The promoter table's second blind spot is not annotation drift at 1.8% - it is one row per symbol at 14.7%, and the case that motivated the question is not displaced at all

Comparison is by gene SYMBOL, so a symbol that was renamed between the two releases appears as 'absent from current' rather than as displaced, which means the 522 absent rows may contain genuine displacements that this method cannot see - the 1.83% is therefore a lower bound on displacement and the 522 are the uncertainty. Alt, random and fix contigs were dropped from the reference. 'Multiple non-overlapping TSSs' is defined as a >=2 kb span between the extreme TSSs of the symbol on the row's chromosome, which counts long first-intron alternate promoters as separate models. Occupancy was NOT re-scored at the alternate windows in this run, so the number of calls that would change is unmeasured - q155's planned step 5 remains undone.

• A 1.83% displacement rate does not mean 1.83% of this project's CONCLUSIONS are safe from annotation age, because the project's conclusions are concentrated in a small number of genes rather than spread over the table. A single displaced row at a load-bearing locus would matter more than 355 at random ones; the load-bearing loci checked here (IRX1, IRX2, AFF1) all have d_nearest of 0 or 9 bp, but that is four genes, not an argument about the rest.
• The 14.69% multi-model figure counts symbols, not calls, and most of those symbols are never used by this project. Treating it as the size of a defect in the project's output would be the same error as treating a name-drop as a dependency, which this corpus has recorded. The honest statement is that it bounds the AMBIGUITY of the table, not the error rate of the findings.

This corrects the framing of q155's hypothesis but not its underlying observation: q155 was right that the scored C4orf36 window is ~42 kb from AFF1 and that a C4orf36 TSS sits within a few hundred bp of AFF1's, and right that this makes the AFF1 window unattributable. Only the diagnosis - 'displaced against a newer annotation' - is wrong. The 278 bp figure is computed from the current release; q155 reported 269 bp, and the small difference was not chased to its source, which is most likely a different transcript chosen as AFF1's representative TSS.

• Calling C4orf36 'not displaced' depends on accepting symbol-level matching: if the 2020 table intended the promoter-proximal C4orf36 model and the release changed which transcript is canonical, then from the table-builder's point of view the row IS pointing at the wrong place, and d_nearest = 0 merely records that the old coordinate still exists somewhere in the new release. The two framings differ in what they treat as the unit of truth, symbol or transcript.

The two censuses are not identical in construction: q153 counted shared windows between table rows and, separately, against a full annotation, while this run counts any foreign current TSS falling inside a row's window. They agree to within 0.2 and 1.0 percentage points respectively, which is close enough to call reproducible but not a re-implementation of the same statistic. The all-curated figure being HIGHER here than q153's (22.63 vs 21.60) is expected, since NR_ transcripts add TSSs.

• Agreement between two censuses that both count TSS proximity in the same genome is weaker evidence than it looks - they share the dominant driver, which is the real density of bidirectional promoters, so they could agree while both mis-specifying the window. This corpus has recorded that two derivations asking the same question measure consistency rather than validity.