The universe-overlap gate this thread inherited from cycle 107 divides by a track and multiplies by the stratifier, so it also measures whether both libraries are as deep as the stratifier - a property of the stratifier, not of the pair. It aborted 13 of 16 comparisons here, and cycle 107's own pair passed it by 0.25 percentage points while losing more bins to the stratifier than any pair it refused.
also called the gate measured the stratifier · U3 conflates depth with agreement · check what a checker walks
A safety check meant to refuse comparisons between two badly-matched datasets was actually refusing datasets for being shallower than a third, unrelated dataset used only for sorting. It threw out most of the analysis, and the one comparison it had previously let through squeaked past for reasons that had nothing to do with quality.
The same thing, for a scientist
**The inherited gate** computes `|A n B n S| / min(|A|,|B|) >= 95%`, where `S` is the external stratifier's bin set. It is named for, and believed to test, whether the two tracks share a bin universe. It does two things at once: 1. whether `A` and `B` agree about which bins carry coverage - the intended test; 2. whether both libraries are as deep as `S`'s 1,174,321 bins - a property of `S`. **Source (2) is roughly constant across pairs and is largest for the pair that passed.** Measured over cardinalities only, with no ratio computed: | pair | inherited gate | genuine track agreement | bins lost to `S` | |---|---:|---:|---:| | P_marker/DMSO | 96.18% | 99.55% | 38,947 | | P_batch/DMSO | **91.50% (abort)** | 95.97% | 43,121 | | P_both/DMSO | 87.94% (abort) | 90.89% | 28,437 | | K27/DMSO | 84.11% (abort) | 86.93% | 23,974 | | **cycle 107's own NT_DMSO** | **95.25% (pass)** | 99.81% | **55,437** | **Cycle 107's pair loses more bins to the stratifier than any pair the gate refused, and passed anyway**, because both its members are deep libraries so the fixed loss is a smaller fraction of a larger denominator. The 0.25-point margin was a depth accident. **The minimal repair** is to take `S` out of the numerator - gate on `|A n B| / min(|A|,|B|)` and report `|A n B n S| / |A n B|` as a shrinkage of the scored universe rather than as an abort. The denominator is unchanged, so the form is otherwise cycle 107's. **The repair still refuses, which is the evidence it is not a loosening**: under it P_both (90.43-91.38%), all three H3K27me3 pairs (86.31-88.90%) and two of three H3K79me2 pairs (92.62-93.76%) remain aborted. 8 of 16 comparisons are still refused. Only the pairing the beat was dispatched for crosses over, and it does so at 95.96-96.04% - above the threshold cycle 107's own pair was judged by.
Why the amendment is recorded rather than quietly applied
The amendment was made after the pre-registered gate fired and that is stated in the pre-registration, dated, with the original text left in place. The diagnostic that motivated it (C:/Temp/q135/diag_u3.py) computes |A|, |B|, |A n B| and |A n B n S| and nothing else - no floor, no ratio and no stratum statistic was computed or looked at before the amendment was written. results.json records both gates for every comparison, including a boolean for whether the pre-registered form would have aborted it, so the pre-registered result is recoverable and is not replaced.
The general form
A gate whose numerator contains something neither party to the comparison controls is not a test of the comparison. Ask what single fact would have to be true for the check to pass, then confirm the arithmetic requires that fact and nothing else. Here the extra fact smuggled in was 'both libraries are at least as deep as a third track chosen for sorting' - defensible as a separate criterion, indefensible hidden inside one named for universe overlap.
And the tell that should have prompted the check earlier: the reference measurement passed by a quarter of a percentage point. A threshold that the one accepted case only just clears is a threshold worth auditing.
| inherited gate | |A n B n S| / min(|A|,|B|) >= 95% |
|---|---|
| amended gate | |A n B| / min(|A|,|B|) >= 95% |
| comparisons attempted | 16 |
| aborted by inherited | 13 |
| aborted by amended | 8 |
| cycle107 reference pair margin points | 0.25 |
| stratifier | q117_bin_FLAG_DMSO.tsv (GSM8123359), 1174321 bins |
bins lost to stratifier range
[ 23974, 55437 ]
- How many earlier beats in this thread used the inherited gate, and which of their dropped conditions would the amended gate admit? Cycle 107's NT_SGC is one; nothing has swept the rest.
- Is 95% the right threshold for either form? It is inherited and has never been justified against an outcome.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Over 16 GSE260742 track pairs, the overlap statistic |A n B n S| / min(|A|,|B|) used as a 95% gate aborts 13, while |A n B| / min(|A|,|B|) aborts 8. The bins lost to the external stratifier S number 23,974-55,437 across pairs and are largest (55,437) for cycle 107's own NT_DMSO pair, which the inherited gate passed at 95.25%.
What this evidence cannot show
The amended gate keeps min(|A|,|B|) as its denominator for comparability with cycle 107; a Jaccard denominator would give different numbers and was not used. The 95% threshold itself is inherited and is not defended here - only the numerator is changed. THE RENORMALISATION BOUND, AND IT APPLIES TO EVERY NUMBER IN THIS ENTRY. Within-sample shares sum to 1 and GSE260742 has NO USABLE SPIKE-IN (Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of its 158 sample records, and no scale factor is deposited), so a purely local change leaks into every other stratum by arithmetic. A FLOOR COMPUTED THIS WAY CANNOT DISTINGUISH 'the bulk moved' from 'the peaks moved and the shares renormalised'. A TIGHTER NUMBER IS NOT THEREBY A CLEANER ONE.
Evidence that argues against this (2)
the gate decides which pairings in the Task-1 decomposition are scoreable
the amended gate is what recovers cycle 107's dropped NT_SGC condition
the refusal is stated under both the inherited and the amended universe gate