One-call SPC review

sw.review() answers the question most people actually have: is this process okay? It selects the right control chart for the data, checks the assumptions behind it, runs capability against specification limits if you have them, and returns a single structured verdict.

import shewhart as sw

rv = sw.review(df, value="torque", lsl=9.95, usl=10.05)
rv.ok          # the gate: stable, capable, no failed checks
rv.headline    # "In control: no rule violations on the imr chart. Cpk 1.41 (capable)."
rv.summary()   # the audit text

review() is pure composition: every number comes from the validated chart and capability functions, and rv.chart / rv.capability hold the full underlying results. The selection logic below is a set of documented conventions, not new statistics.

How the chart is selected

Your data	Call	Selected chart
measurements, no subgroups	review(df, value="x")	I-MR
measurements in subgroups of 2-8	review(df, value="x", subgroup="batch")	Xbar-R
measurements in subgroups of 9+	same	Xbar-S
measurements in subgroups of differing sizes	same	Xbar-S (stair-step)
defective units + inspection size	review(df, defectives="rej", size="insp")	np (constant size) or p (varying)
... with over/underdispersion	same	Laney p' (automatic)
defect counts + inspection size	review(df, defects="dents", size="area")	u, or Laney u' (automatic)
defect counts, constant opportunity	review(df, defects="dents")	c

The subgroup-size cutoff follows Montgomery (R loses efficiency above n=8). The Laney switch fires when sigma_z, the variation of the standardized points, reaches 0.667 or 1.5 - Laney's own reading is that values near 1 mean the classic chart was fine. The sigma_z value and the reason for every selection are part of the verdict; nothing is chosen silently.

EWMA and CUSUM are never auto-selected. They have tuning parameters that belong to you, not to a dispatcher; review() flags the conditions that call for a different design (such as autocorrelation) and leaves the choice to you.

The checks

Check	When it runs	Boundary
sample_size	Phase I only	fail below 10 subgroups, warn below 25 (AIAG)
variation	measurements	fail when all values are identical
normality	measurements, n >= 8	warn above the Anderson-Darling 5% critical value
autocorrelation	individuals, n >= 20	warn when |lag-1 r| exceeds 0.5
overdispersion	attribute charts	sigma_z at or beyond 0.667 / 1.5
binary_data	measurements	warn when values are only 0/1
spec_plausibility	with specs	warn when no observation is inside the limits
target_within_specs	with target= and specs	warn when the target lies outside the limits

Checks that did not run are absent from the verdict, and variation, binary_data, and target_within_specs appear only when they trigger. warn never gates; fail does. The check set is open: later versions may add checks, so pin the version where bit-stable gates matter.

Capability doctrine

With specification limits, capability runs only when the chart is in control - indices from an unstable process are not meaningful, so the verdict withholds them (status: "not_assessed", with a reason code) instead of reporting a number with a footnote nobody reads. The judged index is Cpk when available, Ppk on the percentile path; 1.33 and 1.00 are the conventional boundaries for capable / marginal / inadequate. A marginal process gates ok=False deliberately: whoever accepts Cpk 1.1 should do it looking at the confidence interval, not at a green light.

Fit once, monitor forever

# Phase I: fit and freeze
sw.review(df_2025, value="torque").baseline.save("line3.json")

# Phase II, nightly: judge new data against the frozen limits
import sys
rv = sw.review(df_today, value="torque", limits="line3.json")
sys.exit(0 if rv.ok else 1)

In Phase II the chart comes from the baseline, never re-derived from the data - new-window quirks like a high sigma_z are findings (they appear as checks), not grounds to switch charts mid-monitoring. The sample_size check does not run in Phase II: the 25-subgroup guidance is about fitting limits, and yours are frozen.

The verdict as JSON

rv.to_dict() is built for pipelines and agents - see the API page for the frozen schema and Statistics is not a language task for why this exists:

{
  "ok": false,
  "failures": ["out_of_control"],
  "headline": "Out of control: 2 signal(s) on the xbar_r chart. Capability not assessed (not in control).",
  "selection": {"chart": "xbar_r", "reason": "subgroup size 4 (2-8 -> Xbar-R)"},
  "control": {"status": "out_of_control", "signals": [{"rule": "nelson_1", "points": [17], "labels": ["2026-06-12 14:00:00"]}]},
  "capability": {"status": "not_assessed", "reason": "not_in_control"},
  "checks": [{"name": "normality", "status": "pass", "value": 0.31, "threshold": 0.75}],
  "recommendations": [{"code": "investigate_signals", "message": "...", "call": null}]
}