[geoelectric]

I know it was a throwaway, but be careful about the "untested code never works" comment. It plays into the fallacy that if it works, it's adequately tested.

You talk a lot about testing for specified requirements, but not so much about looking for undesired behavior outside the requirements. That's where the fallacy bites you. It's especially pernicious when you start to firm up a working prototype or are inheriting a legacy production release with a history of good behavior, but which is about to change execution context in some way. Changing the context means hitting unexplored territory and there be dragons.

Of course, you're a professional tester (as am I), and I know this is obvious for you. But for the average PHB or engineer-driven organization trying to prioritize, it's important to be pretty clear about this:

Untested code works great in the 95% case--this is one of the big reasons QA has a credibility problem sometimes, because they're spending $$$ to confirm it already works. It's the 5% case you need to worry about. It can kill your product or, in your case, the patient.

[julienzaegel]

(I'm the author of the aforementioned article)

Yes it was a throwaway :-) But I mean "untested" as not tested at all, by any means; code that has went through automated testing (unit tests and the like) has been tested quite a bit, so when QA receives it, the defect rate is not so overwhelming (5% or so as you say). I maintain that the code I write almost never works if I don't write automated tests for it (which is why I always write tests now :-))

You're right about the importance of "looking for undesired behavior outside the requirements". I'll think about an update of the article about manual tests. Thanks for the input!

[geoelectric]

I was QA prior to the advent of mandatory unit test for commit (I'm actually an SDET or tools/infra SWE nowadays, depending on hat) and, even then, untested code was mostly fine, as evidenced by the fields of green "PASS" spreadsheets.

Those spreadsheets either meant it was good when you got it or you did a crappy job testing it, and either way it was fine before being tested. Regression testing in particular is an exercise in generating "PASS" results, and historically QA practice has been dominated by regression testing.

Luckily, SW practices have changed enough that (as you say) most code is at least minimally tested by the developer and, maybe more importantly, some minimal level of unit testing is expected now as a basic professional standard. Hopefully we never move backwards again.