v0.3.0

Eight skills, one position on the method map each: every position now carries a skill.

New in this release: the description position's skill, sizing-opportunities, and gallop.explore.

• The skill turns a what-happened question into a localised, sized hypothesis: the floor before any story, mix versus rate, the funnel step, segments with the number of cuts declared, and the size as three numbers (the ceiling, the prior-anchored estimate, the MDE). The gap between self-selected groups is never quoted as the prize. Four verdicts, a hypothesis template for the hand-back.
• gallop.explore: mix_rate (exact Kitagawa decomposition with centred per-segment contributions and Benjamini-Hochberg flags), funnel_steps, scan_segments, size_opportunity. A fourth SQL template, segment_rates.sql.tmpl, and the bundled size_opportunity.py.
• choosing-causal-designs gains a guard: two units, one treated, is still DiD.
• Three evals for the new skill; a full run of all 24 on Sonnet is recorded in evals/README.md.

Install: /plugin marketplace add 0trm/gallop then /plugin install gallop@gallop; the package with pip install gallop-pds (the import name is gallop).

v0.2.0

First tagged release. Seven skills, one position on the method map each, and the package underneath.

New in this release: the prediction position's skill, automating-decisions, and gallop.validate.

• The skill decides whether a repeated, at-volume decision belongs to a model, which output it needs (forecast, ranking, allocation), whether targeting needs propensity or uplift, and validates out of time against the base rate at the operating point. Impact claims go to an experiment; offline metrics never enter the prior store.
• gallop.validate scores fitted output and fits nothing: time_split, leakage_screen, baseline_lift, calibration with the Murphy decomposition, qini, mase. Dependencies stay numpy, pandas, scipy.
• skills/automating-decisions/scripts/validate_model.py runs the pass in the skill's order.
• Three judgment evals, ablated on Sonnet.

Install: /plugin marketplace add 0trm/gallop then /plugin install gallop@gallop; the package with pip install gallop-pds (the import name is gallop).