entry_id: RD-2026-004
project_id: PRJ-2026-002
core_activity: CA-01
title: Can we turn source code into natural language?
timezone: Australia/Perth
period:
  started: 2026-07-21
  ended: 2026-07-24
research_effort:
  hours: 32
  basis: estimated
  note: Reconcile this public estimate with private contemporaneous company time records.
source_record:
  codex_session_id: 019f8239-a92e-77a3-bb74-316eb0c1dd9a
  repository: mobily-enterprises/vibe64
  package: packages/progsync
  initial_prototype_commit: 574c5203fd3d0ec86d559aaf920979ab620695cd
  final_v1_checkpoint_commit: c44e33568240f8ed2ee095fbc9d92f434e6f8471
  note: >-
    The complete session transcript, working repositories and model-run logs are
    retained privately. This public record contains no customer material,
    credentials or private remuneration information.
hypothesis:
  recorded_at: "2026-07-21 12:36 +0800"
  statement: >-
    If a Markdown Program preserves public symbols, shared complex types,
    cross-module calls, data flow and observable behaviour while omitting
    private implementation mechanics, fresh isolated model calls should be able
    to reconstruct a working implementation without access to the former source.
primary_success_condition: >-
  Generate every managed ProgSync implementation file from Program and retained
  non-generated inputs, then pass all 73 untouched original tests without
  manually editing generated JavaScript.
controls:
  - Original JavaScript was absent from clean generation repositories.
  - Each file was generated by a fresh module-scoped model invocation.
  - Generated candidates could write only their selected module.
  - Deterministic parsers checked the candidate surface before installation.
  - Untouched original tests were retained as the primary behavioural oracle.
  - Generated tests were recorded but were not allowed to certify themselves.
  - Behavioural equivalence, not byte-for-byte source identity, was assessed.
result: inconclusive
reason: >-
  A smaller 21-module application regenerated and passed its behavioural checks.
  ProgSync itself reached 73/73 after bounded Program and compiler repairs, but
  no frozen, zero-intervention clean self-host run had completed successfully by
  the end of the recorded period.
next_test: >-
  Freeze Program, prompts, model settings and retained inputs; complete two
  independent zero-source self-host generations and require both to pass the
  untouched 73-test oracle without Program changes or implementation repair.
follow_up:
  entry_id: RD-2026-007
  url: /research/rd-2026-007-when-the-compiler-writes-itself/
  note: >-
    The follow-up replaced the implementation-shaped v1 graph with a smaller
    five-file v2 Program architecture. It is a separate experiment and does not
    alter the result recorded here.
