Privacy & data handling

ImportantThe promise

Biotic data is confidential, and some of it is sensitive (e.g. Russian-zone data, exact positions of vulnerable species). BAIT is built so your raw data stays on your machine and never trains a model.

Why an AI agent doesn’t “train on” your data

A coding agent reads files to help you in the moment — it does not update any model’s weights. Training would only happen if both of these were true:

  1. data is sent to a model provider, and
  2. that provider is allowed to retain/train on it.

BAIT closes both doors: it keeps data local, and it reminds you to turn provider training off.

Before you start — the pre-flight

  1. Turn training / data retention OFF for your agent (see below).
  2. The database lives outside any repo (~/IMR_biotic_BES_database/).
  3. You’re in a folder that won’t be committed with data — BAIT’s .gitignore / .claudeignore block data files as a safety net.
  4. For sensitive subsets, agree up front on what may be produced or shared.

Turning training off (verify the current setting)

Menus change, so confirm on your provider’s current privacy / data-controls page:

  • Claude / Claude Code: disable the “help improve Claude” / training option (consumer); API / Team / Enterprise inputs aren’t used for training by default, with zero-data- retention available.
  • OpenAI / Codex / ChatGPT: Data Controls → turn off model-training; API isn’t trained on by default, ZDR available on eligible accounts.
  • Other agents: enable the equivalent “privacy mode” / “do not train” / ZDR setting.

When in doubt, prefer API / enterprise tiers with zero data retention.

What’s safe to share

Output Generally OK to share?
Aggregates, counts, summaries ✅
Model parameters (L50, L∞, K, a/b) ✅
Figures & maps (non-sensitive areas) ✅ usually
Maps of sensitive positions ⚠️ aggregate/jitter; ask first
Raw individual records ❌ keep local
Anything Russian-zone / protected-species, position-level ❌ ask first

Dashboards

Run them locally. Never deploy Biotic data to a public host (e.g. shinyapps.io). If you need hosting, it must be inside IMR infrastructure with access control.

If something leaks

Stop, tell your data manager, delete the exposed copy, and rotate any shared access. Data sent to an external service may persist even after deletion — act per IMR policy.

Full operational detail lives in the repo: skills/biotic-privacy/SKILL.md.