0036 — The chart draws every rule; the plant chooses which raise a hold¶

Status: accepted (2026-09-22, by the maintainer)
Date: 2026-09-22
Deciders: maintainer

Context¶

src/fsmes/services/spc.py runs four Western Electric rules over every characteristic's readings, on the write path, so a rule that trips is acted on whether or not anybody has the chart open. All four have always run and all four have always raised a quality hold — a non-conformance a person works through review and a disposition, which is decision 0027.

Decision 0035 placed "SPC rules 1–4" in tier three — what the product owns — with the note "a rule a plant can switch off is a chart that lies." That reasoning is about the chart, and it holds: a chart drawn with rule 4 switched off, presented as an SPC chart, does not say what it appears to say.

The configuration audit of 2026-09-21 (the list, row Q5) found that 0035 had answered a question it was not asked. Whether a plant may choose which rules raise a hold is a different question from which rules are drawn, and 0035 never asked it. Q5 is the only one of that audit's seventy-five candidates listed as general in quality and the only one it records as blocked on a decision rather than as work: until this is decided, nothing about it can be built.

The rule windows — 2-of-3, 4-of-5, 8-in-a-row — are not in question and stay the product's. A plant that changed them would publish SpcSignal.rule = 3 while meaning something nobody else means by rule 3, which is the tier-three side of 0035's own test. RULE_WINDOW = {1: 1, 2: 3, 3: 5, 4: 8} (services/spc.py) is fixed.

What is at stake in practice: a plant inspecting frequently on a characteristic that drifts slowly gets rule 4 — eight readings in a row on one side of centre — several times a shift. Each one opens a non-conformance that somebody must review, disposition and close. A plant in that position today has two options, and both are bad: work a queue of holds nobody thinks are findings, or stop recording checks.

Options considered¶

Option	For	Against
Draw and record all four always; let the plant choose which raise a hold	The chart never lies — every rule is drawn, every firing is recorded as an `SpcSignal` row, and the verdict still says out of control. What the plant chooses is who gets called, which is a plant's own business and differs honestly between an aerospace cell and a bottling line	Two plants' hold queues are no longer comparable. A plant can quieten a rule and then forget it did
Leave it: all four rules always raise a hold	Nothing to build, nothing to explain, every plant's holds mean the same thing	The plant above still has to choose between a queue it ignores and not recording checks, and a hold queue nobody works is decision 0027 already broken, quietly
Let the plant switch rules off entirely — not drawn, not recorded	One switch rather than two concepts	This is exactly what 0035 §3 refused, and for a reason that has not changed: a chart drawn without rule 4 and labelled an SPC chart is a chart that lies
A severity per rule instead — every rule holds, at the plant's own grade	Nothing disappears; triage sorts itself out	It does not solve the problem, which is the number of records somebody must work, not what they are called. That is Q2, and it is a different change

Decision¶

Every Western Electric rule is evaluated, drawn and recorded on every plant. Which of them raise a quality hold is the plant's, and defaults to all four.

The four rules, their windows and their numbering stay the product's and are not configurable by anything. A firing is always written as an SpcSignal row, always drawn on the chart, and always counted in the verdict: a process that trips rule 3 is out of control on every plant, and capability stays withheld until it is settled. Decision 0027 is unchanged for every rule a plant holds on.

What a plant chooses is [quality] hold_rules in its pack — a list of rule numbers from 1 to 4, compiling to MES_QUALITY_HOLD_RULES. Leaving it out is all four, which is what this product has always done, so a plant that configures nothing behaves exactly as it does now. The key may only ever narrow the set {1, 2, 3, 4}; fsmes pack check refuses anything else, offline, naming the four rules.

An empty list is a real answer and is never silent. A plant may record and draw every rule and raise a hold on none — running SPC as an observation is a plant, not a mistake. It is said out loud rather than allowed to be discovered: the chart payload carries rules and hold_rules on every response, and the SPC screen prints which rules this plant holds on and which are drawn and recorded and raise no hold, on every chart, including the one where all four are held. A signal the plant does not hold on is returned with held: false, which the screen renders as no hold — this plant does not hold on this rule, so a firing that opened nothing is explained rather than noticed.

Scott's words, 2026-09-22: "Draw all four; the plant chooses which hold."

This amends 0035 §3 in place rather than superseding it: the sentence "a rule a plant can switch off is a chart that lies" stays true and is scoped to the chart, which is what it was always about.

Consequences¶

Easier: a plant whose process trips a warning rule several times a shift can keep recording checks and keep the chart honest without working a hold queue it does not believe in. Easier: the audit's Q5 stops blocking, and the rest of quality's configuration section can be built.

Harder: two plants' hold counts are no longer straightforwardly comparable, so a fleet view that counts open non-conformances is counting something each plant defined. The chart is comparable and the SpcSignal rows are comparable; the holds are not, and anything that compares them across plants has to say so.

Harder: a plant can quieten a rule and forget. The mitigation is that nothing hides it — the chart says which rules hold on every load, the Configuration page lists the key and names the pack it lives in, and the rule still fires, is still recorded and still moves the verdict.

To revisit: if a plant turns out to want a severity per rule rather than a hold or nothing — the audit's Q2, spc_major_rules — that is a different key and its own change. This decision deliberately does not pre-empt it.

House rules touched¶

Never invent production and unknown is not zero (1 and 2): nothing here changes a reading, a signal or a verdict. A rule that fires is recorded and drawn on every plant, so the evidence is identical whatever the plant holds on; only who is called changes, and the screen says which rules those are. A firing with no hold is labelled with why there is no hold rather than left as an absence a reader has to interpret.

Config, not code, at plant boundaries (4): which findings are worth somebody's morning is a plant's own judgment, and it was a literal — four for loops with no number in them — until this decision.