Re-score every committed within-sample share ratio at S_FLAG against Q117's eviction quartiles - the bins are already on disk in five scratch directories, no download and no wet work - and start with the H3K79me2 PCGF1 arm, because cycle 83's central null is the one most likely to change
also called the restratification sweep · re-run the H3K79me2 arm stratified
Every number of this kind the project has published averages over sites where the drug works well and sites where it barely works. Splitting them the way this cycle did needs no new data - the files are already on this machine - and could change what several earlier conclusions mean.
The same thing, for a scientist
**Analysis, not wet work, and not even a download.** This cycle found that three histone marks are monotone in fusion-eviction depth at Spearman +/-1.000 while their aggregates over the same 1,228 bins sit near 1.0. **Every committed share ratio in this corpus at that set is an aggregate of that shape** [E1]. **Priority order, with what is already on disk:** 1. **`C:/Temp/q124/q124_bin_{NT,C1}_K79_{DMSO,SGC,VTP}.tsv`** - Q124's ARM 2. Its conclusion, `ana-q124-the-menin-inhibitors-h3k79me2-effect-is-unchanged-in-a-pcgf1-knockout`, is the evidence base for `hyp-q124`, and this beat's companion hypothesis names it as falsifier F2. **Six bin files, zero downloads, one script.** 2. **`C:/Temp/q122/q122_bin_MM_K119_F3_*.tsv` and `q122_bin_MM_K79_F3{,NT}_*.tsv`** - Q122's ARM 1 and ARM 4, including the 0.8473 this beat's pre-registration was built around. 3. **`C:/Temp/q124/q124_bin_TC_K79_*.tsv`** - the timecourse, where Q124 found a non-monotone series and a library-complexity confound. **A gradient would be a different statistic from the one that was confounded**, and Q124's DEVIATION 2 already showed rank statistics survive there. 4. **`C:/Temp/q120` and `C:/Temp/q118`** - the human arms, which need an eviction stratifier built from menin or MLL1-N rather than FLAG, and that is a design question rather than a script. **And the control that must run first**, because it is this beat's companion hypothesis's own falsifier F1: rebuild the quartiles from the FLAG **SGC0946**/DMSO ratio (GSM8123360, already binned in `C:/Temp/q117`) instead of the VTP/DMSO ratio. **If the gradient survives a stratifier built from a different drug, it is a property of the bins and not of the drug**, and the whole sweep is measuring something else [E1].
What would make this experiment itself uninformative
If the FLAG-SGC-stratified control (falsifier F1) reproduces the gradient, the sweep will find gradients everywhere and they will mean nothing - they will be a property of a bin's FLAG dynamic range, which is correlated with coverage, mappability and baseline occupancy. F1 must be scored first and reported beside every result of the sweep, exactly as this cycle's criterion-informativeness clauses were.
And if the sweep finds a gradient in every channel including ones with no mechanistic reason to have one, the same conclusion follows. That is why H3K4me3's OPPOSITE gradient in this cycle is the load-bearing observation and why the sweep must report the SIGN for every channel, not the magnitude.
| design | For each committed within-sample share ratio at S_FLAG in this corpus, recompute it over Q117's four eviction quartiles (cuts 0.0992 / 0.1748 / 0.2717 of the FLAG VTP/DMSO share ratio) and report the four values, the Spearman against eviction depth, and a DMSO-decile-matched null drawn from outside S_FLAG. Score falsifier F1 first: rebuild the quartiles from the FLAG SGC0946/DMSO ratio and report whether the gradient survives. |
|---|---|
| readout | Spearman of the share ratio against eviction depth, with sign, and the four per-quartile values, beside the existing aggregate and a matched null |
| controls | the FLAG-SGC-stratified quartiles (a stratifier from a different drug); the DMSO-decile-matched null; and the sign of every channel, since a gradient in every channel in the same direction is an artefact signature |
| cost tier | low |
| timeline | one script, hours; zero downloads while the scratch directories survive |
| absence class | measured-not-analysed |
| absence class note | Every input is already binned and on local disk from five earlier beats, and every aggregate the sweep would revisit is in a committed results.json in this repository. The analysis has never been run because until this cycle nobody had a reason to think the aggregate was hiding anything. |
| absence class assigned | 2026-09-23 (Q125) |
priority order
[ "Q124 ARM 2 H3K79me2 NT and C1 (the null hyp-q124 rests on)", "Q122 ARM 1 and ARM 4 (F3 and F3_NT, including the 0.8473)", "Q124 timecourse (where a rank statistic already survived a library confound)", "the human arms, which need an eviction stratifier that is a design question" ]
- If the sweep changes cycle 83's H3K79me2 null, it changes hyp-q124, which is one of the three standing reasons this corpus distrusts its own pharmacodynamic markers. A beat should be prepared for the restratification to make the biomarker problem smaller rather than larger, and should say so before running it.
- What is the right eviction stratifier for the human arms, where there is no FLAG track? Menin or MLL1-N VTP/DMSO per bin is the obvious candidate and it is in a channel that is also a readout, which is the circularity the murine design avoids.
the sweep that would test it, with falsifier F1 as its first control
first priority: six bin files already on disk, zero downloads
restratified, the central null may or may not survive