Key findings

ChatGPT Work

Authenticated test turns selected each of gpt-6-astra-wm, gpt-6-sol-wm, and gpt-6-luna-wm. Each browser request contained one user message and the model selection, with no client-visible system, developer, or prompt field. An automatic voice prefetch selected Luna and advertised an empty client tool list, but it was not a completed microphone call. The Work service's instructions and completed-call tool surface remain server-side or otherwise unobserved. Work evidence · Sanitized three-model trace · Voice evidence

Codex GPT-6

The Codex catalog exposes instruction records for Astra, Sol, and Luna. Ten of their eleven model_messages fields are identical, including persistent_instructions; each model has a different base instructions_template. The persistent text applies when that mode is enabled. Model comparison · Astra · Sol · Luna

The provenance inventory hashes all 39 string leaves across the three records and 14 bundled desktop helper templates, and labels catalog defaults separately from conditional, bundled, and observed-active surfaces. Prompt provenance inventory · Helper prompts

Voice

The desktop package contains seven Codex voice prompt strings covering new and resumed conversations, coordination, memory, and realtime sessions. Their tool references describe Codex thread and screen actions; they do not establish a ChatGPT Work voice tool inventory. Codex voice prompts · Tool surface

The earlier Aeon core prompt remains in the historical archive. It is a separate source from the current Codex catalog.