anyflights vs assesslite

AssessLite attacks the structural assumptions behind a causal finding and returns three-way verdicts — stable, unstable, or not resolvable — feeding proceed, conditional or abstain decisions, with an auditable JSON record validated against a shared schema. It runs natively in R and Python against one spec, with the Python engine reproducing R's coxph(ties=breslow) exactly. The entire 0.1.0-through-0.4.0 arc landed inside a single day in July 2026.

The releases are cumulative, each restating the previous feature set and adding to it, so read them as one launch rather than four. The direction across that launch is clear: it started with resampling attacks (permutation, holdout, temporal split, subgroup), turned toward causal identification with declared DAGs and the backdoor criterion, then reached into genuinely dependent data with spatial and interference checks. The correctness work moves in step — the 0.3.0 Bonferroni adjustment fixed a holdout rule that was flagging roughly m times too often with m variants.

The project has repeatedly shipped what it previously listed as future work within days, so the next release most likely converts another declared gap rather than opening a new front.