A Yield Gain That Decayed After Handoff

Separate the technical from the organizational, then test in order of cost and speed.

Technical hypotheses, tested first:

• Catalyst change. Overlay the supplier switch date on the yield chart. Compare yield for batches before and after the switch, holding temperature and raw-material lot constant. A 6% price reduction that costs 2.7 points of yield is not a saving - compute the crossover: at Line 2 volume, 2.7 points of yield is worth far more than 6% of catalyst cost. This is the single fastest test and the highest-probability cause.
• Temperature excursions. Pull reactor temperature by batch for the post-handoff period. Count batches outside 178–182 °C and correlate with yield. Interlock resets are a documented excursion source - count reset events per week.

Organizational hypotheses:

• No named process owner after handoff: 'notify supervisor' is not a reaction plan, so excursions produce notification and no action.
• Operators bypassing the interlock indicates the control was never accepted as legitimate - either it genuinely slows the batch (a design problem to fix) or it was never explained (a training problem). Determine which.
• The control plan covers one X in a process where the catalyst is clearly a significant input - the plan is incomplete.

Two defects.

1. Contaminated limits. Limits calculated from the pilot period include the improved batches - and possibly special causes from the pilot itself. Limits built from a mixture of two process states are inflated, which is precisely why a 2.7-point drop sat inside the chart without signalling. A chart that cannot detect the failure it was built to catch is worse than no chart, because it creates false assurance.
2. Wrong chart family, possibly. Batch yield as a proportion depends on lot size. If lot sizes are constant, an individuals chart on yield is defensible for n = 1 batches. If lot sizes vary materially, a p-chart (or Laney p′) on proportion nonconforming is correct, because the varying n changes the variance and the limits.

Fix:

• Freeze separate phases: baseline (91.4%) and post-improvement (94.5%). Plot phase limits, not one blended set.
• Confirm stability in the post-improvement phase before recalculating limits; then set limits from a clean, stable, representative period that includes normal catalyst and seasonal variation.
• Do not recompute limits on every shift or after every bad week - that converts the chart into a mirror.
• Add the catalyst lot/supplier as a tracked variable, and an individuals chart on reactor temperature with its own limits, since temperature is the X the plan already claims to control.

Control plan rebuilt around ownership and actionability:

• Every X: specification or control limits, measurement method and frequency, sample size, named owner (a person, not a department), and a reaction plan that states the specific action, who takes it, and within what time. 'Notify supervisor' becomes: 'If temperature > 182 °C, stop charging, verify jacket valve position, notify shift lead within 15 minutes, log in the deviation system.'
• Mistake-proofing over vigilance. If the interlock slows the batch, fix the interlock design or the batch logic - do not rely on operators not bypassing it. Add an audit that counts bypass events as a leading indicator.
• Supplier change control. Catalyst supplier changes route through a requalification protocol with a defined yield check; 6% unit savings cannot be approved without the yield test.
• Visual management and layered audits: yield and temperature posted per shift; supervisor audit weekly, process manager monthly, with findings logged.

Closure criteria: 90 days of sustained performance at or above the improved level with no unexplained special causes, control plan signed by the process owner (not the belt), operators trained and assessed, and finance validating the savings against the agreed method. Then hand off with a 30/60/90-day check-back scheduled - the belt leaves, the accountability does not.