The named-but-unwritten queue's top two entries were a web API and a survival endpoint, and MYC was the real answer all along
also called ledger item J · symbol collision registry · REST EFS PDF collision
A tool in this project lists molecules the graph mentions often but has never written up, ranked by how often they appear, and its top two answers were REST and EFS - but in every case checked the graph was saying 'the REST web interface' and 'event-free survival', the commonest survival statistic in childhood cancer, and both happen to be real gene names too so the tool could not tell the difference. With those and two others blocked, the honest top of the list is MYC, a gene mentioned in a hundred entries and never written up.
The same thing, for a scientist
`check-named-but-unwritten.py` ranks HGNC protein-coding symbols by how many node files name them. Its top-ranked results were **REST (217 files)**, **EFS (136)** and **PDF (17)** - respectively the UniProt/EuropePMC REST API, event-free survival, and the file format. Each is a genuine HGNC symbol, so no lexical rule separates them. A machine-readable registry at `data/symbol-collisions.json` now blocks them, and the script consults it. With blocking on, the queue head is **MYC (100), ACTB (44), PTH (36), PICALM (35), IGH (34), MCL1 (33)** - all genuine, several central.
What the registry is, and why prose was not enough
The collisions were already documented - in docs/coverage.md and in brief rules 62-68, as prose, for humans. The ledger's own wording was that they needed recording 'where an automated pass would read them', and that is the whole of the fix. data/symbol-collisions.json is now machine-readable, every row carries the accession, length and verification date, and check-named-but-unwritten.py loads it and refuses to queue a blocked symbol.
Four classes are represented:
| token | HGNC says | this graph means | files |
|---|---|---|---|
| REST | RE1 silencing transcription factor | the REST web API | 217 |
| EFS | embryonal Fyn-associated substrate | event-free survival | 136 |
| PRC1 | protein regulator of cytokinesis 1 | Polycomb Repressive Complex 1 | 78 |
| PZP | pregnancy zone protein | KMT2A's PHD-Zinc-finger-PHD cassette | 21 |
| peptide deformylase, mitochondrial | the file format | 17 |
EFS is the most instructive. The token is frequent because the graph is doing its job: event-free survival is the primary endpoint of essentially every trial cited here. In a clinical knowledge graph, the vocabulary of the domain collides with gene nomenclature precisely where the graph is densest, so this class of error scales with quality rather than with sloppiness.
The registry also carries the alias traps where the danger is a swap rather than a fabrication - CLL1 (CLEC12A vs COLEC10, 12 aa apart), MLL2/MLL4 (KMT2B vs KMT2D), HRAS1 (NRAS vs HRAS, 0 aa apart) - and the non-gene antigens, which have no symbol at all and so can never be queued.
What was underneath
With blocking on, the queue head is finally worth reading:
| symbol | files | why it matters here |
|---|---|---|
| MYC | 100 | the canonical downstream oncogene; named in a hundred entries, no entry of its own |
| ACTB | 44 | mostly a DepMap and qPCR control - real usage, arguably not a gap |
| PTH | 36 | genuine: the revumenib label reports raised intact PTH in 34% of patients |
| PICALM | 35 | a fusion partner in its own right |
| IGH | 34 | genuine: clonal IGH/IGL sequences in lineage-switch tracking |
| MCL1 | 33 | a real, druggable dependency |
| HOXA10, PBX3 | 23 each | core HOXA-programme genes, the heart of this disease |
MYC at 100 files is the finding. This is the sixth instance of the project's recurring shape - the mechanism gets written and the proteins named inside it do not - and it was hidden behind two artefacts for as long as the queue has existed.
A caveat on ACTB that the registry deliberately does not encode: it is a real gene genuinely named 44 times, but almost always as a normalisation control. It is not a collision and is not blocked; whether a loading control deserves an entry is an editorial question, not an identifier one, and conflating the two would corrupt the registry.
- Anyone using the named-but-unwritten queue to plan work before this fix would have been sent to write gene entries for a web API and a survival endpoint. The queue is a work-planning instrument, so its top ranks are the part that actually gets acted on.
- Blocking is silent suppression. If REST ever becomes biologically relevant here, the registry will hide it and nothing in the report will say so - that is a known, accepted cost recorded in E1's counter-evidence, not an oversight.
- MYC is named in 100 entries with no entry. Write it.
- Has a genuine biological use of REST or EFS been suppressed by blocking? A one-off full scan of those 353 occurrences would settle it.
- Should the registry flag rather than block, so suppression is visible in the report?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
The named-but-unwritten queue's three highest-ranked symbols were REST at 217 node files, EFS at 136 and PDF at 17, and in every occurrence inspected these denoted the REST web API, event-free survival and the Portable Document Format respectively rather than the HGNC genes of those names; with the collision registry blocking them the queue's top entry is MYC at 100 files.
data/symbol-collisions.json and the regenerated docs/named-but-unwritten-report.json · Oncofusion knowledge graph (internal) 2026What this evidence cannot show
Run 2026-09-20 against 1,470 entries. **Occurrences were sampled, not exhaustively read**: for each token the first matching context in several entries was inspected, so a genuine biological use of REST or EFS somewhere in the graph would have been missed, and blocking the token would then hide a real gap. **The registry is hand-curated and therefore incomplete** - it records the collisions found, not the collisions that exist, and the queue below the head has not been audited this way.
Evidence that argues against this (1)
- Blocking a token is a destructive fix: REST is a real neuronal transcription factor with published roles in cancer, and if this graph ever acquires a genuine REST claim the registry will suppress it silently. A softer design would flag rather than block and require a human to dismiss. The current choice trades that risk for a queue whose top is readable, and the risk is recorded here rather than in a comment nobody reads. Oncofusion knowledge graph (internal) 2026
THIRD ARTEFACT OF THE SAME SHAPE: 'ALL' read as a gene symbol generated most of the screen's 154,350 first-version candidates, after REST and EFS did the same to the unwritten queue.
SAME WEEK, SAME SHAPE: a token classifier cannot tell a gene symbol from an acronym, which is why REST, EFS and PDF ranked at the top of the other instrument.
THE GAP THAT WAS HIDDEN: MYC at 100 files became the queue's top entry only once REST, EFS and PDF were blocked.
FOURTH INSTRUMENT THIS MONTH TO OVER-FLAG: ALL as a gene symbol, scoping language as negation, an accession invented by a fallback, and now a shared DOI prefix.
ITS TOP THREE WERE ARTEFACTS: REST (217) is the web API, EFS (136) is event-free survival, PDF (17) is the file format. All three are real HGNC symbols, so no lexical rule separated them. Blocked via data/symbol-collisions.json; the honest top is now MYC at 100.
- supports →The named-but-unwritten queue is keyed on gene symbols, so CD15 is invisible to it at 23 entries
THE OTHER HALF OF THE SAME INSTRUMENT PROBLEM: that analysis found what the queue cannot see, this one finds what it wrongly sees. The registry now carries both.
- related to →KMT2A / MLL1 — the gene that breaks
PZP IS NOT A GENE HERE: 21 entries use PZP for KMT2A's PHD-Zinc-finger-PHD cassette, while HGNC assigns it to pregnancy zone protein.
SEVENTH INSTRUMENT THIS MONTH TO OVER-FLAG. Here the first version treated fusion NAMES as species markers and reported a mouse marker in 55% of items; the real figure is 21%.