11 projects, and the number in each one I stopped trusting.

1The assembly passed its clearance check. It passed on nine straight design changes.

2A solid overlap test found the hinges had been interfering the whole time. Nine passes of a weak test is not evidence.

1The critic separated real swarms from generated ones, so the GAN looked like it was training.

2It was separating them by magnitude alone. A bare Tanh output covers under 40% of the real feature range, so the critic never had to look at the behaviour. That is what broke run 1.

1The output looked right, and the tests passed.

2Looking right is not being reproducible. A test walks the syntax tree and fails the build if the engine imports a clock, so the same input cannot quietly produce a different file tomorrow.

1Macro F1 of 84.8% across the entity types.

2Macro F1 averages every class equally, so a rare obligation type carries the same weight as a common one. That is the number I want for compliance, where the rare clause is the one that bites.

1The fused trajectory tracked the truth, so the fusion looked right.

2Each sensor has a different failure mode. VIO fails on blank walls, UWB fails behind metal, and the IMU never fails, it just drifts. The test that matters is dropping one at a time.

1The explosion looked right on screen, and looking right is most of the job in a visualisation.

2Looking right is not the physics being right. A separate NumPy implementation runs the same rules headless and logs the numbers, so the shader and the maths have to agree.

1Best model: Random Forest, test R squared 0.9414. That is the number a dashboard would show.

2Plain linear regression scores 0.9410 on the same split. The forest wins by 0.0004, and its train score of 0.983 against 0.941 on test says it is memorising rather than learning.

1An 87% ATS pass rate means the resumes get through the filter.

2Getting through a keyword filter and being worth an interview are different targets. The score is a parser check, not a judgement of the person.

1The backtracking matcher returns the right answer on every test pattern.

2Right answer, wrong cost. The Thompson NFA simulator gets the same answers in time that does not explode, which is the whole point of a ReDoS check.

1The feature values matched the reference implementation, so the code was correct.

2Correct is not the same as useful. The real test came later, when these features had to separate rock from sand in SkyScout perception.

1The models scored well on the held-out split.

2Molecules are not independent rows. Very similar compounds landing on both sides of a random split make a model look better than it is. Chemistry decides what counts as a fair split.