Most of what I believed about my own systems did not survive measurement.

A benchmark fails silently far more often than it fails loudly, and the failure is almost never in the arithmetic. Across thirteen cases the number was correctly computed every time: what was wrong was the answer key, the population, the headroom, the baseline, or which artifact was measured. The cheapest defence found so far is a second number that must agree.

An index is a decision about what is searchable, not a copy of a repository, and the decision is usually made by accident. What hurt in every case was text nobody would ever search for. But the size of that harm did not generalise: the same proportion of prose did different jobs in different repositories.

Context fills from tool output, not from conversation, and the cost is paid on every subsequent turn rather than once. Continuing a session feels free at each individual moment, which is why the ceiling arrives without warning.

Every apparent GPU problem here was a plumbing problem, not a capacity one, and the two want opposite responses. Twice the accelerator was idle while something upstream fed it badly, and once it was not running at all despite being configured.

Picking the right repository is a retrieval problem with almost no published benchmarks, and the obvious signals do not work on it. Absolute similarity scores failed three separate times as a basis for any threshold. Size bias is real and survives the usual normalisations.