Findings

This section is a running lab-notebook of decided outcomes: experiments we have run, what they showed, and — most importantly — what each result dictates for how the algorithm and the roadmap should proceed.

It is deliberately separate from Performance → Results, which is the head-to-head benchmark scoreboard (tables of wall time and memory vs R DADA2). A finding here is not "how fast are we"; it is "we tested a hypothesis, here is the evidence, and here is the path it opens or closes."

What belongs here

• A hypothesis and a verdict. Each page states what was tested, the A/B setup, the numbers, and a plain conclusion.
• The consequence. Every finding ends with what it means going forward — a recommended default, a lever to pursue, or a direction ruled out.
• Negative results are first-class. A carefully-disproven idea is as valuable as a win because it stops us re-litigating it. Pages say "do not retry" where that applies.
• Correctness framing. Where a change touches inference, ASV concordance is reported (via dev/compare_asvs.py) using set-identity + abundance, not raw n_asv counts — see the concordance guardrail.

Index

• KDIST cutoff decoupling — the learn-errors and dada k-mer-distance cutoffs can be set independently; the dada-stage speedup is safe to take, while pushing the same cutoff into learn-errors perturbs the error model and churns real-abundance ASVs.
• K-mer screen size — --kmer-size has ~no effect on the final chimera-filtered table on either platform; it is a speed/memory knob, not an accuracy knob (k=5 Illumina, k=7 PacBio for speed).
• Band size & platform-aware defaults — the 16/32 Illumina/HiFi band default is vindicated; the two platforms fail band-tightening through opposite mechanisms, so a single global band would be wrong.