Last updated: 2026-09-04 · 5 min read
What Cpk Number Should You Demand? — Process Capability as the Language of Mold Acceptance
Cpk is the process capability index — the number that answers the question “in tolerance?” was never designed to answer: not whether the parts you measured fit the drawing, but whether the process that made them can keep fitting the drawing tomorrow, next week and next million shots. Arithmetic is simple — Cpk compares the process distribution to the nearest specification limit, counting how many standard deviations fit between the process center and that limit.
Key Facts
<>
Flash — Thin excess material that escapes between the mold parting line or around ejector pins. Cause: insufficient clamp force, worn parting line, excessive injection pressure, or contaminated mold surface. Flash thickness: 0.02–0.1mm. Requires secondary trimming to remove. The related guide covers adjacent topics.
What Do Cpk Values Actually Mean?
Five levels, each with its own economics:
| Cpk | Statistical Meaning | Typical Use |
|---|---|---|
| Below 1.00 | Under 3σ from the limit | Not acceptable; improve |
| 1.00–1.33 | 3–4σ; predictable scrap | Tolerated on non-critical only |
| 1.33 | 4σ; ~64 ppm tail risk | Standard acceptance threshold |
| 1.67 | 5σ; tightened risk | Critical characteristics |
| 2.00 | 6σ; six-sigma quality | Rare, high-reliability duty |
The ladder reads as escalating distance between the process center and the specification cliff. Below 1.00, the process distribution already crowds its limit — parts will fall out of tolerance at a rate you can watch, and accepting the tool simply schedules the argument for later. The 1.00–1.33 band is the negotiation zone: sustainable on a loose, non-functional dimension, but on anything functional the band is where improvement actions live.
The 1.33 threshold is the industry’s working compromise — four sigma of protection, tail-risk in tens of parts per million, robust enough to survive normal drift. The 1.67 level costs real effort — tighter process and tighter tool — and earns its keep on dimensions carrying function or safety: a seal face, a bearing bore, a snap engagement.
The 2.00 row calibrates expectations: six sigma on an injection-molded dimension demands variation control of a different order, and demanding it by reflex usually means tolerance was drawn tighter than needed.
How Do You Write Cpk Into Mold Acceptance?
Four steps, in order:
- Choose the characteristics worth capability money. Cpk on every dimension of a drawing is spendable but rarely rational: capability studies concentrate on the characteristics whose failure carries consequence — functional fits, sealing surfaces, engagement features, critical wall locations. A working selection is a dozen or fewer dimensions per tool, each tagged by consequence: which are functional (higher threshold), which are cosmetic-adjacent (standard threshold), which are reference-only (no study at all). This selection is a joint decision in the trial preparation phase, written into the protocol.
- Collect the data under production conditions, honestly sampled. A capability number is only as honest as its sampling conditions. Run the process to steady state before sampling — mold at temperature, material conditioned, parameters at the documented setpoints. Sample enough: the working minimum is around thirty parts per characteristic for a preliminary number, with fifty or more preferred for anything consequential, collected across a time span that lets normal drift occur. Respect subgroups and sequence — consecutive shots grouped, groups spread across the run — so the study can distinguish short-term variation from drift.
- Read the study in order: stability first, then capability. Before any Cpk is quoted, the data goes on a control chart, and a process that shows special-cause signals — points outside limits, runs, trends — has no meaningful capability number, only instability wearing arithmetic. Only a stable study earns the Cpk computation. Cp and Cpk read together tell the story — a high Cp with low Cpk says the variation is fine but the center drifted, fixable by process adjustment; low on both says the variation itself exceeds the tolerance, a tool-and-tolerance conversation.
- Agree the thresholds per characteristic, in writing, before the trial. The acceptance protocol is where Cpk becomes contract: for each studied characteristic, the document names the drawing feature, the sample plan, the stability requirement, the minimum Cpk by consequence class, and the remedy path if a value lands short. Writing thresholds per characteristic prevents the two classic deadlocks: the global “we require 1.67 on everything” that prices out rational programs, and the post-hoc “1.1 is fine for this one” that converts a study into a negotiation.
Parting line — The line where the two mold halves (A-side and B-side) meet. Parting line location affects: gate placement, ejector pin marks, witness lines, and cosmetic appearance. A well-designed parting line minimizes visible witness marks. The related guide covers adjacent topics.
A threshold agreed before data exists is a standard; one negotiated after is a mood. The protocol ends the trial with both parties reading the same arithmetic under the same rules — the only environment where capability numbers build trust.
Frequently Asked Questions
How many parts does a valid Cpk study need?
The working minimum is around thirty parts per characteristic, fifty or more for anything that will anchor an acceptance decision, and the honest answer includes two qualifiers beyond the count. The parts must be sampled across a production-representative window — not a single burst of consecutive shots — because the study needs to see the process’s normal drift, not just its momentary behavior.
And the parts must come from a stable process, because stability is what makes thirty samples a window into a million shots: the index extrapolates from the sample’s variation to the process’s future, and extrapolation is only lawful when the process is in statistical control.
Why not just verify that sample parts are in tolerance?
Because “in tolerance” describes the parts you measured and says nothing about the parts you didn’t — and capability is precisely the bridge from one to the other. A dimensional report on ten good parts proves the tool can make good parts at least once; it does not tell you how much of the process’s natural variation remains between the process center and the specification limit.
The distinction has money attached: a process centered beautifully but sitting 1.2 sigma from its limits can produce ten perfect samples and then scrap at a rate that dwarfs any tooling saving, while the same tool with the same tolerance at 4 sigma produces good parts almost by constitution. In-tolerance verification remains necessary — it confirms the tool is built right — but it is not sufficient, because acceptance is a decision about the future.
Can a Cpk number be gamed?
Yes — which is why the protocol exists and why the number is never read alone. The classic manipulations are sample curation, window selection, condition tuning, and stretching the tolerance before computing — all of which produce beautiful numbers that the production floor cannot reproduce.
The antidotes are procedural, not statistical: the sampling plan fixed in advance; the control chart read before the index; the conditions documented as the production setpoints with the setup sheet attached; and the threshold agreement that removes after-the-fact renegotiation. One more check is cultural and free: ask to see the raw data, not just the index — a shop running honest studies produces the charts instantly, while a shop producing theater needs a day to generate what a real study would have generated anyway.
Key Takeaways
<>
Cpk is the most honest sentence in the acceptance conversation when the conditions are right, and a statistic looking for a lawyer when they are not.
The protocol — characteristics chosen, data collected honestly, stability verified, thresholds agreed — is what makes the difference.
Sources: