Modern science runs on software that most scientists were never trained to use — and can't easily tell whether to believe. Caliper does the analysis for you. And it does the one thing no tool does today: it tells you, honestly, when the answer is solid and when you should ask a human.

It is a ready-to-use agent. The calibrated-trust-and-feedback layer is what makes it worth trusting unattended.

Existing science agents build a curated toolbox and an agent to drive it, then hand you an answer with no bound on being wrong — so a resource-constrained scientist still can't act on it unattended. Caliper closes that gap, across fields.

One core, swappable packs. Ships with a working bio pack (genomics); astro is a skeleton proving the core is domain-agnostic.

Most science-AI agents focus on capability — assembling tools and automating the analysis. Caliper adds the piece they leave out: an honest, calibrated signal of when to trust the result, and the discipline to hand borderline cases back to a human.

Others hand you an answer. Caliper hands you an answer plus a provable bound on how often it is wrong when it doesn't ask for help — and sharpens that judgment every time you correct it.

_{Comparison reflects each category's described design at the time of writing. These are complementary efforts; Caliper's trust layer can in principle sit on top of an existing agent.}

Give Caliper a rule in plain terms — "never let more than 1 in 10 of the answers you hand me unchecked be wrong" — and it keeps it. From a modest set of expert-checked examples it learns exactly how confident it must be before answering on its own; anything below that bar it escalates. The guarantee is finite-sample and distribution-free, and it holds even when the underlying judge is imperfect — because it would rather ask for help than mislead you.

pip install -e .
python examples/bio_demo.py        # full run: analyze → grade confidence → decide
python examples/feedback_loop.py   # watch it grow more confident as feedback arrives
python -m unittest discover -s tests

Provider-agnostic via make_llm(provider=...) or the CALIPER_PROVIDER env var.

from caliper import make_llm
llm = make_llm()                              # default
llm = make_llm("openai")                       # OpenAI
llm = make_llm("openai", model="gpt-5-chat-latest")

Caliper deploys as two cooperating tiers, so your data never leaves your own machines:

• A control host — a small, always-on server (a VM in your private network / VPC, or any low-cost cloud instance). It serves the web app, runs the agent and the trust layer, handles login + audit, and keeps chat history. It stays light — it holds no large data.
• Your private analysis server — where the raw data and the field's tools already live (on-prem or in your VPC). The control host connects to it over a secure connection and runs each analysis step there, inside a confined working directory.

   Browser ──HTTPS──▶  Control host  (web · agent · trust · login + audit)
                            │   per step: secure connection
                            ▼
                     Your private server   (data + tools)
                     • runs each step in a confined workspace
                     • raw data stays here — only small results return

In practice:

• Users reach it at a domain you control, over HTTPS, behind a login (per-user email + password; every sign-in is audited — who, when, and source IP).
• Every analysis step is workspace-confined: it writes only to a dedicated directory and reads your data read-only.
• Large data never leaves your server — only small results and logs flow back.
• History and the experience log are kept as files on both tiers — no database required.

To deploy: point a control host at your private server (its address + an account), set your model provider's API key and a login, and expose a domain over HTTPS. The control host is intentionally tiny; all heavy compute runs on your server, next to the data.

Research preview. Working: thin core, bio pack, calibrated gate (tested), multi-provider models, live feedback recalibration, offline + live demos. Next: exact-validity calibration (Learn-then-Test), a real genomics tool environment and a reproduced published study, distribution-shift robustness, and a fleshed-out astro pack. See docs/DEVELOPERS.md for the architecture and the math.

Apache 2.0 — see LICENSE. Caliper is an independent reimplementation; it reuses no third-party code verbatim. Prior art is credited in NOTICE. Underlying domain tools carry their own licenses; review before commercial use.