Oncofusion

start / graph / ana-no-pi-axis-is-thin-and-the-type-fields-understate-coverage-by-a-median-of-twenty-fold

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

Measuring every PI-named axis structurally and by content: the type fields understate coverage by a median of 19.5x and by up to 81x, no axis falls below 18% of the corpus, and 'which axis is thinnest' is retired as a beat selector because it reports a lower bound as though it were the truth

also called no PI axis is thin · the type fields understate coverage 20-fold · thinnest axis retired as a selector · structural versus content coverage

This project files each entry under a category label. For several rounds the coordinator chose what to work on next by counting those labels and picking the rarest. Counting the actual text instead shows the labels miss most of the relevant work - by twenty to eighty times - so every topic the lab cares about is already covered far more than the labels suggest. The method of picking work by counting labels has been retired.

The same thing, for a scientist

**Statement.** For several cycles beats were selected by asking which of the PI's named axes was thinnest, measured by `type` field or id prefix. **Cycle 113 dispatched an agent on that basis and the agent returned the correction**: `type == "analogy"` gave 17 of 2,552 entries, while a `comparative/` folder held **68** cross-family entries of which exactly **one** carried that type. This measures the same thing for every axis at once. | axis | structural | content | % of corpus | understated | |---|---:|---:|---:|---:| | cell type | 32 | **1,839** | 71.7% | **57.5×** | | cell state | 32 | 624 | 24.3% | 19.5× | | subcellular / subnuclear localization | 50 | 903 | 35.2% | 18.1× | | epigenetics | 22 | **1,782** | 69.5% | **81.0×** | | analogous biology, other fusion families | 24 | 467 | 18.2% | 19.5× | | interactors of the fusion | 313 | 1,059 | 41.3% | **3.4×** | **`T2` fires: the type fields are not a coverage measure.** Median understatement **19.5×**, worst **81×**. **No PI-named axis falls below 18.2% of the corpus by content.** **The one axis that behaves is the diagnostic.** *Interactors* is understated only **3.4×** — and it is the only axis with a maintained structural convention (313 entries under `protein`, `partner_interface` and the `interface-` prefix). **A type field tracks its subject exactly to the extent that somebody has been enforcing it**; elsewhere it records the filing habit of whoever wrote the entry. **Consequence, recorded as a decision: "which axis is thinnest" is retired as a beat selector.** It reports a lower bound as though it were the truth. The ledger already says the live queue is the newest king log's *what is next* section, which names accessions and open questions rather than categories; that remains the selector.

Go deeper
Why this was worth measuring rather than just fixing the one case

Cycle 113's error could have been closed by noting that the comparative/ folder exists. That would have left the method intact and the next axis would have failed the same way — epigenetics is understated 81× and would have been the next candidate for a 'cold axis' dispatch on a corpus where 69.5% of entries discuss it.

The general shape: a categorical field is an authoring choice made one entry at a time with no enforcement. Counting it feels like measuring coverage because it yields a clean number over the whole corpus, which is exactly what makes it attractive when hunting a gap. The same failure has now appeared three times in this thread in different clothes — a checker that walked a filename while reporting on a field, a gate that put the stratifier in its own numerator, and this.

What this does not say

It does not say the corpus is saturated. A content regex counts entries that mention an axis, not entries about it, so the content column is an upper bound exactly as the structural column is a lower one. The honest claim is only that the truth lies between them — which is already enough to retire a selector that reported the lower bound alone.

And it says nothing about the literature. An axis can be saturated here and thin in the world. That is a different question, it is the one the PI's priority actually poses, and answering it needs a comparison against an external inventory of KMT2A::AFF1 biology rather than against this corpus's own filing.

Facts and figures
questiondoes the type field measure the PI's axes, or only how entries were filed?
corpus entries2565
median understatement19.5
worst understatement81
the diagnosticinteractors is understated only 3.4x and is the only axis with a maintained structural convention - a type field tracks its subject exactly to the extent someone enforces it
DECISION RECORDED'which axis is thinnest' is retired as a beat selector. The live queue is the newest king log's what-is-next section, which names accessions and open questions rather than categories.
what this does not saythat the corpus is saturated - the content column is an upper bound just as the structural column is a lower one - and nothing about the literature
absence classmeasured-not-curated
bite backThis entry retires a selector and replaces it with a hand-written queue that has no measurement behind it. If that queue drifts toward whatever the last agent found interesting, coverage will rot in exactly the axes this entry declares healthy, and the content counts here - which are upper bounds - will be quoted as evidence that it did not.
axes
epigenetics22.0epigenetics: 22analogous biology other f…24.0analogous biology other fusion families: 24cell type32.0cell type: 32cell state32.0cell state: 32subcellular/subnuclear lo…50.0subcellular/subnuclear localization: 50interactors of the fusion313.0interactors of the fusion: 313
axes — the numbers behind the chart
{
  "cell type": {
    "structural": 32,
    "content": 1839,
    "pct": 71.7,
    "understated": 57.5
  },
  "cell state": {
    "structural": 32,
    "content": 624,
    "pct": 24.3,
    "understated": 19.5
  },
  "subcellular/subnuclear localization": {
    "structural": 50,
    "content": 903,
    "pct": 35.2,
    "understated": 18.1
  },
  "epigenetics": {
    "structural": 22,
    "content": 1782,
    "pct": 69.5,
    "understated": 81
  },
  "analogous biology other fusion families": {
    "structural": 24,
    "content": 467,
    "pct": 18.2,
    "understated": 19.5
  },
  "interactors of the fusion": {
    "structural": 313,
    "content": 1059,
    "pct": 41.3,
    "understated": 3.4
  }
}
Still open
  1. Should the type fields be ENFORCED rather than abandoned? Interactors at 3.4x shows a maintained convention works. A validator rule requiring an axis tag would make the field usable at the cost of retrofitting 2,565 entries, and this entry does not weigh that.
  2. Is any axis thin against the LITERATURE rather than against this corpus? That is the question the PI's priority actually poses and it needs an external inventory of KMT2A::AFF1 biology to answer.
  3. Replace the content regexes with something defensible. They are the whole measurement and they are one person's word lists; an entry-level classifier or a sample hand-labelled by a second reader would bound their error.
  4. What replaces the retired selector, with a measurement behind it? The newest king log's queue is a list of what the last cycle found interesting, which is not a coverage argument.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Across 2,565 committed node files, structural counts (type field or id prefix) against content counts (subject regex) are: cell type 32 vs 1,839; cell state 32 vs 624; subcellular/subnuclear localization 50 vs 903; epigenetics 22 vs 1,782; analogous biology in other fusion families 24 vs 467; interactors of the fusion 313 vs 1,059. The median understatement factor is 19.5x and the maximum 81.0x. Every axis is present in at least 18.2% of the corpus by content. The least understated axis, interactors at 3.4x, is the only one with a maintained structural convention.

    What this evidence cannot show

    THE CONTENT REGEXES ARE THE COORDINATOR'S AND ARE THE WHOLE MEASUREMENT. A different word list gives different numbers; the lists are printed in the script so they can be disputed. Matching a term anywhere in an entry counts a MENTION, not a subject, so every content figure is an upper bound - 'chromatin' appearing in an entry does not make that entry about epigenetics, and the 69.5% figure is certainly inflated by that. The structural tests are equally crude: id prefixes and type values were chosen by inspection and a mis-specified prefix would understate an axis artificially. An entry can be about several axes so the counts are not a partition and must not be summed. Nothing here measures the literature.

    Evidence that argues against this (3)
    • The conclusion is convenient for me. I made the selection error being analysed, and 'the method was systematically broken' is a more comfortable finding than 'I misread one number'. A sceptic would note that the strongest evidence for the systematic reading - that epigenetics is understated 81x - depends on a regex containing the word 'chromatin', which appears in a great many entries for reasons unrelated to epigenetics as an axis.
    • Retiring the selector removes something and replaces it with a pointer to the newest king log's queue, which is written by the same coordinator that got this wrong. That queue has no measurement behind it at all - it is a list of what the last cycle happened to find interesting. Replacing a bad measurement with no measurement is not obviously an improvement and this entry does not argue that it is.
    • The interactors result cuts against the headline. If a maintained convention gives 3.4x rather than 19.5x, the fix might be to ENFORCE the type fields - a validator rule requiring an axis tag - rather than to abandon them. That would make the field a usable measure going forward, at the cost of work on 2,565 existing entries, and this entry dismisses the option by not raising it.
Connected to