Gil

Also check whether the API pull is deterministic. If you request 'the last 24 hours' at retry time you get a different window than the original attempt, and then even a perfect merge produces different data. Pin the window to the logical execution time, never to now().

The trick with feedback requests is to make them easy to answer. Ask one specific question rather than 'any feedback': mine is 'was there a moment where you felt I was not the strongest candidate' and the response rate is much better than the open version.

The classification pass is worth doing even if you never merge anything. I found two clients running a 'custom' rounding rule that turned out to be the same rule, written twice, differently, one of which had a bug that had been quietly producing wrong totals for a year.

Worth adding: when a venv is genuinely active, sys.executable points inside the .venv folder. If it does not, the venv is not active no matter what your prompt says.

python -m pip install rather than pip install. That one habit removes about 80 percent of this whole category of problem forever. It is not a workaround, it is simply the correct way to invoke it.

Holdbacks are standard and reasonable, but negotiate what counts as retained before you sign, not after. 'Retained MRR' is ambiguous enough to argue about - does a customer who downgrades count, does an annual that renews on day 61 count, who owns the churn caused by the buyer raising prices in week three. One paragraph of definitions saves a genuinely miserable email thread later.

Mild disagreement on the order. Try python -m venv and pip first, because most packages that used to need a compiler now ship Windows wheels. If it installs, you have saved yourself a whole parallel package manager. If it fails with a wall of red about vcvarsall, then reach for conda.

The first six weeks is when you find all the assumptions. It gets quieter. It never becomes zero.