Oncofusion

start / graph / ana-the-two-silent-failure-classes-did-not-propagate-and-the-audit-corrected-itself-twice-first

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

The two silent-failure classes found in cycle 59 did not propagate through the corpus - 13 candidates, all legitimate on reading, and the audit corrected itself twice before it found anything

also called silent failure class sweep · sub-bin resolution audit · replicate provenance audit · no defect found · the audit found its own construction first

Two ways of getting a wrong number without any error message turned up last cycle. This checked whether the same mistake had been made anywhere else in the project, read every candidate, and found none - though the checking tool had to be fixed twice first, in exactly the way it was built to detect.

The same thing, for a scientist

Cycle 59 exposed two failure classes that produce a plausible wrong number with no error, no zero, no empty querytranslation and no payload shortfall - so every guard this project has is blind to them. Both were found by accident. The obvious question is how many other entries carry a figure from the same shape of mistake, and this is the sweep. **Class A - sub-bin resolution.** GSE210480's H3K9 tracks are **1,000 bp binned**: a 147 bp query returns one distinct value and its mean equals a 501 bp interval containing it. Two cycles costed that dataset without checking, and a coordinator brief said *"a ~20 kb displacement on a 147 bp query"*. **Class B - replicate count from accessions.** The eight CTCF bigWigs in that series are **four byte-identical files under eight accessions**, verified by HTTP HEAD: four distinct Content-Lengths, each appearing twice. Counting accessions understates every standard error by √2. **THE VERDICT IS A CLEAN NEGATIVE.** Over 2,103 entries, **154 name a GEO accession**. The triage produced **10 class-A** and **3 class-B** candidates, and every one was read: - The **147 bp family** (five entries) quote 147 bp because that is the *length of canonical KMT2A exon 11*, not a window they measured. - The **100 bp pair** quote *"100 bp tiling"* as **Krivtsov 2019's own stated analysis parameter** - windows from the TSS extending 10 kb into the gene body - correctly attributed to the paper rather than measured from a track. - The **Gothe BLISS pair** cite deposits carrying `.bed.gz` DSB position lists with per-position UMI counts, where sub-base-pair resolution genuinely exists. - `ana-kmt2a-bcr-ctcf-site-coordinate-overlap`'s unobtainable "~500 bp windows" design was already corrected in cycle 59 when the binning was found. - All three class-B hits say "2 replicates" or "4 replicates", which is the **correct** structure for those arms. The four-files-under-eight-accessions problem is in the CTCF arm, and **no entry asserts a replicate count for it**. **So the classes are confined to the two worked cases that exposed them.** That licenses not re-auditing 154 entries by hand. It does **not** license confidence that the classes are absent, because the commonest form of class A - a figure quoted with **no width stated at all** - is invisible to this audit by construction. **AND THE AUDIT CORRECTED ITSELF TWICE BEFORE IT FOUND ANYTHING.** The first run's class A returned 10 hits of which **nine were a regex artefact**: `\b(\d{1,3})\s?bp\b` matched the tail of every comma-separated number, so "5,026 bp" produced "026 bp" and "1,048 bp" produced "048 bp". An instrument reporting the last three digits of a four-digit number as a sub-kilobase width **is the failure class it was written to detect.** Class B returned **110** candidates and was useless until the pattern was restricted to explicit replicate language: bare "n = 3" is a cohort size and "78 samples" is a deposition census, both legitimate and together swamping the signal. 110 → 3.

Facts and figures
scriptanalyses/audit-track-resolution-and-replicate-provenance/run.py
resultsanalyses/audit-track-resolution-and-replicate-provenance/results.json
statusrun
run date2026-09-22
why this ranThe cycle-59 king log committed to sweeping the corpus for these two classes, on the grounds that both produce plausible wrong numbers with no error and both were found by accident.
class B dispositionall three assert '2 replicates' or '4 replicates', the correct structure for those arms; none counts accessions as replicates, and no entry asserts a replicate count for the CTCF arm where the duplication actually is
absence classmeasured-not-analysed
scope
defects_found0.000defects_found: 0class_B_candidates3.000class_B_candidates: 3class_A_candidates10.0class_A_candidates: 10entries_naming_a_geo_acce…154.0entries_naming_a_geo_accession: 154entries_scanned2,103entries_scanned: 2103
scope — the numbers behind the chart
{
  "entries_scanned": 2103,
  "entries_naming_a_geo_accession": 154,
  "class_A_candidates": 10,
  "class_B_candidates": 3,
  "defects_found": 0
}
class A disposition
{
  "the_147_bp_family": "five entries quote 147 bp as the LENGTH of canonical KMT2A exon 11, not as a measured window - legitimate",
  "the_100_bp_pair": "both quote '100 bp tiling' as Krivtsov 2019's own stated analysis parameter, correctly attributed to the paper - legitimate",
  "the_gothe_bliss_pair": "GSE121740/GSE121742 deposit .bed.gz DSB position lists with per-position UMI counts, so sub-bp resolution genuinely exists there - legitimate",
  "ana-kmt2a-bcr-ctcf-site-coordinate-overlap": "its '~500 bp windows' design WAS unobtainable and was already corrected in cycle 59 - a known case, already recorded in the entry"
}
the audit corrected itself twice
{
  "class_A_first_run": "10 hits, NINE a comma-boundary regex artefact: '5,026 bp' matched as '026 bp', '1,048 bp' as '048 bp'. Fixed with a negative lookbehind for a digit or comma.",
  "class_B_first_run": "110 hits and useless; bare 'n = 3' is a cohort and '78 samples' a deposition census. Restricted to explicit replicate language: 110 -> 3.",
  "why_recorded": "an instrument reporting the last three digits of a four-digit number as a sub-kilobase width IS the failure class it was written to detect, and this ledger already tracks instruments here finding their own construction before they find a defect. A clean negative is only trustworthy with its self-corrections on the record beside it."
}
what this audit cannot see
[
  "a figure quoted with NO width stated at all - probably the commonest form of class A, and invisible by construction",
  "an entry that discusses a bin size in prose will match the resolution-OK pattern and be cleared even if its figures came from a different unchecked track",
  "a replicate count implied by structure rather than stated as a number"
]
How this could bite us back
  • This is a quality audit and nominates nothing. Its only clinical relevance is negative: it does not license any additional confidence in figures elsewhere in this graph, because it can only see the forms of the mistake it was built to pattern-match, and it was wrong about nearly everything on its first two runs.
Still open
  1. Is 154 GEO-citing entries the right denominator for class A? A sub-bin figure could appear in any entry quoting a genomic measurement, whether or not it names an accession, and restricting to GEO co-occurrence was a convenience rather than a justified scope.
  2. How would a figure quoted with no stated width be caught at all? That is the commonest likely form of this failure and no instrument here can see it. The only reliable defence is recording the bin size in the analysis output at the time of measurement, which is now done prospectively but cannot be applied retrospectively.
  3. Do any entries quote a magnitude from a RATIO track alongside one from a raw-coverage track? That is a third class from the same cycle - the 4.7x element CTCF from ratio tracks is not comparable with 50x from raw coverage - and it was not audited here.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    A regex triage over all 2,103 entries found 154 naming a GEO accession, 10 class-A candidates (a sub-kilobase figure with no bin size recorded anywhere in the entry) and 3 class-B candidates (an explicit replicate count with no file-identity check recorded). Every one of the 13 was read. None is a defect: the 147 bp figures are canonical KMT2A exon 11's length rather than measured windows; the two 100 bp figures are Krivtsov 2019's own stated tiling parameter, correctly attributed; the two Gothe BLISS entries cite deposits with per-position UMI counts where sub-base-pair resolution genuinely exists; the one genuinely unobtainable design was already corrected in cycle 59; and all three class-B hits state replicate structures that are correct for their arms, with no entry asserting a replicate count for the CTCF arm where the four-files-under-eight-accessions duplication actually sits.

    What this evidence cannot show

    This is a REGEX TRIAGE and both false positives and false negatives are expected, which is why the count is presented as a reading list and every hit was read rather than tabulated. The audit cannot see the commonest likely form of class A - a figure quoted with no width stated at all - and it clears any entry that mentions a bin size in prose even if its figures came from a different, unchecked track. 'Defects found: 0' is therefore a statement about what this instrument can detect, not about the corpus. The citation is the script and its output rather than an external source, because the claim is about this repository.

    Evidence that argues against this (2)
    • The audit's own history argues against trusting its negative. Its first run produced 10 class-A hits of which nine were an artefact of its own regex, and 110 class-B hits that were almost entirely cohort sizes and deposition censuses - so on its first two attempts this instrument was wrong about nearly everything it reported. An instrument that needed two corrections to stop reporting noise is weak evidence that there is no signal.
    • A clean negative is the outcome most likely to be a measurement failure rather than a fact, and this project has a specific record of that: the ledger notes an instrument that flagged 1,717 titles as truncated and was wrong about essentially all of them, and several 'zero mismatch' results that hid the bad case by leaving it out of the denominator. The denominator here is 154 GEO-citing entries, and whether that is the right denominator for class A - which could affect any entry quoting a genomic figure, GEO-cited or not - was not tested.
Connected to