Epigenetics reads 39.2% and looks like the thinnest PI dimension. Roughly half of the shortfall is a missing FIELD NAME over content that is already there, a third is proteins with no chromatin context to record, and the real content gap is about a ninth of the corpus
also called q181 epigenetics triage · is the 39.2% real
A tool that scores how well this project records the chromatin biology of each protein says only 39% of entries have it, which made epigenetics look like the worst-covered topic here. Looking entry by entry, most of the missing 61% is either content that IS written down but under a field name the tool does not recognise, or proteins like cell-surface antigens and Golgi enzymes that have no chromatin biology to record at all.
The same thing, for a scientist
check-pi-dimensions.py reports epigenetics coverage of 39.2% by dedicated field over a 541-entry gene/protein population, the lowest of the PI's five dimensions and unchanged in character since cycle 52. This analysis partitions the same population and asks, for each of the 329 entries WITHOUT a dedicated field, which of three explanations applies: (1) a real biology gap, (2) a schema artefact, (3) a category error. The partition is: 134 entries name a histone mark somewhere in the entry but carry no epigenetics-named key; of the remaining 195, 121 name a chromatin CONCEPT (enhancer, nucleosome, accessibility, Polycomb, remodelling) for which the detector's mark regex has no token at all, and 74 name neither. Hand-reading seeded samples from each stratum puts the decomposition at roughly 48% schema artefact, 33% category error, and 19% real biology gap. The corrected statement is that ~88% of entries either carry epigenetic content or have none to carry, and the real content gap is ~63 of 541 (11.6%), not the 60.8% the headline implies. The 39.2% is ARITHMETICALLY CORRECT AND SHOULD BE CORRECTED IN INTERPRETATION, NOT RETIRED: it accurately measures a field that is genuinely absent, and it has been read as a measure of missing knowledge, which it is not.
What the detector actually tests
check-pi-dimensions.py scores epigenetics two ways and prints both 1:
epi_field--any(_EPI_KEY.search(k) for k in dat). This iterates the KEY NAMES ofdata{}and matchesepigen|histone|methylat|acetylat|ubiquitylat|chromatin_mark|^mark$|^mark_|writer|eraser|reader_of. It never looks at a field's VALUE.epi_text--_EPI_MARK.search(json.dumps(d)), a specific-mark regex (H3K4, H3K27, H3K79, H2A.Z, 5mC, bivalent, ...) over the whole serialised entry.
The headline 39.2% is the first of these. So the question the number answers is "did an author name a key containing one of nine word-stems?" -- not "does this entry know this protein's chromatin context?". Those two questions come apart, and the whole of this analysis is a measurement of how far apart.
The partition, and the three strata that need different answers
Over the same 541-entry population the detector uses (type == "gene" OR id startswith "gene-"; 348 gene + 193 protein) 1:
| stratum | n | % of 541 |
|---|---|---|
A -- carries a key _EPI_KEY matches | 212 | 39.2% |
| B -- no such key, but a specific mark IS named somewhere | 134 | 24.8% |
| C -- neither | 195 | 36.0% |
and C splits again 2:
| n | % of C | |
|---|---|---|
| C-concept -- names a chromatin concept the mark regex has no token for | 121 | 62.1% |
| C-nothing -- names no mark and no chromatin concept | 74 | 37.9% |
Only 74 of 541 entries (13.7%) are silent on chromatin in every vocabulary tested. That is the ceiling on the real gap before any judgement is applied, and it is already far below the 60.8% the headline implies.
Hand-reading each stratum, with intervals
Regexes cannot tell a statement about the entry's own subject from boilerplate about KMT2A, so each stratum was sampled at seed 181 and read by hand 3. Scores are mine and the sampled ids are persisted in the analysis JSON so they can be re-scored:
| what was scored | k/n | point | Wilson 95% |
|---|---|---|---|
| B: the mark statement is ABOUT the entry's own subject | 15/20 | 75.0% | 53.1-88.8 |
| C-concept: substantive chromatin statement about the subject | 12/25 | 48.0% | 30.0-66.5 |
| C-concept: explicit NOT-APPLICABLE or bounded absence | 5/25 | 20.0% | 8.9-39.1 |
| C-nothing: protein family with no chromatin contact | 27/30 | 90.0% | 74.4-96.5 |
Stratum B is the finding. The 15 substantive cases include gene-kat2a (KAT2A/KAT2B acetylate H3K9 and SEC components), gene-kat8 (MOF H4K16 acetyltransferase in the MLL1 complex), gene-kmt2d (H3K4 mono-methyltransferase at enhancers), gene-mllt1 (the YEATS-H3K27ac structure 5J9S), gene-srcap (an H2A.Z-loading motor), gene-rbbp4 (PRC2 assembly and H3K27me3), gene-son (lowers H3K4me3 via menin), gene-atrx (H3K9me3 at ERVs) and gene-brd3 (hyperacetylated chromatin occupancy). These are canonical writers, readers and erasers of the exact marks the detector looks for, each with a cited finding, all scoring zero on the field line. The five that failed were boilerplate about another protein (gene-lmo2's hit is a sentence about KDM5A; gene-picalm's is a citation title) or explicit negatives (gene-ctnnb1, whose title says it is "the one piece of KMT2A biology that is not chromatin").
The decomposition, and what the number should say instead
Applying the sampled rates to the strata gives, over the 329 entries with no dedicated field:
| explanation | ~n | ~% of the shortfall |
|---|---|---|
| (2) schema artefact -- the content is there, the field name is not | 158 | 48.0% |
| (3) category error -- no chromatin context exists to record | 108 | 32.8% |
| (1) real biology gap | 63 | 19.1% |
Corrected headline: ~478 of 541 (88.4%) of entries either carry epigenetic content or have none to carry. The real content gap is ~63 of 541, 11.6%.
These point estimates inherit the sample intervals above and should be quoted as approximate. The claim that does not depend on any sampling is the 74-entry C-nothing stratum, and the eight-entry hard floor in the sibling entry ana-q181-the-only-chromatin-machinery-with-genuinely-nothing-is-the-ptefb-brake.
Why this is not an argument for changing the tool quietly
The instruction that produced this analysis was explicit that altering a measurement tool to improve the number it reports would make the beat worthless, and scripts/check-pi-dimensions.py was not modified. Every figure above was produced by a separate script reading the same population and re-implementing the two regexes verbatim.
Two concrete defects in the detector are reported as findings, with proposed fixes, in ana-q181-the-epigenetics-detector-reads-key-names-and-never-field-values and ana-q181-the-mark-regex-has-no-token-for-enhancer-nucleosome-or-accessibility. The recommendation about what to do with the metric is in hyp-q181-a-two-state-coverage-metric-cannot-express-not-applicable. None has been applied.
| script | analyses/q181-is-the-epigenetics-gap-real/{keys,run,baserate,cohortc,vocab,sample121,final}.py |
|---|---|
| status | run |
| generated | 2026-09-24 |
| tool modified | false |
| tool modification note | scripts/check-pi-dimensions.py was NOT edited. Its two regexes were re-implemented verbatim in a separate script. |
| population | 541 |
| verdict | CORRECT, DO NOT RETIRE. The 39.2% accurately measures field presence and is being read as a measure of missing knowledge. Roughly half the shortfall is notation and a third is proteins with no chromatin context. |
strata — the numbers behind the chart
{
"A_field": 212,
"B_mark_only": 134,
"C_neither": 195,
"C_concept": 121,
"C_nothing": 74
}measurements — the numbers behind the chart
[
{
"quantity": "share of the epigenetics coverage shortfall that is a schema artefact",
"value": 48,
"units": "percent",
"n": 329,
"subject": "entries in the 541 gene/protein population with no _EPI_KEY-matching data key",
"universe": "329 uncovered entries; 134 stratum B scaled by 15/20 hand-read + 121 stratum C-concept scaled by 12/25 hand-read",
"null": null,
"note": "point estimate only; inherits Wilson 53.1-88.8 and 30.0-66.5 from the two samples"
},
{
"quantity": "share of the shortfall that is a category error (no chromatin context to record)",
"value": 32.8,
"units": "percent",
"n": 329,
"subject": "same 329 uncovered entries",
"universe": "121 C-concept scaled by 5/25 explicit NOT-APPLICABLE + 74 C-nothing scaled by 27/30 + 17 attributed from stratum B",
"null": null,
"note": "C-nothing is dominated by cell-surface antigens, Golgi/ER enzymes, septins and rRNA-processing factors"
},
{
"quantity": "share of the shortfall that is a real biology gap",
"value": 19.1,
"units": "percent",
"n": 329,
"subject": "same 329 uncovered entries",
"universe": "residual after the two estimates above",
"null": null,
"note": "= ~63 of 541 entries, 11.6%, against the 60.8% the bare headline implies"
},
{
"quantity": "entries silent on chromatin in EVERY vocabulary tested",
"value": 74,
"units": "entries",
"n": 541,
"subject": "no _EPI_KEY key, no _EPI_MARK token, no chromatin concept term",
"universe": "541 gene/protein entries at commit ac2aeab, 2026-09-24",
"null": null,
"note": "13.7% - the sampling-free ceiling on the real gap before any judgement is applied"
},
{
"quantity": "corrected coverage: carries epigenetic content OR has none to carry",
"value": 88.4,
"units": "percent",
"n": 541,
"subject": "the epigenetics PI dimension, re-stated as a three-state question",
"universe": "212 field + ~158 content-without-field + ~108 not-applicable, of 541",
"null": 39.2,
"note": "the null here is the figure this analysis is correcting, not a statistical null"
}
]- Which reading of the dimension did the PI intend - 'the entry knows this protein is a chromatin protein' or 'the entry records which mark, at which loci, in which cell state'? The decomposition above inverts between the two, and nothing in the standing instruction distinguishes them. This is the single question that most changes the answer.
- Stratum A was never audited. How many of the 212 entries with an epigenetics-named key carry an empty, hedged or 'unknown' value in it? _EPI_KEY matches the key name alone, so a covered entry can say nothing.
- Would a second reader, given the same seed-181 samples and the same three categories, reproduce the 15/20, 12/25 and 27/30? The ids are persisted; nobody has re-scored them.
- Does the same key-name-only test deflate the other four PI dimensions in the same way, and by how much? Localization is measured field-only for a documented reason, but nobody has asked whether its 77.4% has the same notation component.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
check-pi-dimensions.py, run 2026-09-24, reports epigenetics coverage over 541 gene/protein entries as 39.2% by dedicated field (212), 54.7% by a mark named in the body (296) and 64.0% either (346), with the gene split at 37.9% and protein at 41.5%; its epi_field test iterates data{} KEY NAMES only and never inspects a field's value.
What this evidence cannot show
The 541-entry population is a UNION of type=='gene' and id-prefix 'gene-', which the script's own comments flag as a union of two node types whose coverage differs; it also sweeps in six comut-* composite nodes that are not single proteins. The figures are a snapshot at one commit and other beats were writing concurrently.
Evidence that argues against this (1)
- The detector's generosity runs the other way too, and this analysis does not correct for it: _EPI_KEY matches 'writer', 'eraser' and 'methylat' as bare substrings of a key name, so a key like 'writer_of_that_mark' on an entry whose value is 'unknown' scores as covered. Stratum A was never audited for empty or hedged values, so the 212 is an upper bound on genuinely covered entries and the true A is smaller than stated. That error pushes the corrected headline DOWN and was not quantified.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Of the 195 entries with neither a dedicated epigenetics key nor any specific mark token, 121 (62.1%) name at least one chromatin concept for which _EPI_MARK has no token - 'chromatin' bare in 97, enhancer in 25, histone bare in 23, remodelling in 20, Polycomb/PRC in 15, ChIP-seq/CUT&RUN in 13, heterochromatin in 9, nucleosome in 8, chromatin accessibility/ATAC in 6 - leaving 74 entries (13.7% of the 541) silent on chromatin in every vocabulary tested.
What this evidence cannot show
The concept term list is mine and was written before the count but not pre-registered anywhere a reviewer can check independently. 'chromatin' bare is by far the largest contributor (97 of 121) and the hand-read shows it is the least reliable: four of 25 sampled hits were the word appearing in a method description or a paper title rather than a statement about the subject ('remodeling of the human interactome' matched the remodel pattern in gene-ap2b1 and is not chromatin remodelling at all).
Evidence that argues against this (1)
- This measurement argues the detector's vocabulary is too narrow, but the detector's own source comments give a reason for the narrowness: the generic words 'acetylation' and 'methylation' were deliberately excluded because they would match any passing mention and turn the measure into noise. The same objection applies with full force to bare 'chromatin', which is the bulk of my 121. If bare 'chromatin' is dropped, the concept stratum falls from 121 to at most 24 entries on the specific terms, and the schema-artefact share of the shortfall falls with it.
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
In a seed-181 sample of 20 stratum-B entries, 15 carry a mark statement about the entry's own subject rather than imported boilerplate (75.0%, Wilson 53.1-88.8); in a seed-181 sample of 25 stratum-C-concept entries, 12 carry a substantive chromatin statement about the subject and 5 carry an explicit NOT-APPLICABLE or bounded-absence statement; in a seed-181 sample of 30 stratum-C-nothing entries, 27 are proteins of families with no chromatin contact (cell-surface antigens CD22/CD34/CD7/KIT/S1PR4/SIT1, Golgi and ER enzymes, cytoskeletal septins and microtubule proteins, rRNA-processing nucleolar factors, spliceosomal and decay-pathway RNA-binding proteins, and comut-* composite nodes).
What this evidence cannot show
Three samples of n=20, 25 and 30 against strata of 134, 121 and 74. The intervals are wide and the C-concept proportion in particular spans 30-66%, so the 48% used in the decomposition is not distinguishable from anything between a third and two-thirds. Scoring was done by the same agent that framed the categories and that wanted the answer to be interesting, with no second reader; the sampled ids are persisted precisely so a second reader can disagree.
Evidence that argues against this (2)
- The single-scorer design is the weakest part of this analysis and it biases in a known direction. I classified 'the entry states the subject is a chromatin protein' as substantive content, but the PI dimension asks for a protein's chromatin CONTEXT - which mark, at which loci, in which cell state - and almost none of the 15 stratum-B substantive cases carries that at field granularity. Under a stricter reading of what the dimension asks for, most of the 158 I score as schema artefact would revert to a real gap and the decomposition would invert. The two readings differ in what 'coverage' means, not in the counts, and this entry does not settle which the PI intended.
- No second instrument was run against these labels. Rule 92 in this project's standing brief records a classifier that was wrong five times out of nine at the top of its ranking, and the corrective there was hand-labelling - which is all this analysis has. There is no independent check that my hand labels are better than the regex they replace.
supplies the null (42.5% corpus-wide) under the 54.7% epi_text line the triage rests on
the key-versus-value asymmetry is the mechanism behind the schema-artefact share of the shortfall
121 of the 195 empty-stratum entries name a chromatin concept the detector has no token for
the core-axis audit reproduces the schema-artefact dominance on a pre-specified 32-protein list
the sampling-free floor (8 of 122 machinery entries) bounds the real-gap share the triage estimates
the corpus already writes the third state on a neighbouring dimension, which is why the category-error share is invisible here
the ~33% inapplicable and ~48% notational shares are the triage's output and the hypothesis's premise
- related to →Plain-English: chromatin and epigenetics — the sticky notes on the DNA, and which ones matter here
measures how well this corpus records the dimension this concept entry explains