Two Points After Lunch

In control? No. Two Xbar points beyond the UCL are a signal by any rule set, regardless of how few points are affected.

Diagnosis of the pattern

• Xbar out, R in control → a shift in location, not an increase in dispersion. The fill mean moved; the within-subgroup spread did not. That narrows the cause to something acting on the level: a setpoint change, a product or lot change, temperature/CO₂ after an idle period, a valve or pump restart characteristic, or a sensor zero shift.
• Both signals in the same 20-minute post-lunch window is the strongest clue. Random common cause would scatter across the day. A recurring, position-specific event is a predictable input change - exactly what an idle-restart startup sequence produces.
• Do not be reassured by Cpk = 1.35. Capability indices computed on an unstable process describe a mixture, not a process. Report the chart first; capability only after stability is demonstrated.

Correct response at the signal

1. Contain: segregate product from the affected subgroups back to the last in-control point and evaluate against specification.
2. Do not adjust the filler head. Record the current setpoint, then investigate the input side: was the line idle? Was there a product or syrup lot change at the break? What does the first subgroup after restart look like on previous days?
3. Verify with data: pull the post-lunch subgroups for the last 20 days and compare their means with the mid-morning subgroups. If the post-lunch mean is systematically high, the cause is in the restart, not in chance.
4. Only then change something, one variable at a time, and confirm with a short-run check.

The operator's response: adjusting on every signal without diagnosis is tampering.

• Two out-of-control points on an Xbar chart represent a real level shift; adjusting is not inherently wrong. But dialing the head down and re-checking is an unblinded, unrecorded adjustment loop: with no setpoint log, no one can tell whether the next subgroup is high because of the environment or because the previous adjustment was too large.
• Adjust-and-check cycles typically produce more variation - the classic funnel-rule experiment. Watch for a subsequent pattern of alternating points or increased R-chart variation, which would be the adjustment itself creating signals.
• The absence of a documented setpoint makes the process non-reproducible: every shift inherits whatever the last operator dialed in.

The supervisor's action: taping over the alarm destroys the control.

• The alarm is a detection device. Disabling it converts a detectable special cause into an undetected one and converts the control chart into decoration.
• "It kept going off" is itself evidence: either the control limits are wrong (computed from an unstable period, or inherited from a different product), the sampling is mismatched to the process, or a real recurring problem was never fixed. All three are fixable; none are fixed by tape.
• This is a governance failure, not an operator failure: if the alarm fires and no one has authority or time to respond, the reaction plan was never real.

Root cause of both behaviours: a control plan that specified what to measure but not who responds, how fast, and with what authority.

Reaction plan (per signal type)

Signal	First responder, within	Action	Escalate if
1 point beyond Xbar limit (mean shift)	Operator, 15 min	Stop adjusting; record setpoint; contain product since last good subgroup; check idle/lot/temperature inputs; notify line lead	Cause not identified in 30 min or a second signal within 24 h
Trend / 7 points one side / 2 of 3 beyond 2σ	Operator, end of shift	Log the pattern; check for drift in setpoint or material; line lead reviews next shift	Pattern persists 2 shifts
R chart out of control	Line lead, immediately	Check for mixed product, measurement method, or intermittent mechanical fault	Any repeat
Alarm disabled or bypassed by anyone	-	Prohibited. Immediate escalation to the process owner	Always

Other required plan changes

• Sampling frequency: 30-minute sampling with weekly sign-off is mismatched - review the chart daily at shift handover, weekly at the process-owner level.
• Filler-head setpoint log with time, value, and who changed it. Non-negotiable.
• Startup standard work for the post-lunch restart, with a first-subgroup check before releasing product.
• Control limits recalculated from a verified stable period and frozen; documented in the plan with the date and the basis.
• Named owner per control item and a revision date; the plan is a live document.

Verification after 30 days

1. Chart review: 30 days of data, no recurrence of the post-lunch mean shift, R chart in control, limits unchanged.
2. Process audit by someone outside the area: is the alarm functional, is the setpoint log complete for every shift, does the operator on nights describe the reaction plan correctly when asked?
3. Outcome metrics: fill-volume defect rate and giveaway (overfill cost) - a control plan that holds the mean but costs 0.5 mL of giveaway per bottle is not a success.
4. Capability re-check on stable data only, then publish Cpk/Ppk alongside the chart, never alone.
5. Feedback loop: any signal without a documented response in the log is a finding against the plan, not the operator.