Walk up from this file until we find the directory with pyproject.toml.

Point estimates + 95% CIs for bias / RMSE / coverage / width.

Return parent/iter_NN<suffix> with NN one greater than the highest existing.

Load a skill body (everything after the closing `---` of the frontmatter).

Reproduce the (train_idx, test_idx) splits a discrete-treatment DRLearner used at fit time, so that each row's nuisance prediction can be routed

Add covered / abs_error fields when the path produced a usable answer.

First-iteration eval spec per plans/smoketest.md.

Mirror oci_agent.critic._parse_critique's fence extractor.

Mean |error|, RMSE, coverage, mean interval width for one path.

Partition by satisfaction (3 tiers) and print per-estimand rows.

Coverage vs RMSE with reversed x-axis: higher x = lower RMSE = better. Caller passes a hint; we enforce a floor of 0.5 so cross-figure compar

Ask the user whether to revise and run the next iteration. Returns False on EOF / non-tty / negative answer so the loop stops cleanly when st

Coerce `satisfaction` into {estimand: tier}. Accepts the per-estimand dict (new schema) or a single string / legacy `is_satisfactory` boolean

Check whether the 'python3' Jupyter kernelspec points to the current interpreter. This function does not modify the user's Jupyter environme

Compute true ATE/ATT/ATO from the raw zymu CSV. The ATO trimmed-row indices live in a sibling `keep_indices.json` written by the runner — the

58 ACIC-shaped covariates: a mix of continuous, binary, and count types. Real ACIC `x.csv` has a few object-dtype columns; we keep this all-numer

Generate treatment + potential outcomes for one ACIC response replication. Confounding: propensity and outcome both depend on x_1..x_8 (continuou

Read cached critique.json from a run dir, or call the Critic to produce one.

Return (satisfaction_tier, diagnostic_flags) per writing-reports rules.

Annotate each row with `llm_satisfaction` (same value for all 3 estimands of a run). Calls the Critic concurrently when parallel > 1; cached criti

Return the (x_1..x_58, z, y) view the analyst would see — no counterfactuals (y0/y1) and no ground-truth means (mu0/mu1).

Coverage vs RMSE — our ATT vs full ACIC 2016 competition (combined).

Coverage vs RMSE — ATE only, three judge tiers.

Parse a `--treatments` argument like "1", "1,3,5", or "1-77".

Aggregate summary for a multi-study run.

Confusion matrix at the (run x estimand) level: rows = deterministic (ground truth), cols = LLM. Both judges now emit a per-estimand verdict,

Programmatic validation against the writing-specs Validation rules.

One messages.create call. One retry on parse failure. No tools.

Execute one (spec, run_name) via subprocess. Returns (name, ok, last_stderr).

Run a single (treatment, response) pair end-to-end. Writes `<out_dir>/t{treatment}_r{response}/{contrast.json, plan.md, scaffolded/, baseline/

Vary the constant treatment-effect component smoothly across treatments so the smoketest battery sees a non-trivial range of true ATEs (~1.0..5.0)

Doubly-robust (AIPW) pseudo-outcome for the ATE per unit. Y_DR_i = m1(X_i) - m0(X_i) + (T_i / e_i) (Y_i - m1(X_i))

Estimate, standard error, and (1 - alpha) CI from DR pseudo-outcomes. When `centering` is supplied (array, same length as `pseudo`), the SE is

ATT influence-function-based pseudo-outcome. psi_ATT_i = (1 / pi) * [ T_i (Y_i - m0(X_i)) - (1 - T_i) (e_i / (1 - e_i

Compute ATT inverse-propensity weights. Parameters ---------- propensity_scores : array-like Estimated Pr(T=1 | X). tre