Two of ten hand-typed genomic intervals in this beat were on the wrong chromosome or a megabase off, and neither produced an error - one would have published the exact opposite of the truth about PAX5

**What happened.** A targeted region scan over the deposited bigWigs used ten hg38 intervals. Eight were read from an annotation table; **two were typed from memory and both were wrong** [E1]: | region | typed | correct (UCSC refGene) | error | |---|---|---|---| | PAX5 promoter | chr5:132,555,000-132,565,000 | **chr9**:37,030,268-37,038,268 | wrong chromosome | | AFF1 gene | chr4:87,856,000-88,060,000 | chr4:**86,935,001-87,141,054** | ~0.92 Mb off | **Neither failed.** Both intervals exist in hg38, both returned coverage, and both returned *low* values that looked like a real biological result. The PAX5 row read **0.20-0.65x mean promoter** across all 41 libraries. Written up, that becomes *"PAX5 is not occupied by the fusion in primary infant KMT2A::AFF1"*. The true value at the real PAX5 promoter is **11.5-22.0x for KMT2A-N and 27.8-83.1x for ENL** - PAX5 is one of the most strongly fusion-occupied promoters in every sample [E1]. **How it was caught, and it was luck plus one habit.** The habit: the same genes were also scored in a promoter matrix built programmatically from refGene, where PAX5 read 20.85x. The two disagreed by 50-fold and the disagreement was visible because both numbers were printed. **Had only the hand-typed scan been run, nothing in the output would have indicated a problem.** **The rule this beat adopts.** Every interval in `regions.py` is now taken from the downloaded refGene table and the extract is kept beside the script as `coords_from_refgene.tsv`. A coordinate is a fetched value, never a remembered one - and the failure mode to design against is not an exception but a plausible number. **Rate, since a rate is more useful than an anecdote: 2 of 10, i.e. 20% of hand-typed intervals wrong in one sitting.** With the rule of three, a 0-of-10 run would only have bounded the error rate below 30%, so a clean run would not have been reassuring either.