What a read sample can prove

Decision: fetch inspect classifies a run from the first 100,000 reads of each file. A call from that sample may refuse a source only on positive evidence that the source cannot be used. It never refuses on absence of evidence, and every call records the classifier version that made it.

The first reads are not a random sample

Files are written in the order a sequencer, an aligner or an archive produced them, and the first reads are a stratum of that order:

• A coordinate-sorted BAM starts at the start of chr1. SRR10377488's Cell Ranger BAM has 680,000 records in its first 32 MB. All of them are in chr1's first megabase, and 97% are secondary alignments to repeats and to a mitochondrial insertion near 630 kb. Read for strand, they said the 3' library was mostly antisense. The barcodes are unaffected, since every read of every cell carries one. So on a coordinate-sorted BAM the classifier reads barcodes and refuses to measure strand or library type.
• A fastq starts at the first tile of the first lane. Edge tiles carry more Ns and lower quality. That is why barcodes are matched within one substitution, with N as a wildcard.
• A multi-lane run is one lane first. A run whose lanes hold different libraries (a GEX lane and a feature-barcode lane in one SRA run) is called on the first lane's library. That is a limit of the method, and the call says which file it read.

What the first reads are a fair sample of is a library's structure: which reads exist, how long they are, where the barcode starts, what follows the UMI. Those are the same for every read of the file, which is why a few thousand reads already say them.

Refusing a source needs positive evidence

The file layer's refusal path (files) will drop a candidate only on a proven-unusable verdict. The standard for one:

• no read of 30 bases or more besides the barcode read (there is no cDNA read to align), or
• no read of 26 bases or more whose first 16 bases are a listed barcode above background (there is no barcode read). The background is what the list matches by chance: 0.016% exact for the v2 list, 0.16% for v3.

Geometry alone never proves absence. A file of 150-base reads may hold the barcode read sequenced long. A 28-base-only run (SRR30132199) is unusable because it has no cDNA read, not because 28 is an unusual length.

The first spots of an SRA object must account for the run. This is where the first reads stop being the library's structure. DRR878211's object holds 26 + 8-base barcode spots first and its 98-base cDNA reads in spots of their own further in. Its runinfo mean is 66 bases a spot, against the sample's 34. A refusal of an SRA object therefore holds the sampled spot against the run's mean bases per spot, from runinfo. If the sample falls short by more than a base, or the mean is unknown, nothing is refused. A derived fastq is not held to the mean. ENA's lone fastq of SRR29534765 is short of it because it omits a read, and that omission is what is being refused.

A read of 11 to 25 bases may be a barcode no list holds. SRR9105023 is 10x 3' v1: a 98-base cDNA read, a 14-base barcode read and a 10-base UMI read. No v1 list is registered, so "no listed barcode" proves nothing about it. A sample index is 8 or 10 bases and is still set aside.

Fit needs the barcode read and the cDNA read to be mates. DDBJ's fastqs of DRR878211 pair the barcode read with an 8-base index read and put the 98-base reads in a third file of unpaired spots. The first reads' spot names agree 0% of the time. That set is undetermined, never fit.

These three conditions are rule fit-2. The 2026-10 content survey found each of them by reading the files that fit-1 had judged.

A cDNA read is any read of 30 bases or more that is not the barcode read (rule fit-3). fit-2 asked for 50 and refused every copy of SRR34778689, whose 28 + 47-base reads map; the maintainer counts a 47-base cDNA read usable. Thirty bases is still far above an index read. The barcode read is excluded by position, not by length, because some barcode reads are 30 or 31 bases (SRR22434015, SRR15425104). Excluding it corrected fit-2 too, which let the barcode read stand in for the cDNA read: DRR259182's lone 150-base fastq, every read of which starts with a listed barcode, was called fit. A barcode read that long has room for a cDNA read after its barcode and UMI, and may be the two reads joined, so such a set is undetermined. Only a barcode read too short for that (under 58 bases) with nothing beside it is proof.

A refusal needs every listed file read (rule fit-5). ENA lists SRR11040464's _1.fastq.gz with a size and an md5 and serves an empty directory at that path, and answered 403 for one of ERR4065012's fastqs on the day it was read (throttling, it turned out: it answered 200 the next day). Judged on the file that could be read, each set was refused for lacking the read the missing file most likely holds. Half a set is not positive evidence of absence, so such a refusal is undetermined, and its reason says which file was missing. A copy read as another technology is undetermined too (fit-4): a 10x list not holding a Parse or BD Rhapsody barcode says nothing about whether the library is usable.

Each verdict records what it was judged on: every file's format and modal length, the call, the run's mean spot (run_spot), and the spot-name agreement of every pair of files from different sources. fit_from_detail judges a recorded verdict again from those alone and gives what fit gives on the reads, which a test holds it to. A later rule is checked against the same inputs, and only a record lacking one is read again.

A library that is not gene expression gets fit=None. Whether a VDJ or feature-barcode library is usable depends on a pipeline this project does not run.

Undetermined is not unfit. A CRAM read without pysam, an SRA object without sra-tools, or a fetch that failed all leave a source undetermined, and undetermined has no effect on recommendation.

The classifier is a versioned instrument

Every call carries version. A threshold change is a new version, never an edit of an old one, for the same reason a route's RouteEvidence carries a date: a call cited in the KB must be reproducible from its version and its reads. The thresholds are set on the benchmark's design half and reported on its held-out half, by study, seeded. Version 0.2.0's were set on the 2026-10 benchmark: 0 errors in 66 held-out HIGH calls and 0 in 28 MEDIUM, with disputed labels set apart.

The thresholds were first looked at on the gate runs: SRR25056225, SRR11181954, SRR10340945, ERR4265292 and SRR29534765. Those runs are chosen, not sampled, so their studies are excluded from the benchmark.

What the spike settled (S0, 2026-10-19)

Question	Answer
Does sra-tools read one specific object?	Yes. fastq-dump -X N --split-files <url> reads a given object's first spots. SRR25056225's Normalized object reads 28 + 91 bases with real qualities; its Lite object reads the same bases with ? throughout. An object is attributed by its URL, so no quality-type fallback is needed
Does pysam read a range-fetched CRAM prefix?	Yes, for an unaligned CRAM. The first 32 MB of ERR13855116's 27865_1#3.cram read 100,000 records in 0.1 s with fetch(until_eof=True): paired 0x40/0x80, R1 26 bases, R2 96, BC/QT tags. An unaligned CRAM needs no reference. An aligned one needs EBI's registry (REF_PATH)
How stable is a call at 100,000 reads against 200,000?	Completely, on the benchmark: all 240 runs got the same call from both
On what terms are the barcode lists available?	10x's lists are under 10x's terms. They are fetched at runtime from Teichlab/scg_lib_structs at a pinned commit, sha256-checked, and never redistributed. Flex's list ships only inside Cell Ranger, so Flex is not a layout here
A 5' gate run	SRR11181954 (GSE145926). Its metadata says exactly 10x_5p_v1_v2, and its reads are R1 26 bases and R2 100 bases, the same geometry as 3' v2. It is classified 5' v1/v2 at HIGH, with the cDNA read 13% sense

Two premises in the plan were wrong, and both are corrected here:

• ERR4265292's BAM has no 10x_bam_to_fastq header. It was written by fastq_pre_barcodes. It keeps the cell barcode in CR/CY and the UMI in the SAM specification's RX/QX, not 10x's UR/UY. The barcode read is there all the same, 16 + 12 bases, and the reader rebuilds it from whichever UMI tags are present. Its SRA object holds only the 91-base read.
• rename10xrun is GPL-3.0, not MIT. That is the licence of the repository it lives in. The method is reimplemented and cited, as before.