The Prototype That Cannot Be Built Twice: Design Transfer Is a Statistics Problem, Not a Paperwork Problem

Author: Aravind Venkatesh
Manufacturing Engineering Lead
Expertise: New Product Introduction
Service: New Product Introduction
Sector: Diagnostic Devices
Therapeutic Devices
Drug Delivery
Topic: Manufacturing & Production
Development Strategy & Process
Published: 20 April 2026
7 min read

TL;DR

Design transfer fails when a design that works at n=10, built by the engineers who designed it, meets a process that must work at n=100,000, run by operators who did not.
21 CFR 820.30(h) requires that the design be correctly translated into production specifications. The word doing the work is correctly — meaning the specification captures what actually makes the device function, not just what the drawing says.
Process validation (IQ/OQ/PQ) is a statistical claim about a process distribution, not a demonstration that three batches passed. If you cannot state the process capability, you have not validated anything.
Tolerance stack-up, not component tolerance, is what breaks assemblies at volume. A stack that is acceptable at worst case on paper is often unacceptable in practice because the assumed distributions are wrong.
Bringing manufacturing engineering in at design transfer is the single most expensive scheduling decision available to a device programme. The correct entry point is detailed design.

There is a specific moment in most device programmes where confidence collapses. The design has been verified. The clinical feedback is good. The regulatory pathway is mapped. And then the first production-representative build comes back with a 62% yield, and nobody can explain why.

This is not bad luck. It is the predictable consequence of treating design transfer as a documentation handover rather than as the point where a design meets statistics for the first time.

A Prototype Proves Feasibility. It Says Almost Nothing About Manufacturability.

An engineering prototype is built under conditions that will never recur: unlimited time per unit, the designer's own hands, selection of the best components from the bin, and immediate rework of anything that does not fit. Every one of those conditions is a hidden process parameter, and none of them appear in the drawing package.

The classic examples are mundane and expensive:

A part CNC-machined for prototyping cannot be injection moulded as drawn. Wall thickness transitions produce sink marks. Missing draft prevents ejection. Gate location drives a weld line straight through a sealing face. The redesign required is not cosmetic; it changes the part, and therefore may change verification.
An adhesive bond that "works" was cured by an engineer who happened to hold the parts for forty seconds. Surface energy, dispense volume, open time, fixture pressure, and cure profile were never specified because they were never conscious decisions.
A press-fit that assembles beautifully at 22 °C in an air-conditioned lab behaves differently in a Chennai production hall in May, because both the polymer and the operator's grip strength have changed.

Design transfer, done properly, is the activity that makes these implicit parameters explicit, bounded, and controlled.

Process Validation Is a Statement About a Distribution

Installation, operational, and performance qualification are widely executed as a documentation exercise: run three batches, confirm they pass, sign the report. That satisfies a checklist and validates nothing.

The substantive question is: what is the distribution of the output, and how much of it sits inside specification?

That question is answered by process capability. For a process with a two-sided specification:

Cp compares the specification width to the process spread — how tight the process is, irrespective of where it sits.
Cpk additionally accounts for how far the process mean has drifted from the specification centre.

A process with Cp = 2.0 and Cpk = 0.8 is a precise process that is badly centred, and it will produce defects. A process with Cpk ≥ 1.33 is conventionally treated as capable for non-critical characteristics; critical-to-quality characteristics on Class II and Class III devices are frequently held to ≥ 1.67.

The practical implication: three passing batches with no capability data is not evidence of a validated process. It is three samples from an unknown distribution. If the process is running at Cpk = 0.9, three batches will very often all pass, and the fourth will not.

Operational qualification is where this gets decided. OQ should deliberately run the process at the edges of its parameter ranges — the worst-case combinations — to establish where the process window actually is. A DOE across the two or three parameters that drive the critical output typically costs less than one failed PPQ and produces vastly more information.

Tolerance Stack-Up Is Where Assemblies Actually Fail

Component tolerances are usually fine. Assemblies fail because tolerances accumulate.

Two methods, two very different answers:

Worst-case stack-up sums the extremes. It guarantees fit but is punishingly conservative, and often drives component tolerances so tight that unit cost becomes uncompetitive.
Statistical (RSS) stack-up takes the root sum of squares of the individual tolerances, on the assumption that component dimensions are independent and approximately normal. It yields a far more realistic stack — provided the assumptions hold.

They frequently do not hold. Two parts from the same multi-cavity mould are not independent; cavity-to-cavity variation is systematic. A supplier who is 100% inspecting and sorting to specification does not deliver a normal distribution — they deliver a truncated one, sometimes bimodal if they are also shipping rework. Applying RSS to a sorted, correlated population produces a confident number that is wrong.

The discipline here is unglamorous: know which cavity each part came from, know whether your supplier sorts, and validate the stack empirically on production-representative parts before you commit to tooling.

Design Validation Requires Real Production Units, and That Constrains the Schedule

FDA design control regulation requires validation on initial production units, lots, or batches, or their equivalents. This is not a formality. It is the requirement that forces the schedule to sequence correctly:

Design outputs frozen
Production tooling and processes established
Process validation executed
Initial production units built under the validated process
Design validation performed on those units

Teams routinely try to run step 5 in parallel with step 2, using hand-built units, and then discover that the moulded part differs from the machined part in a way that matters — different surface finish changing adhesion, different crystallinity changing stiffness, different residual stress changing dimensional stability after sterilisation.

Where NPI Programmes Commonly Break Down

Manufacturing engineering arrives at design transfer. By then, every decision that determines manufacturability has already been made. Manufacturing engineers belong in detailed design, where their input still costs nothing.

PFMEA treated as a deliverable rather than an analysis. A PFMEA rushed to close a phase gate lists obvious failure modes with optimistic detection ratings and produces no actions. A useful PFMEA is uncomfortable to write, because it names the steps where the process depends on operator skill and nobody wants to admit that.

No distinction between critical and non-critical characteristics. If every dimension on the drawing is treated as equally important, inspection cost explodes and attention is spread evenly across features that do not matter and features that do. Identify the characteristics that drive function, safety, and fit; control those tightly and relax the rest.

Supplier process changes that are never communicated. A supplier who improves their own process — new resin lot, new machine, new plating line — has changed your product. Purchasing controls under 21 CFR 820.50 and ISO 13485 clause 7.4 exist to make change notification contractual, and they only work if the quality agreement is specific about what constitutes a notifiable change.

Assuming yield improves automatically with volume. It improves with learning, and learning requires that failures be characterised rather than reworked. A line that scraps and moves on generates units. A line that performs failure analysis on scrap generates process knowledge.

The Localisation Dimension

For India-based manufacturers, design transfer carries an additional strategic weight. India still imports an estimated 70–80% of its medical devices, with the government's Scheme for Strengthening Medical Device Industry, launched in November 2024 with a ₹500 crore outlay, explicitly targeting key components and accessories. Under the Production Linked Incentive scheme, 22 greenfield manufacturing projects have been commissioned and production has started on more than 55 device types, with cumulative eligible sales of approximately ₹12,344 crore and exports of about ₹5,869 crore as of September 2025.

The bottleneck in that transition is rarely capital and rarely demand. It is process capability — the ability to demonstrate, with data, that a domestically manufactured device holds specification across lots, shifts, and seasons. That is precisely what a properly executed design transfer produces.

The Framing That Helps

Design transfer is the point at which a device stops being a set of intentions and becomes a repeatable physical process with a measurable distribution. Everything before it is a claim. Everything after it is evidence.

Teams that internalise this stop asking "did the batch pass?" and start asking "what is the capability, where is the mean, and what makes it drift?" Those are answerable questions, and the answers survive scale.

RhythmRx supports new product introduction from detailed design through process validation and production ramp, with manufacturing engineering engaged before design freeze rather than after it.

Aravind Venkatesh is Manufacturing Engineering Lead at RhythmRx, working on design transfer, process validation and tooling qualification for high-volume device assemblies.

Sources

21 CFR 820.30(g) and 820.30(h) — design validation and design transfer; 21 CFR 820.50 — purchasing controls; 21 CFR 820.75 — process validation.
ISO 13485:2016 clauses 7.3.8 (design transfer), 7.4 (purchasing), 7.5.6 (validation of processes).
Ministry of Chemicals and Fertilizers / Department of Pharmaceuticals — Production Linked Incentive Scheme for Medical Devices progress data, September 2025; Scheme for Strengthening Medical Device Industry, November 2024.
Invest India / IBEF — India medical device import dependence estimates, 2025.