Deep Research Intake Preparer

• fixed-base limitations motivating VMS,
• model-based preference over model-free or data-driven methods,
• non-geometric versus geometric model-based approaches,
• energy-based modelling as the chosen path,
• port-Hamiltonian interaction control,
• safe interaction and passivity.

Default Bucket Counts

Adjust counts based on gaps in the current papers/relevant/ corpus.

Intake Handoff Instructions

After delivering the prompt, tell the user:

• Run the prompt in ChatGPT Deep Research or equivalent.
• For each selected paper, create a folder under papers/download/<id>/ using papers/download/999/ as the template.
• Each folder should contain the paper file, a reference.bib, and a summary.md with only the title, abstract, and conclusion filled in.
• The agent completes the remaining summary.md sections once the folders are in place (Stage 4B).

Common Mistakes

• Generating a prompt that does not specify an output format — the user then receives results that cannot be converted to intake folders.
• Omitting anchor papers — without them, the search may return off-topic work that uses different terminology.
• Setting bucket counts too high with no selection criteria — leads to downloads that swamp Stage 4B without improving the corpus quality.