git:20261027.bf61c98 to git:20261101.32e95ee

13 added, 10 removed. Audit A to A.

# Working on ADA

- ADA is a Python package (`ada/`) with a command line (`ada`), a research

- workflow (`ada/research/`: `ada init`, `ada build`, `ada verify`) and a

- Streamlit app (`app.py`, `ada/ui.py`, `ada/home.py`).

+ ADA is a Python package (`ada/`) with a command line (`ada`), tool

+ definitions for language models (`ada/tools.py`), an MCP server

+ (`ada/mcp.py`) and a chat command (`ada/chat.py`).

[docs/architecture.md](docs/architecture.md) says which module does what.

+ The research workflow, the modelling mode and the browser app are in

+ [ADA Lab](https://github.com/saineshnakra/ada-lab); `app.py` here only runs

+ ADA Lab's app for the hosted demo.

## Setup and checks

```bash

- pip install -e ".[app,ai,dev,model]"

+ pip install -e ".[dev]"

ruff check .

python -m unittest discover -s tests

```

- One test file: `python -m unittest discover -s tests -p "test_research_build.py"`.

+ One test file: `python -m unittest discover -s tests -p "test_claims.py"`.

CI runs the same checks on Python 3.11, 3.12 and 3.13, plus a job that builds

- the wheel and runs `ada init`, `ada build` and `ada verify` from it.

+ the wheel and runs `ada claim`, `ada ask`, `ada scan` and `ada mcp` from it.

## Conventions

- - Calculations live in the engine modules and `ada/research/`, never in the

- app; the app only draws results.

+ - Calculations live in the engine modules, never in the command line, the

+ tools or the MCP server; those only wrap the engine.

- Every number comes from pandas, SciPy or statsmodels and is shown with its

calculation. When something can't be computed honestly, refuse with a

reason instead of guessing.

- - Builds are byte-stable: the same study.yml and data give the same files.

- Don't put dates, temporary paths or random state in outputs.

+ - Results are deterministic: the same data and question give the same

+ answer. Don't put dates, temporary paths or random state in outputs.

- Code is formatted by hand: ruff check only, line length 110. Don't run

`ruff format`.

- Tests use `unittest`. A new rule needs a test, including how it behaves on

thin data.

- Docs, captions and messages are plain English in the present tense, with

no marketing words.

- Commit messages are one sentence that says what changed, such as "Skip

empty numeric columns in the value checks instead of failing on them".