, the conventions of the toolchain. The notation requires study to read and study to write. People who want a piece of software written either learn the notation themselves or hire someone who has.
The result has been that the people who can produce software are a narrow population. Domain experts who know what software should do — legal aid attorneys who understand a benefits application, clinicians who understand a diagnostic conversation, teachers who understand how students learn, accountants who understand a tax form — typically cannot produce the software themselves. They describe what they need to a programmer; the programmer translates the description into a notation; the notation gets compiled to machine instructions; the resulting software approximates what was described, with the approximations introduced by the translation step.
This arrangement was imposed by the cost of producing software. The translation step was necessary because there was no way to skip it. Writing detailed specifications was just as much work as writing code, and the code had to be written, so the specifications were typically written informally, in the form of conversations between domain experts and programmers, and the result was that domain expertise rarely survived the translation cleanly.
What changed
Cognitive automation changed the cost of the translation step.
A model that can read a detailed specification in natural language and execute against it does not need a programmer in the middle. The specification is the source code. The execution is whatever the model does when the model reads the specification and runs the work the specification describes.
This is not a metaphor. A framework — Ora’s name for a structured cognitive specification — is a literal artifact: input format, processing steps, decision points, output format, written in natural language. The model reads it. The harness around the model executes the specification step by step. The output is what the framework’s design intended.
The translation step has not disappeared; it has moved. The translation now happens at runtime, inside the harness, between the natural-language framework and the model’s response to it. The translator is not a human programmer; it is the model plus the harness’s structural commitments. The specification stays in natural language, where the domain expert wrote it.
Who can contribute
This rearranges who can produce software at the consumer-facing layer.
A legal aid attorney can contribute a framework for benefits applications because she knows what the form requires. She does not need to write code. She writes the specification — what inputs the framework takes, what steps it walks the user through, what decisions it surfaces, what output it produces. The specification is written in the kind of careful prose she already writes when she trains paralegals. The framework is the specification.
A clinician can contribute a framework for medical-information synthesis because he knows what a clinically useful synthesis looks like. He does not need to write code. He writes the specification.
A teacher can contribute a framework for a particular kind of learning support because they know what their students need. They do not need to write code. They write the specification.
The contribution barrier is domain expertise, not technical skill. This inverts the conventional open-source contribution model in a way that opens contribution to populations conventional open source has not reached.
What this does to commercial software at the consumer layer
The economic moat that has protected commercial software at the consumer layer is the cost of producing it. Intuit’s TurboTax, the various small-business accounting suites, the legal-document services, the basic medical-information products, the household financial planning tools — these all justify their pricing on the cost of having produced them and the ongoing cost of maintaining them. The price-to-cost ratio is sustained by structural advantages: format lock-in, distribution control, brand recognition, network effects.
When the cost of producing a framework that does the same underlying work approaches zero — because the framework is a natural-language specification that the model reads and executes — the structural advantages have to do all the work. Some of them can. Format lock-in is real where formats are proprietary and industry-standard at once. Network effects are real where the value of the software comes from other people using the same software. Distribution control is real where commercial channels gate access to the user base.
But many commercial software products at the consumer layer are not protected by any of these. They are protected only by the cost of having produced them, which is no longer the protection it used to be. A framework that lets a household file its own taxes for free dissolves a chokepoint that has been lobbied to remain in place for decades. A framework for routine legal documents dissolves the cost gap between paying a paralegal and producing the document yourself. A framework for basic financial planning dissolves the advisory-fee structure that prices out moderate-income households.
The Foundation’s framework library is not a project to write a few useful pieces of software. It is a project to make the production of software at the consumer layer accessible to the populations who know what the software should do. Whatever exists in the library at launch may not exist when people need it most, because the disruption itself disrupts framework-development capacity. Whatever does not exist at launch will be produced afterward, by domain experts contributing what they know.
What this does not do
Natural language is the source code for cognitive specifications. It is not the source code for arbitrary software. The model executes the framework against accumulated state, structured input, an adversarial verification pipeline, and the rest of the harness that converts probabilistic outputs into reliable execution. The framework specifies cognitive work; the harness executes it. Software that does not reduce to cognitive work — operating systems, network protocols, low-level computing infrastructure, real-time control systems — remains in the conventional notation, written by programmers, for the foreseeable future.
The contribution model also does not eliminate the need for review. Frameworks that meet the Foundation’s specifications are subject to review by domain experts and active library contributors. The review checks whether the framework does what it claims to do, whether it documents its limitations clearly, whether it conforms to the format conventions the library requires. Multiple frameworks for the same task can coexist; the library is not exclusive. But contributions that don’t meet the specifications don’t enter the canonical library.
And the model still has to be capable enough to execute the specification. Frameworks that exceed the model’s capability — frameworks that ask the model to do work the model genuinely cannot do — fail in the way single-pass model use fails. The architecture is not magic. It is what becomes possible when the cost of writing detailed cognitive specifications drops far enough that the specifications themselves can be the unit of contribution.
Why public domain
The contribution model only works if the artifacts contributed cannot be enclosed.
A framework library that requires permission to use is not a public good. A framework library that requires attribution can become a contested attribution claim. A framework library released under a copyleft license preserves openness through copyright enforcement, which is a tool that can be challenged in jurisdictions where copyright does not apply or applies differently than the author intended.
CC0 1.0 Universal — the strongest public-domain dedication available — takes the licensing apparatus off the table. There is no permission to grant. There is no attribution to enforce. There is no condition that can be challenged. Anyone can use, modify, distribute, build upon, fork, or republish any framework, for any purpose, without permission, attribution, or fee.
The summary
When the cost of producing software approaches the cost of writing a detailed specification, the specification is the source code, and the people who can write detailed specifications about what software should do — domain experts in their own fields — become the people who can produce software at the consumer-facing layer. The framework library is the operational form of that recognition. Public-domain dedication is what makes the contribution model durable. This is what natural language programming is, and it is the structural shift the Foundation’s framework library is built on.