Adjudicating the unwritten-gene queue by claim membership diverges from it at Spearman 0.36 - and then puts four acronym collisions in its own top five, so divergence was not improvement
also called claim-membership queue adjudication · NES RPE CD4 CPM are collisions · a structural test inherits the collision problem
A list of genes the project has mentioned but never written up is ranked by how often each is mentioned. Ranking it instead by whether anyone made a claim about the gene gave a different order - but its top entries turned out to be abbreviations that only look like gene names.
The same thing, for a scientist
Coverage item 6 owed was 'adjudicate the queue's residue symbols by content'. I replaced the prose heuristic with a structural test: does the symbol appear inside an `evidence[].claim` field, whole-word and case-sensitive, rather than anywhere in the entry? Pre-registered: Spearman below 0.6 or >40% zero-claim means the queue ranks the wrong thing and the claim count is the shortlist. **Spearman came in at 0.3557 and only 4 of 150 carry no claim at all** - so the first half fired and the second did not. But the pre-registered conclusion is WRONG, because reading the matching claims shows the new ranking's top five are **NES** (nuclear export sequence), **RPE** (retinal pigment epithelium), **CD4** (a flow-cytometry marker), **CPM** (counts per million) and **RPA2**. Four collisions and one real gene - and the real one is RPA2, which q195 had already identified independently as the only genuine AF4-specific gap. The durable output is not a ranking but four new rows in `data/symbol-collisions.json`.
The test, and why it looked like the right one
The queue's own note says frequency is not importance: 'a protein named in sixty files because it appears in one copied roster is not a gap'. The obvious fix is to ask whether the corpus asserts anything about the symbol, and that is structural rather than heuristic - every entry has an evidence array whose items carry a claim string. No window to tune, no keyword list. Matching was whole-word and case-sensitive specifically to avoid the WAS/was error this project has already made.
It did diverge from the queue, sharply: Spearman(files named, claim-carrying entries) = 0.3557, with only 10 of 20 top-20 members shared.
And then it reproduced the collision problem it was meant to solve
| rank | symbol | files | claim entries | what it actually is |
|---|---|---|---|---|
| 1 | RPE | 67 | 32 | retinal pigment epithelium / hTERT-RPE1 |
| 2 | NES | 48 | 26 | nuclear export sequence |
| 3 | CD4 | 44 | 22 | immunophenotype marker |
| 4 | VHL | 50 | 21 | not adjudicated here |
| 5 | RPA2 | 26 | 20 | a real gene |
| 7 | CPM | 39 | 17 | counts per million |
Verified by reading the matching claim text, verbatim:
- NES: '...RESPONSIVE NUCLEAR EXPORT SEQUENCE at their extreme carboxy termini; the NES is conserved in HDAC7...'
- RPE: '...replicating the direction found in RPE epithelium...'
- CD4: '...pro-/pre-T, CD4/CD8 double-negative, CD7-positive...'
- CPM: '...its RNA-seq baseline (0.39-1.02 log2 CPM)...'
- RPA2: '...colocalisation with RAD51AP1 and RPA2 in multiple nuclear foci on DNA damage...' and '...recruits CtIP as the end-resection effector activating RPA2/RAD51...'
An acronym appears in a claim as readily as a gene does. Claim membership measures load-bearingness, which is a real property, but it is orthogonal to whether the token is a gene symbol at all. My pre-registered statement - 'the claim count is the shortlist' - is withdrawn: it is a differently-wrong ranking, not a better one. CPM is the worst of the four, because a units token lands inside quantitative claims, which is exactly where a reader trusts the text most.
What actually gets fixed
A collision documented only in prose is not documented - this project wrote the SEM trap into an entry rather than the registry every beat greps, and it fired again one cycle later. So the output here is four new rows in data/symbol-collisions.json (216 -> 220): NES, RPE, CD4 and CPM, each with block_autogeneration: true, the colliding meaning, and the verbatim claim text that proves it.
Verified, not assumed: re-running check-named-but-unwritten.py afterwards confirms all four are gone from the queue. The queue length stays at 150 because it is capped and lower-ranked symbols roll up - so the count looking unchanged is expected and is not evidence the fix failed.
And the one real gene in the top five, RPA2, needed no shortlist to find: q195 reached it from the significance side, and it is being written this cycle as Q192. Two independent routes converged on the same single protein, which is a better argument for RPA2 than either ranking.
| queue symbols | 150 |
|---|---|
| entries scanned | 3095 |
| spearman files vs claim entries | 0.3557 |
| top20 overlap | 10 |
| n zero claim | 4 |
| collisions registry | 216 -> 220 rows |
| real gene in top5 | RPA2 - independently identified by q195 and dispatched as Q192 |
| withdrawn conclusion | the pre-registered 'claim count is the shortlist' verdict |
| preregistered | rho < 0.6 or >40% zero-claim => queue ranks the wrong thing |
zero claim symbols
[ "NEB", "FBL", "APEX2", "CAPS" ]
top5 by claims
[ "RPE", "NES", "CD4", "VHL", "RPA2" ]
collisions found
[ "RPE", "NES", "CD4", "CPM" ]
- Is VHL at rank 4 (50 files, 21 claim entries) a genuine gene subject or a fifth collision? Not adjudicated here.
- How many of the remaining 146 queue symbols are collisions? Four were found by reading only the top five, which is a high hit rate and suggests the registry is far from complete.
- Is there a test that separates gene mentions from same-string abbreviations without reading? Neither file count nor claim membership does it, and both failed on the same tokens.
- Do the four zero-claim symbols (NEB, FBL, APEX2, CAPS) indicate anything, or are they simply named in rosters? They are the one group this analysis did not read.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.internal
Ranking the 150-symbol unwritten-gene queue by the number of distinct entries naming the symbol inside an evidence[].claim field, whole-word and case-sensitive over 3,095 entries, gives Spearman 0.3557 against the queue's file-count ranking, shares only 10 of 20 top-20 members, and leaves just 4 of 150 symbols with no claim-carrying entry (NEB, FBL, APEX2, CAPS).
q196 claim-membership adjudication of the 150-symbol unwritten-gene queue · Oncofusion KG 2026What this evidence cannot show
Claim membership is computed from the `claim` field only, not from caveats, counter-evidence notes or section prose, so a symbol argued about in a section but never in a claim string scores zero here. The Spearman is over 150 points including many ties at low counts, which inflates instability in the tail.
Evidence that argues against this (1)
- The low correlation was read as evidence that the queue ranks the wrong thing. It is equally consistent with BOTH rankings being wrong, which is what the collision adjudication then showed. A divergence between two instruments says at least one is wrong; it never says which, and the pre-registration should have required a labelled positive set rather than treating divergence as improvement. Oncofusion KG 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.internal
Four of the five highest-ranked symbols by claim membership are not gene mentions: RPE is retinal pigment epithelium, NES is nuclear export sequence, CD4 is a flow-cytometry immunophenotype marker and CPM is the counts-per-million RNA-seq unit; only RPA2 is a genuine gene subject. All four are now registered in data/symbol-collisions.json with block_autogeneration and verbatim claim text, and re-running the queue tool confirms all four are excluded.
data/symbol-collisions.json, rows 217-220 (NES, RPE, CD4, CPM) · Oncofusion KG 2026What this evidence cannot show
Each verdict rests on reading two matching claims per symbol, not all of them, so a minority of mentions could be genuine gene references - particularly CD4, where the gene and the marker are the same molecule. The CD4 row says only that a queue proposing gene-cd4 is proposing an entry no claim in this corpus needs, not that CD4 biology is irrelevant. VHL at rank 4 was NOT adjudicated and may be genuine.
Evidence that argues against this (1)
- Registering a symbol with block_autogeneration is a destructive act for future sweeps: if any of these four is later the subject of real index-fusion biology, the registry will now hide it, and the entry that would have been written never gets proposed. The rows carry the colliding meaning rather than a blanket veto for that reason, but the hazard is real and is the mirror image of the one being fixed. Oncofusion KG 2026
independently reaches RPA2: it is the only real gene in the claim-membership top five, and q195 reached it from the significance side
same instrument, next defect: after the negative controls were excluded, four of the queue top five are acronym collisions
the entry that makes NES mean nuclear export sequence throughout this corpus, which is why NES ranked second as a gene gap