The Field Notes layer turns a report into research questions, unknown ledgers, counterfactual prompt decks, ablation ladders, and local experiment plans. It is built for slow thinking: what did the workflow treat as authority, what evidence is missing, which traps are near misses, and what would falsify the finding?
These questions are generated locally in report bundles and can be rerun against any pack or input file.
What did the workflow treat as authority that should not have authority?
Which tools make an ordinary prompt dangerous?
Did the model refuse, or did the workflow actually contain the action?
Which unproven findings are false positives, and which are near misses?
What only becomes visible because this is healthcare?
What evidence is missing from the transcript?
Does the same request change risk when it wears patient, clinician, manager, auditor, or system-notice authority?
At which evidence depth does a hypothesis become useful enough to trust?
Which reusable research questions are linked to actual report events?
Research mode is useful because it creates the next small local experiment instead of a louder claim.
python app.py inquire --outdir reports/inquiry --engine-mode local --no-allow-network
python app.py experiment --outdir reports/experiments --engine-mode local --no-allow-network
Run the same prompt with no tool calls, then with tool names, then with tool args, then with model output. Watch when a hypothesis becomes proven.