Changelog

0.2.0 (2026-09-26)

Checks on the diagram itself.

Codebook: ca.py codebook drafts what each column means and when it was recorded; the expert confirms. Unconfirmed meanings lower the grade.
Reverse causation: the interview asks whether past outcomes drove who got the action; outcome_baseline names the before-the-action measure, and its absence is flagged.
Alternative diagrams: re-estimates under user-named alternatives (alternatives, with illustrative for trap demonstrations), each control dropped, and each left-out column added. New report section "What if our diagram is wrong?".
Structure second opinion: a light PC-algorithm search flags collider-like controls and unused columns linked to both action and outcome, as questions.
Planted-effect simulation on the user's own data (sim_reps by budget); failure lowers the grade.
Placebo uses 5 shuffles. Report diagram now draws every node and added arrow at full width. "Who gains most" only names a group when the difference is real.
New example and eval: AI-training case with a misleading mediator name, a collider and reverse causation.

First public release.

Guided workflow: data profile, plain-language interview, causal diagram sign-off, pre-registered plan.
Estimators: doubly robust ML (main), double ML, causal forest, regression, propensity weighting, g-computation for amount treatments.
Diagnostics: overlap, balance, placebo, random common cause, subsample stability, Cinelli–Hazlett sensitivity, bad-control illustration.
Abstention: grade D with Manski / Manski–Pepper bounds, instrument bounds and complier effect.
Optional foundation-model cross-checks: CausalPFN (local) and TabPFN (hosted, opt-in).
Designed HTML report (ca.py report) opening with a data overview (shape, column roles, distributions, sample rows; report_sample_rows: 0 hides the rows), causal diagram drawn and explained in words right after the data (ca.py dag), test sizing (ca.py power).

0.1.0 (2026-09-26)

First public release.

Guided workflow: data profile, plain-language interview, causal diagram sign-off, pre-registered plan.

Estimators: doubly robust ML (main), double ML, causal forest, regression, propensity weighting, g-computation for amount treatments.

Diagnostics: overlap, balance, placebo, random common cause, subsample stability, Cinelli–Hazlett sensitivity, bad-control illustration.

Abstention: grade D with Manski / Manski–Pepper bounds, instrument bounds and complier effect.

Optional foundation-model cross-checks: CausalPFN (local) and TabPFN (hosted, opt-in).

Designed HTML report (ca.py report), causal diagram (ca.py dag), test sizing (ca.py power).