Research project / PRJ-2026-002

Program as Human-Readable Source

Research into whether a Markdown description of a program can become a readable source representation from which AI can build and maintain a working implementation.

running FY 2026-27
Project periodJul 2026 to Jun 2027
Core activities1
Recorded experiments2
Field of research4602: Artificial intelligence

Project objective

Determine whether ordinary software can be expressed as readable Markdown that preserves its meaningful algorithms, data structures, public functions and cross-file data flow while leaving target-language syntax and private implementation mechanics to an AI translator.

The immediate application is Vibe64. Vibe64 already lets a person ask an AI to build software, but the result is still conventional source code. This project asks whether the lasting source can instead be something that a non-specialist can read, review and change.

Why experimentation is required

Natural language is not a formally defined programming language, and current language models do not behave like deterministic compilers. A concise description may be readable but omit information needed to reconstruct the program; a sufficiently complete description may become as difficult to read as source code. It was not possible to determine in advance where that boundary would fall, whether generated implementations would retain observable behaviour, or whether failures could be corrected without turning the Markdown into pseudocode.

Core activity CA-01

Reconstructing software from human-readable Program modules tests whether file-level Markdown modules contain enough meaning for isolated AI translators to create or update conventional implementations that satisfy independent behavioural checks.

The work covers the Program format, bounded translation context, shared types, deterministic interface projections, candidate validation, incremental synchronization and clean reconstruction. Ordinary Vibe64 product integration is recorded separately from this experimental activity.

Records retained

  • dated hypotheses, design decisions and experiment transcripts;
  • versioned Program, prompt and compiler changes;
  • source and Program repositories used for clean-room reconstruction;
  • generated candidates, rejection diagnostics and accepted checkpoints;
  • independent and generated test suites;
  • complete test results and failure-cluster notes;
  • commits and private time and expenditure records.

Current position

The v2 experiment has demonstrated one complete source-free reconstruction from four production Program modules and one shared type registry. The mature and independently generated ProgSync implementations passed the same 55-case public oracle, and the generated compiler converged with Codex unavailable.

The experiment also found a clear synthesis defect despite sufficient Program. The verified managed implementation must therefore remain durable source alongside Program, reconciled against their last accepted pair rather than treated as disposable compiler output. Repeated clean reconstruction, generality beyond ProgSync and direct testing with the people expected to read and edit Program remain unproven.

Entity considerations

Ownership, control and financial risk

Project-level facts relevant to determining which entity conducted the R&D activities for its own benefit.

01

Effective ownership

Mobily Enterprises developed the prototype in its Vibe64 codebase and can use, modify, publish and integrate the resulting open-source work. The supporting company records retain the relevant contractual and IP details.

02

Direction and control

Mobily Enterprises selected the research questions, designed the trials, set the acceptance tests and decided when to reject, revise or continue an experiment.

03

Financial burden

Mobily Enterprises bore the researchers' time and model-compute costs for the work recorded here. Detailed remuneration, expenditure and any apportionment records are retained privately rather than on this public site.

Research entries

Experiments in this project

Each entry covers one experiment or material iteration and keeps its own dates and supporting records.

RD-2026-007 concluded

When the compiler writes itself, what is the source?

We replaced an implementation-shaped natural-language corpus with five small Program files, rebuilt ProgSync without its former source, and tested the result against an independent 55-case oracle. The build worked, but one clear synthesis error showed why verified generated code cannot be disposable.

23 Jul to 23 Jul 2026 1 days 15 research hours Core
#ai#natural-language#program#progsync
RD-2026-004 concluded

Can we turn source code into natural language?

We tested whether readable Markdown could preserve enough of a program for isolated AI translators to reconstruct working software. A 21-module trial succeeded; self-hosting exposed deeper contract failures and left the central claim unproven.

21 Jul to 24 Jul 2026 4 days 32 research hours · estimated Core
#ai#natural-language#program#progsync