Oncofusion

start / graph / hyp-q181-a-two-state-coverage-metric-cannot-express-not-applicable

Hypothesismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q181-epigenetics-dimension

Every PI-dimension coverage line in this project is two-valued, and a third of what it counts as missing is a state the metric has no way to express - so the epigenetics figure will not improve by filling fields, and should not be made to

also called q181 recommendation · three-state coverage

The score for each topic has only two settings: recorded, or not recorded. But there are three real situations - we know it, nobody has measured it, and there is nothing to know because the protein never touches chromatin. Lumping the last two together is why the chromatin score looks so bad, and filling in fields would hide the problem rather than fix it.

The same thing, for a scientist

Hypothesis: the epigenetics line's 39.2% is not a statement about missing knowledge but about a metric whose codomain is too small. Three states exist in the corpus - RECORDED, UNMEASURED (bounded absence), NOT APPLICABLE (category statement) - and this project has already evolved vocabulary for all three on the localization dimension (124 nodes), while the metric collapses the last two into 'uncovered'. This beat's triage estimates that ~33% of the epigenetics shortfall is NOT APPLICABLE and ~48% is content present under an unrecognised field name, leaving ~19% as a real gap. The prediction that follows is falsifiable and cheap: if the metric were three-valued and read data VALUES as well as key names, the epigenetics line would move from 39.2% to roughly 85-90% without a single new literature fact being added - and the residual, ~11% of the population, would be the list worth dispatching a beat on. If instead the line moved only a few points, the triage is wrong and the gap is real.

Go deeper
The claim, stated so it can be scored
  1. check-pi-dimensions.py is three-valued in its subject matter and two-valued in its output. Every one of the five PI dimension lines prints covered/not-covered.
  2. The corpus already writes the third state and does so on the localization dimension only -- 124 nodes, with two entries explicitly distinguishing 'a bounded absence' from 'a NOT APPLICABLE' [ana-q181-the-not-applicable-convention-exists-for-localization-and-not-for-epigenetics].
  3. Therefore the right response to a low epigenetics figure is not to fill fields. Filling an epigenetic_readout on gene-fut9, a Golgi fucosyltransferase, would raise the number and degrade the graph. This is the failure mode the dispatch that produced this beat named in advance and it is real: 108 of the 329 uncovered entries are estimated to be in that class.
  4. And the residual matters. ~63 entries, and a sampling-free floor of 8 chromatin-machinery entries, are a genuine gap. That list is small enough to be a beat and is named in ana-q181-the-only-chromatin-machinery-with-genuinely-nothing-is-the-ptefb-brake.
The recommended change, which has NOT been made

scripts/check-pi-dimensions.py is unmodified. This is a proposal for the coordinator, not an applied patch, and the reason is stated in the dispatch: changing a measurement tool to improve the number it reports would make this beat worthless.

Schema side (SCHEMA.md, which this beat also did not edit):

  • Add epigenetics_not_applicable as a recognised spelling, whose value must be a reason, not a boolean -- "single-pass Golgi membrane enzyme; no nucleosome contact in any published system".
  • Keep epigenetic_bounded_absence for the distinct unmeasured state, and say in the schema that the two are different.

Reporter side:

  • Print three numbers per dimension -- recorded / declared-inapplicable / neither -- and treat only the third as the gap.
  • Add an epi_value line reading _EPI_MARK over data{} values only, printed separately from epi_field, never merged.
  • Print the near-miss key list so a new spelling is caught the cycle it appears.

Process side, and this is the one that generalises: the metric's current shape means every dimension figure in the ledger is an underestimate by an unknown amount, because none of them can express inapplicability. Localization at 77.4% and cell type/state at 59.7% have never been triaged the way epigenetics has here.

Why this might be wrong

The strongest objection is that it defines the problem away. A three-state metric lets an author discharge a dimension by declaring it inapplicable, and nothing audits that declaration. If NOT APPLICABLE becomes cheap, the epigenetics line becomes a measure of how many authors used the escape hatch. The localization dimension has carried the convention for several cycles without visible abuse, but 124 nodes is a small sample of practice and nobody has spot-checked whether any of those declarations is wrong.

The second objection is the one recorded in this beat's own parent analysis: the PI dimension may mean "which mark, at which loci, in which cell state" rather than "is this a chromatin protein". Under the strict reading, most of what this beat classifies as schema artefact is a real gap, the recommendation is wrong, and the 39.2% is roughly right for the wrong reason. Nothing in the standing instruction distinguishes the two readings, and this hypothesis does not settle it.

Facts and figures
statementThe epigenetics coverage figure is limited by the metric's two-valued codomain rather than by missing knowledge; roughly a third of what it scores as missing is a NOT APPLICABLE state the metric cannot express, and roughly half is content present under a field name the metric does not recognise.
origindata-pattern
prior probability0.7
predictions
[
  "A three-state reporter that also reads data{} VALUES would move the epigenetics line from 39.2% to roughly 85-90% with no new literature fact added.",
  "The residual after that change is ~11% of the 541-entry population, and the chromatin-machinery members of it are the 8 named in ana-q181-the-only-chromatin-machinery-with-genuinely-nothing-is-the-ptefb-brake.",
  "The same triage applied to the localization (77.4%) and cell type/state (59.7%) lines would find a smaller but non-zero inapplicability component, because those dimensions ALREADY carry the NOT APPLICABLE convention and epigenetics does not.",
  "If an epigenetics_not_applicable field is added with no audit requirement, the share of entries using it will exceed this beat's 33% estimate within three cycles - the escape-hatch failure mode."
]
experiments
[]
analyses
[
  "ana-q181-the-thinnest-pi-dimension-is-half-a-notation-problem",
  "ana-q181-the-not-applicable-convention-exists-for-localization-and-not-for-epigenetics",
  "ana-q181-the-only-chromatin-machinery-with-genuinely-nothing-is-the-ptefb-brake",
  "ana-q181-rescore-the-epigenetics-line-three-ways-before-anyone-fills-a-field"
]
falsifiers
[
  "Re-running check-pi-dimensions.py over the same 541 entries with (a) _EPI_MARK applied to data{} values and (b) a declared-inapplicable state, and finding the recorded-or-inapplicable total BELOW 70% rather than the predicted 85-90%. That would mean the content is not there and the gap is real. Scoreable on the corpus as of commit ac2aeab; the threshold is 70% of 541 = 379 entries.",
  "A second reader re-scoring the persisted seed-181 samples and returning a stratum-B subject-relevance rate below 8/20 (40%), against my 15/20. That collapses the schema-artefact share below a third and makes 'fill the fields' the right response after all.",
  "Reading all eight hard-floor entries end to end and finding chromatin content in five or more of them. That would mean even the sampling-free floor is notational and the real gap is near zero - which would REFUTE this hypothesis in the direction of it being too conservative, and is the more likely of the two failure directions.",
  "A spot-check of 20 of the 124 existing NOT APPLICABLE localization declarations finding 4 or more that are wrong - i.e. the protein does have a published subnuclear address. That would show the third state is already being abused on the dimension that has it, and the recommendation to extend it to epigenetics should be withdrawn."
]
How this could bite us back
  • A three-state metric gives every author an escape hatch that nothing audits. If declaring inapplicability is cheaper than researching a mark, the epigenetics line stops measuring knowledge and starts measuring how often authors reached for the hatch - and it will look like it improved while getting worse. Any adoption should pair the new field with a required free-text reason and a periodic spot-check, which is a real cost this recommendation is imposing on someone.
  • Raising a coverage figure from 39% to ~88% by redefinition invites a reader to conclude the dimension is done, and a number that closes a line of work is precisely the one Rule 90 says needs the highest duty of reproducibility. If the ledger absorbs '88%' without the residual list beside it, this beat will have suppressed work on the PI's own priority exactly as the stale interactor figure did for seven cycles.
  • The residual gap this recommendation isolates - the P-TEFb reserve/release arm - is the part of elongation control a menin-inhibitor programme would most want annotated if resistance runs through P-TEFb availability. Declaring the dimension nearly complete while that specific block is empty is the worst possible place for the false reassurance to land.
Still open
  1. Which reading of the epigenetics dimension did the PI intend - 'is this a chromatin protein' or 'which mark, where, in which cell state'? Under the strict reading this hypothesis is wrong and the 39.2% is approximately right.
  2. Has any of the 124 existing NOT APPLICABLE localization declarations ever been audited? If the third state is already unaudited on the dimension that has it, extending it to epigenetics extends an unverified practice.
  3. Do the localization (77.4%) and cell type/state (59.7%) figures have the same notation and inapplicability components? Neither has been triaged, and both are quoted in the ledger.
Every source behind this entry (2)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    check-pi-dimensions.py prints every PI dimension as a two-valued covered/not-covered proportion with no state for inapplicability, while 124 nodes in the same corpus carry an explicit NOT APPLICABLE category statement in data{} on the localization dimension and two of them (gene-smarcd1, gene-bicra) explicitly contrast that state with a bounded absence.

    What this evidence cannot show

    expected_absent_terms names tokens that must NOT appear in the cited script for the claim to hold - the claim is that the reporter has no inapplicability state, so its support is an absence in the source. The 124 figure comes from a case-insensitive phrase scan of which 8 hits were read in full.

    Evidence that argues against this (1)
    • The script's own header argues the opposite of a defect: 'This is a REPORTER, not a gate. It exits 0 always. A missing localization is not an error - for many proteins the subnuclear answer is genuinely unpublished, and several entries in this graph correctly record that as a bounded absence.' The authors knew about the inapplicable and unmeasured states and deliberately chose a generous two-valued reporter with a prose disclaimer rather than a three-valued one. The defect this hypothesis names may therefore be a documented design choice, and the real failure is that the disclaimer does not travel with the number into the ledger.
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    Applying this beat's triage rates, ~108 of the 329 entries with no epigenetics field are estimated to be inapplicable rather than unwritten, and 25 of the 195 entries with neither a field nor a mark already carry an explicit NOT APPLICABLE declaration for localization - including gene-cd19, gene-cd33, gene-cd34, gene-fut4, gene-fut9, gene-elovl1, gene-kras, gene-nras, gene-vim and gene-mapre1.

    What this evidence cannot show

    The ~108 is a point estimate built from three hand-read samples of n=20, 25 and 30 with Wilson intervals of 53.1-88.8, 30.0-66.5 and 74.4-96.5 respectively; propagating those intervals would put the inapplicable share somewhere between roughly 70 and 150 of 329. It should not be quoted as a measured count.

    Evidence that argues against this (1)
    • The 25-entry overlap is the load-bearing concrete evidence here and it does not carry the weight placed on it: a NOT APPLICABLE on subnuclear localization is not a NOT APPLICABLE on epigenetics. Five of the 25 - gene-ccnt2, gene-cdk12, gene-hexim2, gene-larp7, gene-mepce - are P-TEFb-axis proteins where the chromatin call is genuinely contested, and gene-skp2 and gene-igf2bp2 are not obviously inapplicable either. Reading the overlap as 25 answered questions would itself be the Rule 89 error of crediting a dimension on evidence about a neighbouring one.
Connected to