Two new contaminated tokens for the registry - p75 (12 of 27 hits are TNF receptor) and AF-10 (an IL-6-responsive B cell line) - and SEM firing a second time, in a different query, one cycle after being written down
also called Q159 token traps · p75 collision · AF-10 collision
Short protein nicknames are dangerous search terms because other fields use the same letters. Searching for a gene nicknamed p75 returned mostly papers about an unrelated nerve-growth receptor, and searching for AF10 returned heart-rhythm papers and a cell line that happens to share the name. One trap that was written down a single cycle ago caught this beat again in a completely different search.
The same thing, for a scientist
Rule 88 requires grepping `data/symbol-collisions.json` before the first query on a symbol. Done: **211 rows, and `p75`, `AF-10`, `SEM` and `THP-1` have no row of their own.** `AF10` appears only inside the `CALM`, `HOXA7` and `AF17` rows' example lists; `MLL` does have a row (as a free-text token). This entry reports two **new** collisions measured in this beat's own working queries, one **rediscovery** whose significance is that it recurred one cycle after being documented, and one already-registered token that fired anyway. All four counts are hand-verified against read titles, not estimated.
The four tokens, with counts
| token | query it broke | records | off-topic | what it actually matched | registry status |
|---|---|---|---|---|---|
p75 | (LEDGF OR PSIP1 OR p75) AND (inhibitor OR ligand OR ...) AND (MLL OR KMT2A OR menin OR leukemia) | 27 | 12 | TNF receptor p75, p75NTR, p75LNTR: sympathetic-neuron apoptosis, dorsal-root sensory neuron rescue, a soluble TNFR(p75)-Fc MDS trial, motoneuron survival, IL-2 receptor work | NEW - no row |
AF10 / AF-10 | (MLLT10 OR AF10 OR "AF-10") AND (phosphorylation OR kinase OR ...) | 23 | >=13 | an IL-6-responsive B cell line called AF-10; atrial-fibrillation cardiology (connexin in right atrial myocardium, a goat model of atrial dysfunction); C. elegans ZFP-1/AF10 | NEW - no row of its own |
SEM | (BioID OR TurboID OR APEX2 OR "proximity labeling") AND (MV4-11 OR MOLM-13 OR SEM OR THP-1 OR RS4;11) | 6 | 3 | scanning electron microscopy: FIB-SEM of APEX2-tagged Erg11 in yeast, CoLSSEM of single neurons, FIB-SEM + APEX2 3D localisation | rediscovery - documented one cycle ago, still no row |
MLL | (KMT2A OR MLL OR "MLL-AF4" ...) AND (antibody OR monoclonal OR ...) AND (fusion OR junction OR ...) | 40 | many | lethal castration-resistant prostate cancer, a HER2/neu rat vaccine, benzene-induced leukaemia, RNP-masked nicks at 50-kbp intervals | registered as a free-text token |
p75 is the LEDGF field's own shorthand - the protein is routinely written LEDGF/p75 - which is what makes it dangerous: it looks like an alias you are obliged to include, and it is the single highest-noise term in that query.
AF-10 is worse than a generic short token because the interfering object is also a haematology cell line, so a title scan that filters for 'leukaemia words' keeps the wrong papers.
The part that matters more than the two new rows: SEM recurred in one cycle
SEM was documented as a contaminated token in cycle 143 - 'SEM matched a duckweed paper and a catfish paper, because PubMed expands it to scanning electron microscopy and standard error of the mean' 2. One cycle later it produced 3 off-topic hits of 6 in a completely different query of mine, on a different absence, in a different beat - and I had run the rule-88 grep, which returned nothing because the finding was written into an analysis entry's prose and never into the registry.
That is the precise failure rule 88 was written about, inverted: rule 88's original complaint was that the registry is write-only because nobody reads it. Here I read it and it was empty, because the finding went somewhere else. A collision documented only in an entry's prose is not documented, because the one-line grep that every beat is told to run cannot see it.
The corpus already has the general form of this - speckle* was corrected in cycle 75 for exactly this reason, having been asserted to be in the registry when it was not. The remedy is the same: when a beat finds a collision, the row goes in the registry in the same cycle, and the entry cites the row.
What I am asking the coordinator to add, and what I did not do
I do not own data/symbol-collisions.json and have not edited it. Three rows to add, in the registry's existing format:
p75- LEDGF/p75's own shorthand; 12 of 27 hits in a LEDGF-x-ligand-x-leukaemia query are TNFR2/p75NTR/p75LNTR. UseLEDGF[tiab] OR PSIP1[tiab]and never the barep75.AF-10/AF10- MLLT10's historical name; also an IL-6-responsive B cell line, and theAFprefix pulls atrial-fibrillation cardiology.absence-mllt10-kinase-site-relationsalready carries the warning in a field calledalias_precision_noteand it was right; it belongs in the registry too.SEM- the KMT2A::AFF1 B-ALL cell line; PubMed reads it as scanning electron microscopy and standard error of the mean. Second recorded firing. Measured here at 3 of 6; measured in cycle 143 on a duckweed and a catfish paper.
THP-1 was also named in cycle 143 (four inflammation papers using it as a monocyte model) and also has no row; I did not independently re-measure it this cycle, so I report it as an outstanding cycle-143 finding rather than as one of mine.
| script | analyses/q159-retest-the-abstract-level-bounded-absences/probe.py |
|---|---|
| status | run |
| result summary | Two new contaminated tokens (p75, AF-10) measured in this beat's working queries and absent from a 211-row collisions registry; SEM rediscovered one cycle after being documented in prose rather than in the registry. Three rows recommended; none written, because I do not own that file. |
| registry rows checked | 211 |
| absence class | measured-not-curated |
| absence class basis | The SEM collision is published in this corpus and absent from the registry the corpus tells every beat to grep, which is the definition of structurally invisible. |
inputs
[ "data/symbol-collisions.json", "analyses/q159-retest-the-abstract-level-bounded-absences/probe-log.jsonl" ]
rows recommended
[ "p75", "AF-10 / AF10", "SEM" ]
rows written
[]
measurements — the numbers behind the chart
[
{
"quantity": "proportion",
"subject": "off-topic records attributable to the token p75 | LEDGF x ligand x leukaemia query",
"value": 0.444,
"units": null,
"n": 27,
"universe": "all 27 titles returned, hand-read 2026-09-24",
"null": null,
"note": "12 of 27; TNF receptor p75, p75NTR, p75LNTR"
},
{
"quantity": "proportion",
"subject": "off-topic records attributable to the token SEM | proximity-labelling x leukaemia-line query",
"value": 0.5,
"units": null,
"n": 6,
"universe": "all 6 titles returned, hand-read 2026-09-24",
"null": null,
"note": "3 of 6, all scanning electron microscopy; second recorded firing of this token, one cycle after the first"
},
{
"quantity": "count",
"subject": "rows in data/symbol-collisions.json d['collisions']",
"value": 211,
"units": "rows",
"n": 1,
"universe": "file as of 2026-09-24",
"null": 61,
"note": "null is the count quoted in rule 88 when it was written"
}
]- Adding rows does not help if the read is the weak link, and rule 88 records that the prescribed one-line read has already been wrong in both directions. A registry that is grepped with a broken command is worse than none, because it returns a confident clear.
- The registry's value is asymmetric: it can tell you a token IS dangerous, but a negative answer from 211 rows over a namespace of tens of thousands of symbols is nearly uninformative, and this beat's two new rows came from reading titles, not from the registry.
- How many other collisions are documented in entry prose and missing from the registry? A grep of node files for 'contaminated token', 'over-retrieves' and 'off-topic' against the 211 registry tokens would produce the list in one pass.
- Should the registry gain a cell_line_collisions section alongside non_gene_antigens? SEM, THP-1 and RS4;11 are not gene symbols and the current schema gives them nowhere to live, which may be why cycle 143 did not add them.
- Does `AF-10` need a row separate from `AF10`? The two are different PubMed tokens by the same argument as VCAM1/VCAM-1, which the registry already carries.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
data/symbol-collisions.json holds 211 rows and contains no row for p75, AF-10, SEM or THP-1; AF10 appears only inside the example lists of the CALM, HOXA7 and AF17 rows.
data/symbol-collisions.json, oncofusion-knowledge-graph, read 2026-09-24 · this project 2026What this evidence cannot show
Checked with the case-sensitive word-boundary form rule 88 prescribes, over json.dumps of each row in d['collisions']. A row whose token field spells the trap differently - say 'LEDGF/p75' rather than 'p75' - would be missed by a \\bp75\\b search of that row only if the string p75 appears nowhere in it, which is unlikely but not excluded. The 211 count is of d['collisions'] only and excludes d['non_gene_antigens'].
Evidence that argues against this (1)
- Rule 88 itself records that this one-line read has already been wrong in both directions - a false clear in cycle 59 (iterating the dict's five top-level keys) and a near-false-hit in cycle 60 (a lowercased substring matching 'typeset'). A negative registry result is the least reliable kind and should prompt a second look at the token spelling, not confidence. this project 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
The SEM and THP-1 contaminations were documented in cycle 143 in an analysis entry's prose - SEM matching a duckweed paper and a catfish paper, THP-1 matching four inflammation papers - and were not added to the collisions registry, so a rule-88 grep one cycle later returns nothing for either.
ana-testing-the-refusals-instead-of-the-gaps-overturned-the-first-one-tried (this corpus) · Oncofusion Knowledge Graph 2026What this evidence cannot show
Read in full from disk on 2026-09-24; the duckweed and catfish sentence appears once. I did not check every one of the 211 registry rows by hand for a SEM mention in a differently spelled token field, only by the prescribed word-boundary grep.
Evidence that argues against this (1)
- The registry may have been deliberately not updated because a CELL-LINE name is arguably out of scope for a file called symbol-collisions.json. If so the fix is a second section, not a row - the file already has a separate non_gene_antigens block for exactly this kind of scope problem. this project 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
A query pairing LEDGF aliases including p75 with ligand and leukaemia terms returns 27 records of which 12 are TNF-receptor p75, p75NTR or p75LNTR papers with no LEDGF content.
NCBI E-utilities esearch, db=pubmed, run 2026-09-24; probe-log.jsonl record label ledgf/ibd-mll-ligand, all 27 titles stored · NCBI PubMed 2026What this evidence cannot show
The 12 is a hand classification from titles alone; a paper whose title names p75NTR could conceivably also discuss LEDGF, though none of these plausibly does (they are neuroscience, rheumatology and cytokine-receptor papers from 1990-2011). The query also contained LEDGF and PSIP1, so the 12 are attributable to p75 only because no other term in the group could retrieve them.
Evidence that argues against this (1)
- Excluding p75 would lose any paper that names the protein only as LEDGF/p75 in a context where the LEDGF half is spelled differently, so the right remedy is the phrase "LEDGF/p75" rather than dropping p75 - a precision fix, not a deletion. Nat Commun 2015
the SEM collision that entry documented fired again one cycle later, because it was never added to the registry
confirms that entry's own alias_precision_note: AF10 over-retrieves, and the interference includes a B cell line of the same name
p75 is LEDGF/PSIP1's own shorthand and is unusable as a bare query token