1. Is this the data the factor was fitted to?

   The paper's worked example lists AVLP532 -> CL110 as
   10.45 scaled male synapses against 20 female.
   This table holds male 18.0, female 20.0.
   18.0 x 0.581 = 10.458

2. Fitting the same PCA

   variant                                               n   PCA slope
   all edges, raw counts                         3,761,792      0.8413
   both sides nonzero                            1,553,602      0.8423
   noise box removed (male>10, female>8)           374,024      0.8431
   male-side cut only                              559,472      0.8427
   square-root counts                            3,761,792      0.8234
   log counts                                    3,761,792      0.8608
   hemisphere-split proxy (halved)               3,761,792      0.8413
   input-normalised                              3,619,455      0.9576

3. Does a filtering choice explain the gap?

     min male   min female          n   PCA slope
            0            0  3,761,792      0.8413
            2            2  1,023,495      0.8425
            5            4    625,060      0.8428
           11            8    388,946      0.8431
           20           15    255,175      0.8435
           50           40    119,793      0.8442
          100           80     60,719      0.8452
          200          160     26,312      0.8468
          500          400      6,600      0.8521

   slope stays between 0.8413 and 0.8521 while the cut moves over
   a five-hundredfold range. Filtering does not move it toward 0.581.

4. What the factor is supposed to do, and what each choice does

   After alignment the two datasets should sit on top of each other,
   so about half of edges should look stronger in each sex.

                           factor    male/female  % edges male stronger
         published 0.581 (0.5810)          0.768                  32.3%
       PCA on this table (0.8431)          1.114                  57.7%
            sum matching (0.7568)          1.000                  50.1%
              no scaling (1.0000)          1.321                  67.8%

   Edges whose apparent direction depends on which factor is used:
   98,990 of 388,946 (25.5%), every one of them female-stronger under
   the published factor and male-stronger under the refitted one.

5. What this does not show

   The released table carries only edge totals. The paper fitted on
   weights split by hemisphere, after excluding nerve-cord
   connections and restricting to named superclasses. That exact
   table is not published, so a preprocessing step that is invisible
   here could account for the gap. What can be said is that the
   slope of the published cloud is 0.843, that it does not move
   under any filtering or transform tried, and that the difference
   from 0.581 shifts the male-female balance in one direction.

wrote D:\fly-research\results\c2_pca_factor.json
