3 · USE CASE

Six months later: the reported sequence is shifted

‹ back to “Reference trust v2: what was built”
What the reads carry — and what is reported51-bp duplication, chr13:28608240, 120 reads, zero mismatchesreads120 readsG | AAACTCCCATTTGAGATCATAT…CGTAGAAGTAreported0 readsG | GAAACTCCCATTTGAGATCATA…ACGTAGAAGTanchor | inserted basesThe reported insert is the anchor base plus the inserted bases without thelast one: a sequence no read contains.Why it mattersRead literally, a shifted ALT changes the protein: in #428 an in-frameinsertion turns into a premature stop. At the production setting(--ref_trust 0.9) the reference copy hides 1-base shifts of pure duplications.On 228 lab BAMs at --ref_trust 0.9 (calls ≤ 60 bp, ≥ 5 reads)current code0 / 52scratch fix, both slices20 / 48read-derived sequences found in the readsWhy every check passed2022CODERead slice written one base early; its output was replaced by thereference copyApr 2026MERGEPR 67 reports the read-derived sequence: the shift becomes visibleApr 12REPORTRegression explains a 1-base shift: “neither is ‘wrong’ … the sameunderlying duplication read at slightly shifted alignmentpositions”Jul 28TEST DATAhg38 fixture: expected output carries the shifted alleleOct 3DEFECT#428: simulated data — an analyst model reads the VCF literally,wrong proteinOct 12DEFECT#438: claims for a talk checked against the reads — 0 reads carrythe alleleOct 14TEST DATAClean synthetic reproduction; artifacts and commands in #438Consistency with the old version is not correctness:the reads are the ground truth