Roadmap¶
Released versions record what shipped. Everything below the line is intent, not commitment, and the ordering matters more than the version numbers attached to it.
0.1 — foundation (released 2026-08-10)¶
- direct fuzzy calibration, logistic and piecewise
- crisp calibration
- typed fuzzy-set expressions
- necessity/sufficiency parameters of fit
- complete binary truth tables
- exact classical Quine-McCluskey
- conservative and parsimonious csQCA/fsQCA solutions
- pandas-native result objects
- parity harness against R
QCA
0.2 — parity and robustness (released 2026-08-11)¶
- correct intermediate solutions, with easy/difficult counterfactual classification matching R
- multi-value QCA
- enhanced necessity analysis: supersets, trivialness, relevance of necessity, SUIN conditions
- solution-specific unique coverage
- robustness sweeps over cutoffs and case removal
- Schneider–Rohlfing case typology
- calibration diagnostics
- prime-implicant chart inspection, so a solution can explain itself
- R parity extended to intermediate solutions, necessity screens, per-term fits and multi-value models, with the two known divergences pinned by tests
Unreleased¶
- exact minimisation roughly two orders of magnitude faster, via a bitmask cube representation
- phase-level benchmark harness
- a complexity warning raised before the exponential search, not after
Next¶
A second minimisation engine. A CCubes/eQMC-style backend, selected explicitly rather than substituted silently, so a fast approximate answer is never mistaken for the exact one the current engine guarantees.
Cross-language validation as a running check. The R fixtures are committed golden values today; the generator should run on a schedule so divergences surface when the reference implementation moves, rather than when someone next looks.
Simulation. Generating data with a known causal structure is what makes it possible to ask whether the method recovers what is actually there — coverage of the true solution, behaviour under noise and limited diversity.
Visualisation. XY plots, truth-table and chart rendering. Deliberately after simulation: a plot of an unvalidated result is a confident-looking wrong answer.
Provenance. Recording the calibration anchors, cutoffs and version that produced a result, so a published analysis can be reproduced from the artefact rather than from a description of it.
1.0¶
- a public API settled deliberately rather than by accretion, and then frozen
- temporal QCA (tQCA)
- multi-outcome models
- optional R-compatible calibration snapping, for replicating an existing analysis exactly
- benchmark corpus
- documentation rebuilt around tasks rather than modules
- published software paper