[s4mstruthers]

What changed

Two corrections to the Inductive Miner's sequence cut (openprocess/mining/discovery/inductive.py), both taken from S.J.J. Leemans, Robust Process Mining with Guarantees (PhD thesis, TU/e 2017), which is what ProM's and PM4Py's implementations follow:

1. The every-pair test (maximal_sequence_cut, thesis p. 188). The definition of a sequence cut requires that every activity of an earlier group reaches every activity of a later group, and none reaches back. The old code merged groups and validated the cut with an any-pair test, so it accepted {b} | {a, c} | {d} for [<b,b,c,c,c,c>, <b,a,d,d>] although c reaches nothing; the definition gives {b} | {a, c, d}. Now computed as the thesis does: one group per activity, merge any two that hold a pair that reaches both ways or neither way, until nothing merges, then order by reachability and validate every pair.
2. The strict sequence cut (merge_skippable_groups, thesis p. 233). After the maximal cut, a group that some trace gets past without visiting absorbs the neighbouring groups that can only be entered through it or only leave through it, so an optional stretch is one skippable block: [<b>^4, <b,b,a,c>] gives →(↺(b,τ), ×(τ, →(a,c))) instead of →(↺(b,τ), ×(τ,a), ×(τ,c)), which also allowed <b,a> and <b,c>.

Around it: the How tab shows a line under a sequence cut when the strict step merged groups; the thesis is a reference on the IM topic (references.py, docs/references.md regenerated), so Cite on the box lists it; the module docstring explains the strict step; two tests pin the two examples and check the textbook log is unchanged; tests/cross_check_inductive_miner.py records both baselines; CHANGELOG under 0.11.2.

Effect

Cross-check against PM4Py 2.7.23 (seed 2026, 6 textbook + 400 random logs):

	identical trees	differ	of which less precise	guarantees hold
before	369 / 406	37	30	406 / 406
after	401 / 406	5	2	406 / 406

Three other seeds (300 random logs each): 302, 301 and 303 identical. The remaining differences are all fall-throughs on logs that have no cut at all, where the two implementations try the fall-throughs in a different order; precision is equal or goes either way.

This changes discovered models wherever a log has a skippable stretch. The textbook logs and every existing test expectation are unchanged. A saved analysis with such a log will show "This run differs from the record" after the update, which is the badge doing its job.

Checked

• Non-GUI suite with pandas/SciPy/matplotlib/PM4Py present: 297 passed.
• GUI suite offscreen: 132 passed (the root-only autosave test deselected, as in 0.11.2: pandas converters found by either name, what reproducible promises, docs/design.md, IM cross-check #61).

🤖 Generated with Claude Code

https://claude.ai/code/session_01CxRhDznFAbwU3x78bugqqR

---

Generated by Claude Code