API and grammar

This page is the complete public surface of shewhart. It is deliberately small, and it is frozen: from 0.1 on, no public name, signature, default value, or string alias is removed or changes meaning within a major version. Renames, if ever, keep permanent aliases. Defaults never change silently, because changed defaults change numbers, and this is an audit tool.

One import, everywhere:

import shewhart as sw

Control charts

Call	Chart	Notes
`sw.imr(data, value=, rules=, limits=)`	Individuals & moving range	the default chart for n=1 data
`sw.xbar_r(df, value=, subgroup=, rules=, limits=)`	Xbar-R	subgrouped, limits from R-bar
`sw.xbar_s(df, value=, subgroup=, rules=, limits=)`	Xbar-S	subgrouped, limits from S-bar
`sw.p_chart(df, defectives=, size=, rules=, limits=)`	p	varying sizes, stair-step limits
`sw.np_chart(df, defectives=, size=, rules=, limits=)`	np	constant size
`sw.c_chart(data, defects=, rules=, limits=)`	c	counts per unit
`sw.u_chart(df, defects=, size=, rules=, limits=)`	u	rates, varying sizes
`sw.ewma(data, value=, lam=, k=, center=, sigma=, asymptotic=, limits=)`	EWMA	exact limits by default
`sw.run_chart(data, value=, alpha=)`	run chart	four runs tests
`sw.pareto(data, by=, weights=)`	Pareto analysis	counts or weighted (e.g. cost)
`sw.cusum(data, value=, k=, h=, center=, sigma=, limits=)`	tabular CUSUM	decision interval, no run rules
`sw.laney_p(df, defectives=, size=, rules=, limits=)`	Laney p'	overdispersion-robust, reports sigma_z
`sw.laney_u(df, defects=, size=, rules=, limits=)`	Laney u'	overdispersion-robust rates

Analyses

Call	Analysis	Notes
`sw.review(data, value=\\|defectives=\\|defects=, subgroup=, size=, lsl=, usl=, target=, rules=, limits=)`	one-call review	selects the chart, checks assumptions, returns a Review
`sw.capability(data, value=, lsl=, usl=, target=, subgroup=, confidence=, dist=, transform=)`	Cp/Cpk/Pp/Ppk/Cpm with confidence intervals	non-normal via `dist=` (percentile method) or `transform="boxcox"`
`sw.gauge_rr(...)`	ANOVA gauge R&R (AIAG)	planned (0.2)
`sw.type1(...)`	Type 1 gauge study (Cg/Cgk)	planned (0.2)
`sw.attribute_agreement(...)`	attribute agreement (kappa)	planned (0.2)
`sw.tolerance_interval(data, value=, coverage=, confidence=, method=)`	tolerance intervals	normal (Howe k2) and nonparametric (Wilks)
`sw.screen(...)`	fleet screening over many characteristics	planned (0.3)
`sw.monitor(...)`	drift monitoring with chart semantics	planned (0.3)

Planned names are part of the frozen grammar: they will appear exactly as written here.

Everything returns a Result

r.ok          # bool: no signals; use as exit code
r.stats       # named scalars: centers, limits, indices
r.table       # tidy per-point DataFrame incl. signal flags
r.signals     # tuple of structured violations (rule, chart, points, note)
r.meta        # provenance: n, version, input hash, timestamp, source
r.baseline    # frozen parameters; .save(path) / Baseline.load(path)
r.summary()   # plain-text verdict
r.plot(ax=None)
r.to_html(path=None, title=None)
r.to_dict()   # JSON-safe, integer-versioned schema

Reports over several analyses:

sw.report([r1, r2, r3], "weekly.html", title="Line 3 weekly")

The review verdict

sw.review() returns a Review, not a Result: it composes a chart, the assumption checks, and (with specification limits) a capability study into one gate. The underlying Result objects stay reachable for drill-down:

rv = sw.review(df, value="torque", lsl=9.95, usl=10.05)
rv.ok            # True iff failures is empty - the gate
rv.failures      # machine-readable causes, e.g. ("out_of_control",)
rv.headline      # a short deterministic verdict line
rv.chart         # the underlying chart Result
rv.capability    # the capability Result, or None
rv.checks        # tuple of Check(name, status, value, threshold, note)
rv.baseline      # passthrough: rv.baseline.save("line3.json")
rv.summary(); rv.plot(); rv.to_html(); rv.to_dict()

rv.to_dict() is the JSON verdict (schema 1): ok, failures, headline, params (the call echoed, with limits as "fitted" or "frozen"), selection (chart and reason), control (status, stats, signals with index labels), capability (always present; status "not_assessed" carries a reason code such as no_spec_limits, not_in_control, unstable), checks, recommendations (code, message, call), baseline, meta. Numeric fields are finite or null - the JSON never contains NaN.

Two covenant rules for consumers:

The check set and every enum are open: minor versions may add checks (which can tighten the gate) and enum values. Treat unknown values conservatively, and pin the library version where bit-stable gates matter.
Fields are never removed or renamed, and failures is empty exactly when ok is true.

String aliases (stable forever)

Rule sets: "nelson", "western_electric", "none"
Sigma estimators (capability/charts vocabulary follows the AIAG manuals and the major commercial packages): average_mr, pooled, rbar, sbar
Registry: sw.chart("imr", ...) dispatches by alias and is the entry point reserved for third-party plugins (shewhart-<name> packages).

Conventions

Data is always the first positional argument; every column or option is a keyword with a full word (value=, subgroup=, lsl=, rules=).
On a DatetimeIndex, subgroup= also accepts a time window: a fixed one such as "15min" or "1h", or a calendar one such as "W", "ME", or "QE". Calendar windows produce subgroups of differing sizes, which xbar_s and review() handle with stair-step limits.
Sigma estimators are selected by name: sw.imr(..., method="average_mr" | "median_mr"), sw.xbar_r(..., method="rbar"), sw.xbar_s(..., method="sbar" | "pooled"). Estimator aliases, like all string aliases, never change meaning.
No **kwargs anywhere public.
Errors teach: every exception ends with a corrected, runnable example.