King verification of q181: the epigenetics field is NOT hollow (1 of 218 empty), the mark regex has three further blind spots including chromatin accessibility in 25 entries, and my own attempted correction of q181 was wrong
also called q184 · stratum A audit · king check of the epigenetics instrument
A check of this project's own epigenetics coverage tool. The tool was accused of under-counting, and it does - it only ever looks at the NAMES of data fields, never at what is written inside them, and it has no word for chromatin accessibility even though the project reasons about that constantly. But the opposite worry turned out to be unfounded: the entries it counts as covered really are covered, with only one of 218 empty. Along the way I nearly published a correction of the agent's report that was itself wrong, and caught it by re-checking my own pattern.
The same thing, for a scientist
q181 delivered its own weakest points rather than hiding them: *"stratum A was never audited — `_EPI_KEY` matches key names, so some of the 212 may be empty behind the key"*. This audits it, and re-derives the instrument claims from the regexes read out of the script rather than restated. 🟢 **Stratum A is not hollow, and that is a clean confirmation q181 could not give itself.** Of the 218 gene/protein entries with a matching `data{}` key, **exactly 1 (0.5%) has all its matched keys empty** [E1]. The field-based coverage figure is not inflated by empty keys, so the undercount runs in one direction only. 🔵 **74 of 218 have a matched key whose value names no mark by a strict test — and inspecting them, that is not a defect.** They carry deliberate content: `gene-hdac5`'s `non_histone_substrate`, `gene-smc3`'s `acetylation_axis_not_claimed`, `gene-mef2c`'s `epigenetic_direction_conflict`. **A strict 'the value must name a histone mark' test would have called all 74 defective. It is the test that is wrong, not the entries** — the same error the tool under audit makes, one level down [E1]. 🔴 **Three further blind spots in `_EPI_MARK`, beyond the accessibility gap q181 named** [E2]. Among entries the pattern scores as naming no mark: | notation it cannot see | entries | |---|---:| | chromatin accessibility (ATAC / DNase / accessib / FAIRE) | **25** | | a mark referenced by node id, e.g. `mark-h2bk34ub` | 4 | | hyphenated `CpG-island` where the pattern has `CpG island` | 2 | | untokenised residues (H2BK34, H3K122, H4K20, H3K64) | 1 | **32 entries are rescued by one of these.** The accessibility gap is the serious one: the PI's dimension statement names it, this corpus reasons about it constantly — *the z-FLT3 enhancer is ATAC-closed in KMT2A-AF4* — and the instrument has no token for it. 🔴 **And a correction of my own, caught before publication.** I measured *"the epigenetics third state already exists, 89 entries (15.0%)"* and was about to record it as a correction of q181's claim that epigenetics lacks that state. **It is wrong.** My pattern included a generic `NOT APPLICABLE`, which matched the corpus's **localization** convention. Separated: the **epigenetics-specific** negative (*"no histone-modifying activity"*) appears in **8 entries (1.3%)** — `gene-eaf1`, `eaf2`, `ice1`, `ice2`, `nelfb`, `nelfe` and two more — against **81** formal `NOT APPLICABLE` uses on localization, with **zero overlap** [E3]. **q181 was right: epigenetics has the state as an ad-hoc phrasing in 8 places, not as a convention.** Its recommendation stands unaltered.
What the corrected recommendation is
q181 recommended that the corpus adopt a third state for epigenetics — neither has a mark nor unknown, but has none to have. Confirmed: it is written 8 times, ad hoc, in free text, while localization has a formal spelling used 81 times.
So the fix is two-sided, and only one side was visible from either report alone:
- The corpus should adopt the state formally for epigenetics, as q181 said.
- The checker must recognise it when written — today an entry saying "none — no histone-modifying activity" scores identically to an entry that says nothing, because
_EPI_MARKfinds no mark in either. Adopting the convention without fixing the checker would make the coverage figure WORSE, by adding entries that state their status and are still counted as gaps.
That interaction is the reason to fix the instrument before running any filling campaign.
Why the denominators differ from q181's, and why that is not a disagreement
q181 measured 541 gene/protein entries; this measures 593. The corpus gained entries between the two runs — both this cycle's agents were writing while the audit ran, and total nodes moved from 3,031 to 3,059 during it.
The rates agree closely where they should: mark-but-no-field is 25.0% here against q181's 24.8%, and the key-versus-value count is 74 here against its 71. The base rate it flagged reproduces exactly: 1,300 of 3,059 nodes name a histone mark, 42.5%.
A coverage figure measured against a moving corpus should be quoted with its node count, and neither report did. Any comparison between cycles that ignores this will read growth as regression.
| script | inline king verification, cycle 159; regexes extracted verbatim from scripts/check-pi-dimensions.py |
|---|---|
| status | executed |
| verifies | the q181 beat |
| n gene protein entries | 593 |
| q181 denominator | 541 |
| stratum a total | 218 |
| stratum a all keys empty | 1 |
| stratum a key without mark in value | 74 |
| epigenetics third state entries | 8 |
| localization not applicable entries | 81 |
| overlap | 0 |
| king withdrawn claim | 89 entries (15.0%) carry an epigenetics third state - wrong, conflated with localization |
| verdict | the field is not hollow; the instrument undercounts in at least four distinct ways; q181's third-state recommendation stands and the checker must be fixed before it is adopted |
mark regex blind spots — the numbers behind the chart
{
"chromatin_accessibility": 25,
"mark_node_id_reference": 4,
"hyphenated_cpg_island": 2,
"untokenised_residue": 1,
"entries_rescued": 32
}reproduced from q181 — the numbers behind the chart
{
"mark_but_no_field_pct": 25,
"q181_reported": 24.8,
"key_vs_value_count": 74,
"q181_reported_count": 71,
"histone_mark_base_rate": 0.425,
"q181_reported_base_rate": 0.425
}measurements — the numbers behind the chart
[
{
"quantity": "share",
"subject": "stratum A entries whose matched keys are all empty",
"universe": "218 gene/protein entries with a matching data key",
"value": 0.005
},
{
"quantity": "count",
"subject": "entries naming chromatin accessibility and invisible to _EPI_MARK",
"universe": "593 gene/protein entries",
"value": 25
},
{
"quantity": "count",
"subject": "entries carrying an epigenetics-specific third state",
"universe": "593 gene/protein entries",
"value": 8
},
{
"quantity": "count",
"subject": "entries carrying the formal NOT APPLICABLE localization state",
"universe": "593 gene/protein entries",
"value": 81
}
]- Fix _EPI_MARK and _EPI_KEY together, then re-run: read data{} VALUES as well as key names, add an accessibility token, resolve mark-* node-id references, and recognise the 'no histone-modifying activity' state. Adopting the third state WITHOUT the checker fix would make the reported coverage worse, because entries stating their status would still count as gaps.
- Open the other 62 of the 74 stratum-A entries whose matched key holds no mark. Twelve were read and all twelve were substantive; the generalisation to 74 is a sample.
- Should chromatin accessibility count toward the epigenetics dimension at all? It is a state of a locus rather than a modification carried by a protein, and widening the pattern could inflate coverage with mentions. Decide it explicitly before changing the regex.
- Quote the node count beside every coverage figure. q181 measured 541 gene/protein entries and this measured 593 an hour later, because agents were writing throughout. Any cycle-to-cycle comparison that ignores the denominator will read growth as regression.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.software
Of 593 gene/protein entries, 218 have at least one data{} key matching _EPI_KEY. Of those 218, exactly 1 (0.5%) has all matched keys empty by a test treating null, empty string, empty list/dict and the tokens na/n-a/none/unknown/tbd/- as empty. 74 of 218 (33.9%) have a matched key whose value contains no _EPI_MARK match; inspection of those shows deliberate content including gene-hdac5 non_histone_substrate, gene-smc3 acetylation_axis_not_claimed and gene-mef2c epigenetic_direction_conflict.
What this evidence cannot show
The emptiness test is the King's own and treats a handful of placeholder tokens as empty; a differently-worded placeholder would be counted as content. The 74 were inspected by reading the first twelve, not all of them, so 'deliberate content' is an impression from a sample rather than a per-entry verdict.
Evidence that argues against this (2)
- Reading 12 of 74 and generalising is exactly the sampling this entry criticises elsewhere. The other 62 were not opened, and if a meaningful fraction of them are boilerplate rather than considered content, the claim that stratum A is substantive weakens correspondingly.
- 0.5% empty measures only whether a key has a VALUE, not whether the value is correct or current. An entry with a confidently wrong chromatin annotation scores as covered here, and this audit has nothing to say about the accuracy of the 218.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.software
_EPI_MARK, read verbatim from scripts/check-pi-dimensions.py, matches H3K4|H3K27|H3K36|H3K79|H3K9|H4K16|H2A.?Z|H2AK119|H2BK120|H3K14|H3K18|H3K23|H3K56|H4K5|H4K8|H4K12|H3R2|H3R8|H3R17|H3R26|H4R3|arginine methylation|5mC|5hmC|CpG island|DNA methylation|bivalent. It contains no token for chromatin accessibility. Among gene/protein entries it scores as naming no mark, 25 name an accessibility concept (ATAC, DNase, accessib, FAIRE, open chromatin), 4 reference a mark by node id such as mark-h2bk34ub, 2 use the hyphenated form CpG-island where the pattern has CpG island, and 1 names an untokenised residue; 32 entries are rescued by at least one.
What this evidence cannot show
The variant patterns are the King's own and could over-match: 'nucleosome' and 'accessib' appear in prose that is not an accessibility measurement, so 25 is an upper bound on genuine accessibility annotation. The hyphenation and node-id counts are small enough to be noise.
Evidence that argues against this (1)
- Adding an accessibility token would change what the dimension MEASURES, not just how well it measures it. Chromatin accessibility is a state of a locus, not a modification carried by a protein, and a gene entry mentioning ATAC is not necessarily making an epigenetic claim about that gene. Widening the pattern could inflate coverage with mentions rather than annotations - the same failure the existing comment in the script warns against for the words 'acetylation' and 'methylation'.
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.software
A King measurement of 89 entries (15.0%) carrying an epigenetics 'third state' was wrong and is withdrawn. The pattern used included a generic NOT APPLICABLE, which matches the corpus's localization convention. Separating them: the epigenetics-specific negative, 'no histone-modifying activity' or 'none - no histone', appears in 8 gene/protein entries (1.3%) - gene-eaf1, gene-eaf2, gene-ice1, gene-ice2, gene-nelfb, gene-nelfe and two others - while generic NOT APPLICABLE appears in 81, with zero entries carrying both.
What this evidence cannot show
The epigenetics-specific pattern is two phrasings; a third wording would be missed, so 8 is a floor. The zero overlap is informative but follows partly from the two conventions belonging to different dimensions and different beats.
Evidence that argues against this (1)
- Recording a caught self-error is worth little if the same standard was not applied to the rest of this entry. E1's emptiness test and E2's variant patterns are also King-written and were NOT independently re-derived the way this one was after it failed - they were checked once. The asymmetry is real: this error was caught because its conclusion contradicted an agent, which made it worth a second look.
king verification of the q181 beat: stratum A audited and clean, three further regex blind spots, and one king correction withdrawn
king verification of the q181 beat: stratum A audited and clean, three further regex blind spots, and one king correction withdrawn
king verification of the q181 beat: stratum A audited and clean, three further regex blind spots, and one king correction withdrawn