RETRACTED: 'Sixteen percent of GEO sample labels contradict their own metadata' was an off-by-one in my own parser - GEO's brief format puts the title before the accession, so every sample was scored against the next sample's title; the true rate is ZERO, the deposit named as worst offender has none, and the same bug made me 'correct' an agent whose accessions were right

# RETRACTED 2026-09-24 (cycle 121). THE 16.4% WAS A BUG IN MY OWN PARSER, NOT A PROPERTY OF ANY DEPOSIT. **The true rate is ZERO — 0 of 666 comparable pairs. GSE159608 — reported below as the worst offender at 46 of 119 antibody and 9 of 18 treatment contradictions — has ZERO.** **The error.** In GEO's brief format `!Sample_title` comes *before* `!Sample_geo_accession`: ``` !Sample_title = ALL1_Dot1L_AutoCUTnTag_(...) <- line 2 !Sample_geo_accession = GSM4835651 <- line 3 !Sample_characteristics_ch1 = ... antibody: Dot1L <- line 12 ``` My parser opened a new record only on the accession line, so **every sample was given the NEXT sample's title and compared against its own characteristics.** That manufactures a uniform one-record lag — and I reported that lag as the deposits' signature defect. **Every specific example below is an artefact.** `GSM5139712` is not titled VTP against EPZ-5676 characteristics; `GSM4835705` is not titled `KOPN8_H3K27me3` against *IgG* and *K562*; `GSM4835659` is not H3K4me3 against H3K4me1. In each case the title shown belonged to the following sample. **The manual check confirmed the bug to itself.** I read three mismatches 'in full' to satisfy `U2`, using an `awk` that began at the accession line and ran past the next title — **the same off-by-one**. A verification that shares the defect it is verifying is not a verification, and this one produced the confident sentence *"Real contradictions, not tokeniser artefacts"*. **`U2` named the category and I checked the wrong axis.** It says *"a mismatch I manufacture by failing to normalise is my defect, not the deposit's"* — so I checked **normalisation** and never checked **record boundaries**. The clause was right; my reading of it was too narrow. **Found by the q140 agent**, dispatched this cycle on a brief that asserted this entry's claim as established fact. It refused the premise and tested the parser. **That is four of the last six briefs containing a false premise that the agent caught**, and the first where the false premise was a published finding of mine rather than a ledger line. **FINAL FIGURE: ZERO.** After the parser was fixed the count fell to 7 of 656, and reading all seven individually showed **every one was this script's own tokeniser**, not a deposit: | apparent mismatch | what the raw record says | |---|---| | 4x `MLL1C` in title vs `MLL1` in characteristics | title `HCB1_ChIP_MLL1c_DMSO_REP1`, characteristics `antibody: MLL1 (c terminus)`, Active Motif 61296 — **the same antibody** | | 3x `EPZ5676` in title vs `DMSO` in characteristics | title `SEM_0umEPZ5676_Rep1`, characteristics `treated with: DMSO for 7 days` — **0 µM IS the vehicle arm** | Both normalisations were added and the final count is **0 of 666 comparable pairs across 32 series and 2,312 samples**. **The deposits agree with themselves completely.** 🟠 **Stated against my own interest:** those two normalisations were written *after* seeing which cases failed, which is a degree of freedom I did not control. Both are unambiguous on the raw records — a zero dose is a vehicle and a C-terminal MLL1 antibody is MLL1c — but a reader should know the sequence was 16.4% → 1.1% → 0% with me editing the tokeniser between the last two steps. 🔴 **AND THE SAME BUG MADE ME 'CORRECT' AN AGENT THAT WAS RIGHT.** In cycle 120 an agent reported GSE159608's SEM arms as **VTP `GSM5139713/14` against DMSO `GSM5139709/10`**. Using the shifted parse I wrote that it had *"two of its four sample accessions wrong"*, and put that into the entry's caveats, the king log and the commit message. **The agent's accessions were exactly right and mine were shifted by one.** The true map, read from raw line numbers: | arm | accessions | |---|---| | H3K4me3 DMSO | `GSM5139709`, `GSM5139710` | | H3K4me3 EPZ-5676 | `GSM5139711`, `GSM5139712` | | H3K4me3 VTP-50469 | `GSM5139713`, `GSM5139714` | | Menin (no drug arm) | `GSM5139719`, `GSM5139720` | **A parser defect that only shifted numbers also inverted a judgement about who was reliable.** That is the more expensive half of this error, and it is the half that had no pre-registered clause. **What survives:** the `U5` independence test (title and embedded filename agree 156/156 and are therefore one source, not two) is a comparison between two fields of the *same* record and is unaffected. The 7 residual mismatches are real and are listed in the corrected `results.json`. **What this reverses downstream:** GSE159608's SEM `DMSO` / `EPZ-5676` / `VTP-50469` arm labels **can be taken at face value**, so the cycle-120 correction to `absence-no-menin-inhibitor-occupancy-dataset-in-an-infant-derived-kmt2a-aff1-line` is no longer bounded to *the dataset exists*. --- **THE ORIGINAL ENTRY FOLLOWS, UNCHANGED EXCEPT FOR THIS BANNER, BECAUSE THE REASONING IS THE POINT.** **Statement.** A GEO sample title is free text, it is the label every downstream analysis keys on — including this project's, in cycle 116 — and nothing compares it to the sample's own `characteristics`. Comparing them across the 32 cached series, 2,312 samples: | | comparable | disagree | |---|---:|---:| | antibody | 252 | **62 (24.6%)** | | treatment | 334 | **34 (10.2%)** | | **total** | **586** | **96 (16.4%)** | **Pre-registered `A2` fires.** 14 series contradict themselves in a way not explained by `U1`'s pasted-template artefact, which accounted for 2 of 32. **The contradictions are not near-misses.** `GSM4835659` is titled `ALL1_H3K4me3_AutoCUTnTag` and its characteristics read *"cut&tag antibody: H3K4me1 Thermo cat# 710795"* — a different antibody with a catalogue number. `GSM4835705` is titled `KOPN8_H3K27me3` against *"antibody: IgG"* and *"cell type: K562"* — **wrong antibody and wrong cell line**. `GSM5139712` is titled `SEM_H3K4me3_VTP50469` against *"SEM ... with 30 uM EPZ-5676 for 72hrs"* — **the two drugs whose separation this project spent cycles 83 through 117 establishing.** **GSE159608 is the outlier and it is the one this cycle needed**: 46 of 119 antibody and 9 of 18 treatment assignments contradict, 55 of the 96 total. Excluding it the rate falls to roughly 9%, so *"GEO is 16% unreliable"* is the wrong generalisation — **some deposits are badly inconsistent and the only way to know is to check the one in front of you.** **THE PART I EXPECTED TO RESCUE THIS, AND DID NOT.** Each title carries a parenthesised original filename in a third vocabulary — `(20200324_DJ_Hs_K4m3_TB13_0323)`. It sides with the title **39 times to 0**, which reads as decisive vindication of titles. **`U5` was pre-registered against exactly that reading and it fires: title and filename never disagree with each other across 156 comparable pairs.** A field that never contradicts another is a field derived from it. **The third vote is one vote counted twice, and it adjudicates nothing.** **Consequence.** GSE159608 does contain SEM CUT&Tag under a menin inhibitor, which refutes one sentence in `absence-no-menin-inhibitor-occupancy-dataset-in-an-infant-derived-kmt2a-aff1-line`. **But which sample is the VTP arm cannot be decided from metadata**, so the refutation is bounded to *the dataset exists*. Its narrower claim — no menin **occupancy** readout under drug in an infant line — survives, because the SEM menin samples are DMSO-only.