system-open-questions

questions that don’t have a home yet, or that need thinking before they can become tasks. linked from system-map. when a question resolves, note the resolution briefly and leave it here – the history of how questions close is interesting.

supervisor vs study companion

what is the meaningful difference between the supervisor bot and the study companion bot?

study companion bot

• can a local model grounded in vault context approximate the calibration function, or does that require something closer to the full project understanding?
• what does the “holding the research frame” function require technically - is it system prompt, retrieval context, or something about how the model reasons?
• which of these functions are most degraded by lower capability models, and which survive?
• is the dialogic quality (thinking against output rather than receiving it) something that can be designed for, or does it emerge from the interaction?

shared context layer

• Chunking strategy: by heading, by paragraph, or fixed token windows?
• Which embedding model? (nomic-embed-text via Ollama is a reasonable local default)

confidence bot

• How does the Confidence Bot determine “shareable” threshold?
  This is partly a prompt engineering question, partly a research question.

liveness vs performance

is phd-live live, or also performative? does it need an audience, to be the latter? what would it take to move from live to performative?

how does this connect to the research questions that I initially laid out? supervision-290426

what to share publicly

obsidian config, research infrastructure, bots — what is the actual pedagogical intent of making these public? who is the imagined reader / learner?