pilot8 vs pilot16: the junction flags cost no worm counts, on the one worm the two runs share

[lhqing]

Measured 2026-08-24 on ircbc. A measurement is not a decision — what is settled here is what the
two runs contain. What it implies for the plate is not.

The one-sentence answer

Between pilot16 and pilot8 the worm counts did not move (+0.21% on x_sum, +13 genes) while the
phantom junctions collapsed — novel splices −70%, longest N gap 1.33 Mb → 49 kb, and every spliced
alignment in the intron-free bacterium gone — but the matched set is one worm, so this is a
consistency check on the plate-wide splice-flags gate rather than independent evidence.

1. pilot8 ran clean

ss3-pilot8-chimera-8f5991a42944, job 198678 on fat01 (160 cores), 2026-08-24 03:41:03 → 05:44:58
CST, rc=0, 8 of 8 samples finished, no missing deliverables, no alerts. Peak PSS 7.86 GB, peak
results/ 83.2 GB. WORKFLOW_VERSION 2026.8.25.

Memory is PSS summed over the job's descendants, never RSS: RSS counts the shared genome segment once
per attached STAR job, and that is the figure the August run got wrong twice.

2. The matched set is one worm

pilot16 and pilot16r2 are the same 16 N2 animals — day1_N2_1–8 and day17_N2_2,6..12. pilot8 is 8
animals spread across ages and strains. The intersection is exactly day1_N2_1.

The compare directory built alongside pilot8 does not use pilot16 as its before arm — before.per_cell.tsv
is byte-identical to the 784 plate's rows for those 8 cells. For day1_N2_1 that distinction is empty:
pilot16r2's own QC bundle gives share_kept_ecHT115 = 0.001518658890148116, matching the before column
to the last printed digit, and pilot16 v1 and pilot16r2 have identical STAR output for this cell. So the
before column is pilot16 here, and the arm is named honestly.

3. day1_N2_1, pilot16 → pilot8

	pilot16	pilot8	delta
input reads	9,039,826	9,039,826	identical
uniquely mapped	60.49%	60.55%	+0.06 pp
multi-loci	32.44%	32.30%	−0.14 pp
unmapped too short	6.76%	6.84%	+0.08 pp
splices, total	1,067,486	1,068,705	+0.11%
splices annotated (sjdb)	94.79%	98.44%	novel −70%
junctions (SJ.out)	44,601	45,031	novel −9.8%
max N gap, ce11	1,326,638	49,271	cap holds
spliced %, ce11	9.368%	9.336%	−0.03 pp
spliced %, ecHT115	1.185%	0.000%	258 → 0 records
max N gap, ecHT115	191,584	0	—
bacterial share of kept	0.1519%	0.1440%	−5.2% rel
ce11 x_sum	376,955	377,746	+0.21%
ce11 genes	11,913	11,926	+13
ecHT115 x_sum	887	887	0.00%
STAR wall	7m08s	2m13s	3.2x

Splitting is unchanged in shape: records_in 18,079,652 both arms, kept ce11 16,759,056 → 16,748,532
and ecHT115 25,490 → 24,154, dropped.unmapped 1,295,106 → 1,306,966.

The ecHT115 zeros are not the cap doing the work. A 191 kb gap would be excluded by a 50 kb cap, but
gaps under it would survive and they did not. The anchor filters did the discriminating; the cap is the
backstop, which is what splice_args.py says it is. day9_N2_9 is the counterexample that proves the
point — 9 spliced ecHT115 records survive at a 16,093 bp gap, comfortably inside the cap.

The 3.2x speedup is suggestive, not a benchmark: 16 cells against 8 on the same node is a different
run shape. The mechanism is real — the cap re-derives winBinNbits to 14, dropping per-step window reach
from ~590 kb to ~147 kb — but this number does not measure it.

4. The variable is not one flag

pilot16r2 ran STAR with no splice flags at all. pilot8 adds five, plus the junction attributes:

--outFilterType BySJout --alignSJoverhangMin 8 --alignSJDBoverhangMin 1
--alignIntronMax 50000 --alignMatesGapMax 50000
--outSAMattributes NH HI AS nM jM jI

WORKFLOW_VERSION also moved 2026.8.17 → 2026.8.25 across eight bumps — the rRNA counter at 8.24, the
whitelist path at 8.25, the per-cell QC record at 8.21, the index lookup's shared path at 8.22. Anything
here that is not the flags is carried by one of those.

The index is not a confounder. pilot16r2's results/index/ce11_ecHT115 is a symlink to the same
shared star_wormbase_ws298+refseq_rs_2025_06_26 that pilot8 read, so the annotated-splice jump cannot
be an sjdb difference.

The rRNA columns are new at 2026.8.24 and have no before value — they are carried as the empty string
rather than as zero, because a new measurement is not a change.

5. What this does not settle

n = 1. One worm, at the youngest age, in the strain with the lowest bacterial load. It agrees with the
plate-wide result that the junction flags cost no worm counts, and agreement at n=1 is a consistency check,
not a second measurement.

The wider table in the same directory compares all 8 pilot8 cells against the 784 plate, which ran at
WORKFLOW_VERSION 2026.8.18 and is therefore also pre-flags — the same generational comparison at n=8.
The pattern holds there: ce11 x_sum within ±0.21% on all 8 cells, every ce11 max N gap under 50,000, and
ecHT115 spliced % at exactly 0 in 4 of the 8.

Whether any of this changes a figure in the aging series is a question for the full remap, not for eight
worms.

Method and where the files are

compare/collect_after.sh folds the after arm with the same three scripts the before arm was folded by —
one script per measurement, both arms — so a difference between the arms cannot be a difference between
two collectors. join.py merges them into comparison.tsv.

Everything cited moved to aging_SS3/archive/ on 2026-08-24 ahead of the full-plate remap:
archive/pilot8/compare/comparison.tsv, archive/pilot8/pilot8-198678.out,
archive/pilot16r2/chimera/seqforge/pipeline/ss3-p16r2-chimera-365020090ac3/.