Posts and articles

The result is seven principles for humans and machines working together responsibly.

I'll be honest: this is preliminary, independent, and I'm sure parts of it are wrong.

That's why I'm sharing it now.

I need your help to make it better.

Read it, then tell me which principle you'd challenge and what I've missed.

It produces more of them, faster, than any team ever could.

But an output is not an answer.

Someone still has to check it.

Someone has to know which failures matter.

Someone has to decide whether a number is right, or whether it should not be reported at all.

And someone has to stand behind it when a regulator or a reviewer asks the hard question.

That someone is a human.

We usually hear "commodity" and think cheap and interchangeable.

But the original commodities were the things every economy ran on.

Grain, oil, steel.

The things nothing works without.

In the age of AI, that's human judgment.

The value isn't in the human who produces the output, because the model does that now.

The value is in the human who verifies it.

I spend my days building tools for cell and gene therapy research, where a confident wrong number is worse than no number at all.

That work has made one thing very clear to me.

The more AI we use, the more valuable the people who check it become.

Software is becoming abundant.

Judgment you can hold accountable is not.

Is that the right way to see it, or am I missing something?

The tricky part is the label.

"1 mg" might mean 1 mg of the thing you want, or 1 mg of everything in the vial, stabilizers included.

And the powder takes up space once it dissolves, so 1 mL of liquid doesn't give you 1 mL of solution.

Imagine a recipe asking for a cup of chocolate when the bag you bought is half sugar.

You measure perfectly, the cake still comes out wrong, and nothing on the plate tells you why.

A spreadsheet just does the division.

This tool asks what the label number really means before it gives you an answer.

Then it writes your answers next to the result, so whoever uses that solution later knows how it was made.

Here's why that matters: a wrong solution looks exactly like a right one.

It runs in your browser, nothing you type leaves it, and there's no account to create.

It's for research use only.

Is there a calculation in your lab you've stopped trusting?

Tell me which one, because that's how we pick what to build next.

A spreadsheet prints that number anyway and says nothing.

Our tool stops and tells you why it won't work.

If there is a way around the problem, it shows you the extra step.

If there isn't, it says so plainly instead of guessing.

It also asks what you mean before it does any math.

"Make it up to 100" can mean three different things, and each one gives a different answer.

It then shows you the amount you asked for and the amount you will really end up with.

So who is it for?

Anyone who mixes solutions at a lab bench.

Students learning the basics, researchers running experiments, and lab leads who check other people's work.

It's free, and there's no sign-up.

It runs in your web browser, and what you type is never sent to us.

The code is public, so anyone can check how it works.

The link is in the first comment.

Try it with a dilution you think is simple, and tell me what happens.

And one practical step: rely on private, local, open-source AI where you can, so power doesn't sit in the hands of a few companies.

(Yes, the image is AI-generated. The race isn't.)

The Molarity Converter for Biologics will not return anything until all three questions are answered. "Not recorded" is an accepted answer for provenance, and it prints on the output. That is the point of it.

Free, open source under Apache 2.0, runs entirely in your browser. No account, no data collection. Research use, not qualified for GxP decision-making.

Built with AI assistance. I set the question and signed off on what shipped.

Every one of those produces a confident number, and the checks you'd normally run come back clean.

One has an R² of exactly 1.000 and reports every value four times too low. Another gives statistics numerically identical to a correct calibration while every number is off by 30%.

Not similar. Identical.

So we built the thing that sits next to the calculator rather than replacing it. Same arithmetic, but it checks whether the calibration underneath is sound, and when it isn't, it refuses to print the number rather than printing it with a warning nobody reads.

Free, runs in your browser, nothing you type leaves your machine.

Check out the free tool here: https://lnkd.in/gcwqJYED

P.S. This is what I mean by "Pharma 5.0. Human-Machine Collaboration." Not machines making the call. Machines do the work of assembling the evidence and are honest about where that evidence runs out, so the scientist can make a better decision than they could alone.

So we tested for failure instead of correctness. We built calibrations in which we already knew the true answer, then measured how far the tool could be off. We rebuilt the whole calculation a second time in a different language and ran both against the same inputs, because reading code tells you what you meant to write, not what you wrote. And we attacked the live site directly, because the version people actually use is the one that matters.

That found three real defects. One had been live for three days. None were caught by reading the code, and none by the existing tests.

All three slipped through because our fixtures assumed the data would be tidier than real data is.

The tool is free and runs in your browser. Try to break it, and tell me where it gives way: benchtools.ligant.ai