CodexProLong separates temporary model context from persistent research state.
problem + verifier
↓
Codex research turn ──→ Exa / Paperclip
↓
problem-specific program or executable world model
↓
local experiment ──→ frozen verification
↓
append-only journal + files + checkpoint
↺ next context
Every material observation, hypothesis, experiment, score, failure, and handoff was appended to a filesystem journal. Codex could search that history programmatically and keep working files across context rollovers. This follows the core insight of PRO-LONG: preserve the full trajectory outside the context window and let a coding agent retrieve what it needs with code.
The journal was not a polished knowledge base. Its value came from fidelity: failed attempts stayed available, successful programs remained executable, and a later context could inspect the evidence behind an earlier conclusion.
Different problems produced different machinery. Codex could write a parser, optimizer, simulator, exact checker, continuation method, or search program when that representation was useful. François Chollet describes this broader pattern as “LLM-guided on-the-fly synthesis of a symbolic world model”: the model makes its theory executable, tests it, and revises the program when the theory fails.
The research loop could act autonomously inside its local workspace. External submissions required human approval. Verification used pinned evaluator hashes, and public claims were reduced to solution IDs, scores, receipt hashes, and explicit integrity labels.
This document describes the system at a high level. The production controller, prompts, session transcripts, raw journal, research corpus, solver source, checkpoints, and executable world models are not included in the current release tree. Earlier source-bearing revisions remain available under their original MIT grant; see NOTICE.md for the licensing boundary.