The Two-Cell Pilot

Allocation

• Assign the method to 2 of 6 cells; keep the other 4 as the comparison group rather than relying on a before/after only. This gives a concurrent control, which protects against demand mix and seasonality.
• Cell 3 must be handled explicitly. Its operator already uses a partial version, so it is contaminated. Either (a) exclude Cell 3 from the treatment and comparison sets, or (b) treat Cell 3 as already-treated and use it to check whether the full method beats the partial one. Option (a) is the defensible default, but say so in the design rather than discovering it later.
• Choose the two cells to be similar to the control cells on baseline changeover time and product mix - not the two best cells.
• Randomize the assignment if the cells are genuinely comparable; otherwise match pairs and randomize within pairs, then model cell as a blocking factor.

Sample size

• Treated events: 2 cells × 2 changeovers/week × 3 weeks = 12.
• Baseline: 6 cells × 2/week × 4 weeks of historical data = up to 48 events (or 8 per control cell during the pilot).
• With changeover SD ≈ 8 min, detecting a 20-min reduction at α = 0.05, 80% power needs roughly 3 events per group (n ≈ 2(1.96+0.84)² σ²/Δ² = 2 × 7.84 × 64/400 ≈ 2.5). Twelve treated events is ample.
• Detecting a 5-min reduction would need ~40 per group - so state now that the pilot can only resolve a large effect. Do not later claim a small effect "wasn't significant" as proof of no effect.

What to measure (per changeover, every event)

• Changeover time with an operational definition: last good part of the previous run to first good part of the next run. Not "when the crew started."
• Setup scrap and first-hour OEE - a faster changeover that damages the first 50 parts is not an improvement.
• The 4-minute pre-staging time, recorded separately as internal or external depending on whether the line is running.

The 4-minute argument is the crux, and both supervisors are half right.

• If the 4 minutes are performed while the line is still running (external setup), they are not a cost to changeover at all - the entire point of SMED discipline is moving work from internal to external time. The metric must reflect that.
• If the 4 minutes are on the clock after the line stops, the net gain is (baseline − new changeover) − 4. Payback at a 46 → 30 min change is 16 − 4 = 12 net minutes; at 46 → 42 it is 4 − 4 = 0. The method pays back in the first case and not the second - so define, in advance, the minimum net gain that justifies rollout (e.g., 10 net minutes, about 20% of baseline).

Pre-registered success criteria (write before week 1, circulate to the plant manager)

1. Mean changeover time reduced by ≥ 10 net minutes versus concurrent controls over weeks 2–3 (week 1 is learning-curve and analyzed separately, not discarded).
2. Setup scrap and first-hour OEE not degraded beyond a pre-set limit.
3. 95% CI on the net reduction excludes zero and excludes values below the practical threshold.
4. A standard work document exists and a different shift, not trained by the originating operator, can reproduce the time within ±5 min. If only the inventor can do it, the improvement is not a process improvement.

Contamination controls: announce nothing until baseline is captured; the originating operator must not train night shift during the pilot; record who performed each changeover so operator effects can be separated from method effects; keep a diary of any other changes made during the three weeks.

Respond by converting "shows improvement" into a decision rule now. A vague success bar will be interpreted by whoever is most optimistic at the end of week 3.

Evidence required for rollout

• The four pre-registered criteria are met, with the net-gain CI reported, not just a "before vs after" average.
• Reproducibility demonstrated by a shift/crew other than the source operator and cell.
• The effect was measured against concurrent controls, so a plant-wide volume dip cannot be mistaken for a changeover gain.
• Any labor offset (the 4 minutes) is costed into the business case, not ignored.

Why the timeline is too tight Three weeks of pilot plus verification plus training material plus a second-crew trial does not fit inside "next month." Say so with the schedule: pilot weeks 1–3, analysis week 4, standard work and training package weeks 4–5, second-wave pilot in 2 more cells weeks 6–8, decision at week 9. Six cells in one step offers no fallback if the method fails on a cell with different tooling or product mix.

Recommended path: stagger, don't sweep.

1. Roll out to 2 additional cells first (a replication wave), keeping 2 as controls.
2. Only convert the final 2 cells once the replication wave reproduces the gain.
3. Build in a 30-day post-rollout check per cell against its own baseline.

Judgement point: the plant manager is asking a schedule question with a quality question underneath. Answer the quality question with a decision rule and the schedule question with a staged plan - refusing outright loses the sponsor; agreeing outright makes the pilot decorative.