Privacy and sensitive data
The statistics run on your computer; the AI model does not. Here is exactly what is shared, and how to protect sensitive data.
What leaves your computer, and what does not
Stays on your computer
- Your data files.
- The statistics: R, the models and the simulations all run locally.
- Reports, figures and tables, saved in your project folder.
Sent to the AI model provider
- What you type to the agent.
- Anything the agent reads or prints: if it opens a data file to check it, those rows are sent; so are summaries, column names, error messages, estimates and figures it looks at.
The AI model (Claude) runs on the provider's servers, not on your computer. Treat the conversation like an email to a careful colleague at another institution: fine for most ecological data, not fine for data you are not allowed to share.
Private mode
For sensitive data, type “Turn on private mode” in Claude Code. The agent
asks which folders hold sensitive data (by default data_raw and data) and then
works from structure and results without reading raw records:
- it blocks itself from opening, searching or printing files in those folders;
- it still runs CascadiaTools on them, because the toolkit prints summaries and results, not records;
- it sees only column names, numbers of rows, value ranges, missing-value counts, data-check messages and aggregated results.
To turn it off, ask the agent to turn off private mode.
It works by matching patterns in the agent's actions. It is a strong safeguard against accidents, not a security boundary: a command written to get around it could still print data. Summaries, estimates and figures are still sent. Review reports before you share them. For the strictest cases, run the toolkit without AI (below).
Locations of sensitive species
Den sites, nests, roosts and locations of rare or poached species need care. Two simple protections, usable together:
- Shift and rotate the coordinates before analysis. Add a secret offset to every
easting and northing and, if you like, rotate the whole layout by a secret angle. Do this to
every spatial file in the same way: traps or cameras, detections, GPS fixes, search tracks and
habitat grids. Work in projected coordinates (metres, such as UTM), not latitude and longitude.
Spatial capture-recapture, movement and other spatial models depend only on distances and on
which habitat value sits where, so the estimates (density per km², home-range size,
habitat effects) do not change. Maps will show the shifted coordinates; transform them back
yourself for internal use. Habitat rasters (
.tiffiles) are easy to shift but awkward to rotate, so if you have rasters, use a shift only. - Turn on private mode so the agent never reads the records themselves.
Many analyses (occupancy without a spatial term, N-mixture, survival, growth) do not need coordinates at all. Leave them out of the files you give the agent.
Running without AI
Everything the agent does can be done by hand. The modules are R code with a command-line
interface, and the protocols (data intake, planted truth, convergence, model checking,
interpretation) are written checklists. Rscript plugins/cascadia-tools/core/run.R list
shows the modules, and each module's worked example lists the commands that produced it. Nothing
is sent anywhere. You need to be comfortable running R commands.
Check the provider's data terms
How the AI provider stores and uses conversations depends on your account type (personal, team, enterprise or API) and its settings, and these terms change. Before you work with data under a data-sharing agreement, a permit condition or human-subjects rules, read the current terms for your account, and ask your institution whether it has an agreement with the provider. If in doubt, use private mode or run without AI.
