The photos that didn't match the sample
The photos landed at 7:40 on a Tuesday morning in March 2024 — fourteen of them, in no particular order, from a warehouse in Rotterdam.
The first one was just a housing. Face up on a pallet, under bad fluorescent light. Same mold as the unit on my desk. Same texture. Same color chip, or close enough that I wouldn't have caught it in person.
Except there was a line of flashing running down the left side, and the front bezel was a warmer gray than the one I'd signed off in December.
I zoomed in on the LED window. Stress whitening around the screw boss. That's not shipping damage. That's a tool running cold, or a molder pushing cycle time to hit a ship date.
Twelve hundred units. First production run of our outdoor CPE line. And I had exactly one sample — sitting on my desk — and no sealed reference unit at the factory.
That's when I knew it was going to be a bad week.
How we got there
Context, because the housing turned out to be the least of it.
We'd spent about six months sourcing an outdoor CPE router / WiFi bridge for an integrator doing hospitality and campus work across the Benelux and Germany. Their requirement list was short and specific: outdoor-grade enclosure, PoE in, 4G LTE fallback with an unlocked modem, a real 300 Mbps-class downlink, dual-band WiFi bridging, and actual authentication controls — because these were going onto networks that touched guest traffic.
We shortlisted three factories. One came in 9% under the next quote. On a 1,200-unit order that was roughly $8,600 (a number I would come to resent).
Nine percent is not a red flag. That's the problem. Nine percent is plausible. It could be a better tool, a bigger press, a cheaper freight lane, a thinner margin because they wanted the reference account. It could also be a gasket compound, a firmware engineer, and an afternoon of testing somebody decided not to do.
At LinkScope Sourcing 20 we had a rule about exactly this: don't buy the quote, buy the acceptance criteria. We had the rule. We didn't have the paperwork.
I approved the sample in December. I approved it on a Thursday because I was on a plane Friday, and I did the review by eye and by hand instead of against a written checklist. Over 2024 I rejected 29% of first articles across our category. That Thursday, I was the reason one of them got through.
Problem one: the gasket nobody wrote down
The datasheet said IP65. That's a real rating with a real definition — IEC 60529, dust-tight plus protection against water jets. It isn't a vibe. It's a test.
The sample used a continuous molded gasket with a compression lip. One piece, no joint.
Production used a spliced gasket with a butt joint, and the joint sat at the bottom edge of the door — the one place water actually pools.
The vendor's position was that it was the same material. It was. Same compound, same durometer. It just wasn't the same part. Material compliance and design compliance aren't the same thing, and I'd written a spec that only covered the first one.
We ran a jet test on three units from the first hundred off the line. Two had visible ingress at the joint after four minutes on our bench. The third was fine, which is somehow worse — it means the defect is intermittent, which means it shows up in year two in the field instead of on your bench. (I've stopped being reassured by the good unit in the sample.)
Second thing we found while the door was open: the breather membrane was gone. The sample had one. Production didn't. That's a condensation problem — the enclosure pulls in humid air overnight and it has nowhere to go. You won't see it in a week. You'll see it in eighteen months, when someone's outdoor wifi CPE router stops passing traffic after a cold snap.
Problem two: the admin panel
Our test engineer opened the web UI on a production unit and asked me to come look.
Same admin credentials on all 1,200 units. WPS enabled by default. Remote management reachable from the WAN side, enabled by default, pointing at the factory's own DDNS.
Nobody at the factory thought this was a defect. From their side it was a default configuration. From mine it was a fleet of devices about to be handed to a hotel, where the person doing the install is a facilities tech who will never change a password in his life.
Here's the thing about CPE router authentication security: it isn't a feature you list on a datasheet next to antenna gain. It's a set of defaults, and defaults are where most deployments actually live. If the device ships open and the buyer has to close it, you've shipped the risk to the buyer.
It's also not just my opinion anymore. ETSI EN 303 645 has been out since 2020 and treats universal default passwords as a baseline failure. The UK's PSTI regime became enforceable in April 2024 (guidance still being revised, as of early 2025) and bans them outright for consumer connectable products. California's SB-327 has required unique passwords and reasonable security features since 2020. Whether a specific outdoor CPE falls inside a specific regime depends on the market and the product class — read the current text, don't take my summary of it. But the direction is one-way, and a distributor who ships from one country into another carries the strictest rule along with the box.
We didn't ship those units. Not with that firmware.
Problem three: what 300 Mbps and Cat6 actually mean
This one wasn't the factory's fault. It was ours.
The listing copy for the 4G CPE router (unlocked, 300 Mbps class) said 300 Mbps. That number is real — it's the LTE Category 6 peak downlink figure, and it assumes carrier aggregation, 256QAM, and a signal environment that exists mostly in a lab and occasionally on a rooftop with direct line of sight.
And yes, the Cat6 in a Cat6 CPE router part number refers to LTE Category 6, not Cat6 cabling. I've had that argument with a distributor twice. The first time I assumed he was joking.
A reseller took our number, put 'up to 300 Mbps' on a slide, and a customer read it as 300 Mbps everywhere on the property. Nobody lied. Nobody in that chain was being dishonest. The claim just travelled further than the qualification did.
Whose fault was that? Mine.
Per FTC guidance on advertising claims (ftc.gov), a performance claim needs a reasonable basis before it runs, and any qualifying disclosure has to be clear and conspicuous — a footnote under a headline doesn't fix the headline. That's the standard I now apply to our own datasheets before they leave the building.
So we rewrote the line. Peak downlink 300 Mbps (LTE Cat 6, 2CA, 256QAM, lab conditions). Typical throughput depends on band, load, and signal. Boring. Correct.
What it cost
We held 380 units — just under a third of the run — and released 820 after the firmware rebuild.
The factory replaced the gaskets and reflashed at their own cost, about $9,400 on their side. That isn't a win for us. A factory that eats $9,400 to fix your order will find a different $9,400 to save on the next one, and you'll be back here in eight months.
Our side:
- $7,100 in expedited air freight to get the 820 released units to Rotterdam inside the integrator's install window
- $1,850 for a third-party re-test of the enclosure, because I wasn't going to take the factory's word twice
- About $900 in internal hours spread across three weeks
- And 180 units pushed to the next quarter when 18 sites slipped — roughly $5,700 in margin I never got back
Call it $15,550, against the $8,600 we thought we'd saved by taking the 9% quote.
Roughly $1.80 out the door for every dollar we saved on paper. And the factory looked at the same math and drew the opposite conclusion — that our spec was unreasonable — which is why they're off our approved list.
What I changed
I'm not going to pretend this had a tidy ending. The integrator moved phase two to a different supplier. It took most of a year to get back into that conversation.
What did change is the process. Five things, all boring, all cheap:
1. A sealed golden sample lives at the factory. Signed across the seal by me and by their QA lead. Any production revision means a new sample and a new signature. We keep the twin, they keep the reference.
2. Firmware is a spec line item with a hash. If the build on the dock doesn't match the approved hash, nothing gets unloaded. This surfaced a problem we didn't know we had — production was running a November build against a September approval, and the changelog between them was a verbal description sent over WeChat.
3. Acceptance criteria go in the PO, not the datasheet. 'IP65' is not an acceptance criterion. 'No visible ingress after a 12.5 L/min jet from 3 m per IEC 60529 enclosure test 5, on three units drawn from the first hundred, breather membrane present, gasket joint documented in the drawing' — that's an acceptance criterion. It's ugly and it works.
4. Default credentials are a rejectable defect. Unique per-device admin password on the label, WPS off, WAN-side management off unless the customer asks in writing. No exceptions, no 'we'll cover it in the integrator guide.'
5. Every number on a datasheet gets a qualifier or it doesn't ship. Peak, typical, tested, or theoretical. Pick one and say it out loud.
None of that costs anything. The gasket rework, the air freight, the slipped quarter — that's what costs money. The five-line version of this is now page one of the LinkScope Sourcing 20 supplier packet, mostly because I got tired of explaining it verbally.
Which is the part that took me too long to internalize. For years I treated vendor selection as a price question with a quality rider on top. It's the other way round. The unit price is the smallest number in the file. The expensive part is everything that happens after the pallet lands and the photos start coming in.
I still have that housing on my desk. Flashing line, wrong gray, LED window trimmed out. Most useful thing in my office.
