CELLPROOF's method comes within ±0.5% of a US national lab's meter data

This is the first CELLPROOF Report, covering the stretch from our founding in September to early October. In that time we haven't tested a single customer's car, and that's deliberate. One question had to be settled first: can the figures we measure be trusted?

This issue's headline answer comes from the meter of a US national laboratory.

The problem

The most direct way to find out what a battery still has in it is to see how much energy it takes in. That's exactly what DEEP does: over one night's AC charge, it adds up the energy flowing into the battery, moment by moment.

It sounds simple. The hard part is getting it right every time. If the arithmetic is even slightly off, every figure after it is off too, and nothing in the result would give it away.

So before touching our first customer's car, we needed a ruler everyone accepts, to measure our own ruler against.

How we checked

Argonne National Laboratory in the US has published complete data for a 2020 Model 3 on a test bench^[1]: the car's own signals, recorded alongside a power analyser on the battery. Each charge is a long haul of 8 to 12 hours.

A power analyser measures the energy that actually flows into the battery, regardless of what the car estimates, which makes it the perfect answer key. We used only the car's signals, worked out the energy in with our own method, and compared it with the laboratory's readings. We checked 3 charges first; then, in the audit, we brought out 2 charges we had never used, to see whether the result held.

The result

On the 5 AC charges that meet DEEP's measurement conditions, the difference ranged from −0.24% to +0.48%; not once did it leave ±0.5%^[1]. Four full charges of the same car agree with each other within ±0.2%^[1]. Put another way: test the same car twice and the figure doesn't jump around.

LFP batteries behave a little differently from nickel-based ones, so we also turned to an LFP Model 3 published by the Technical University of Munich^[2]. The car had an ultra-slow charge when new and another after about 26,000 km, each taking nearly 57 hours. Our method's battery capacity came within 0.1% of the paper's figure both times^[2].

What it means if you're buying

Put simply, the ruler we measure energy with is straight. The DEEP results you'll see start from a calculation that agrees with a laboratory meter.

To be clear, though: this validates the method, on public laboratory and research data from abroad, not on cars in Taiwan. How the certificate's battery health figure is worked out, and its error range, will be published once validation on cars in Taiwan is done.

Other progress

Beyond the headline, here's what else happened in this period. Every piece of it serves the same goal: figures that stand up to questions.

Write the criteria first, then look for our own mistakes

The real danger in validation is fooling yourself, so this audit followed the researchers' rule: what counts as a pass was written down and saved before any new data was fetched. The exam was 27 drives of real cars we had never read, and every decoding rule held. In other words, the way we read a car doesn't only work on data we already know.

Testing the tests

Having lots of tests doesn't mean they catch mistakes, so we broke the software on purpose to see whether the tests noticed, and added new tests wherever they didn't. The database that holds certificates gets the same treatment: before every update it is rebuilt from scratch on our own machine and checked rule by rule, and only goes live when everything passes. You'll never see this work, but every figure on a certificate has to get through it.

2026 firmware included in our validation

Cars update their firmware over the air, and the format of their battery signals can quietly change with it. If our reading doesn't keep up, the worst case is reading wrongly and getting something that still looks plausible. We checked public recordings of real cars on 2026 firmware item by item, and what needed changing is live. When a format is new to us, the system refuses to read it: better to leave that item blank than to print a figure we aren't sure of.

Receive only, never send

People often ask: could it touch my car? Our answer is a number: while recording, the signals we send to the car number zero. The logging equipment only listens and never transmits, not even an acknowledgement, and the recording software won't start until it has confirmed receive-only mode. Our test equipment has arrived and is being set up.

From the website to cloud analysis

A certificate is only useful if others can check it for themselves. In September the website went live with our two tests, SWIFT and DEEP, and certificates became verifiable online by number or VIN. Once a test is uploaded, the analysis runs by itself in a data centre in Taiwan, then goes to a person for review, and only then is a certificate issued; with the demo car, analysis finished about 1 minute after upload.

This issue in numbers

What's next

Next, we take this ruler onto Taiwan's roads.

1. 01Coming up

   Our first test on a car in Taiwan

   Recording on a car in Taiwan with our own equipment, and checking each reading against what the car's screen shows.

2. 02Coming up

   The battery health calculation, finalised

   Setting how the certificate's battery health figure is worked out, from DEEP tests on cars in Taiwan, and publishing its error range.