Directions

This chapter lists possible extensions. None is currently implemented or scheduled.

Discovery¶

• Multiple breakpoints. The residual-shift pass tests one break on the first temporal column. Segmented regression with several breaks (Muggeo-style iteration, or PELT-style dynamic programming) would catch ramp-then-level patterns that one break averages away.

• Wider itemsets and self-sufficiency. Association mining is bounded to pairs of nominal values. Three-item conditions, self-sufficiency filtering (Webb 2010), and significance-corrected discovery would widen the rule tier without touching the statement schema.

• Subgroup discovery and contrast sets. Beam search with WRAcc over discretized conditions — the CN2-SD lineage — would generalize the “what’s different about these rows” question beyond pairs, and contrast sets against natural segments (before/after a detected change, missing vs. present) would give the changes category a follow-up pass.

• Conditional anomalies. A GritBot-style pass would identify values that are rare within a context. It would need validation that accounts for searching across many candidate contexts.

• Multivariate identities. The identity probe is pairwise affine today; y ≈ a·x₁ + b·x₂ and product/ratio forms would cover the total = quantity × price family the screen currently flags as merely predictable.

• Inclusion dependencies. Value containment between columns could identify lookup or enumeration relationships across tables. Multi-table profiling could also compare profiles for drift.

Semantics and time¶

• Holiday calendars. Gap semantics are weekday-aware; real calendars (regional holidays) would stop flagging public holidays as ETL gaps. The current finding says holidays not modeled for exactly this reason.

• Intraday structure. Datetime columns are analyzed at day granularity; sessions, time-of-day gaps, and sub-daily cadences are open.

• Semantic patterns. Roles are behavioral today (identifier-like, categorical). Libraries for recognizing emails, phones, and codes could add semantic labels, with additional maintenance requirements.

Interfaces¶

• Supervised mode. A targets= argument — profile this column against the rest, with the same passes and validation rules — would focus discovery on specified targets.

• More predicate transforms. The rule schema supports the day_index transform; month, weekday, and window transforms would express calendar conditions in rules.

• MCP server. An agent-facing tool wrapper around profile so any MCP client can call the profiler directly, with the report’s token-tiered views doing the truncation.

• Prediction engine. The predictability screen is a purpose-built histogram GBM. An alternative engine or ensemble could produce the same dependencies/predictive statements, with its method and validation recorded in evidence.

Requirements for new passes¶

Any new pass should follow the current implementation requirements: bounded compute disclosed in coverage, statements separated from methods, validation labeled for what it is, and a negative-control test before any new claim ships.