Oncofusion

start / graph / ana-discussed-is-draftable-for-69-percent-and-the-clause-i-predicted-would-decide-it-did-not

Analysisestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.

43 of the 62 discussed-but-unwritten interactome partners (69%) are genuinely draftable from the committed corpus, so cycle 93's decision to work them before the 34 never-read ones has a real cost basis - but the clause that decided it was the citation clause I thought was safe, not the proxy I flagged as weak, and the 12 RNA-processing failures turn out to be a promiscuity census that should NOT be promoted partner by partner

also called discussed is draftable for 69 percent · the clause I called weak was not the one that decided · the splicing block is a census, not a gap

The previous cycle decided to finish the partner proteins the project has already read about before starting the ones it has never read, on the grounds that the first group is cheaper, and this checked whether that word was true: for forty-three of the sixty-two the project's own files really do contain enough to write a page, so the decision holds. The nineteen that fail are mostly RNA-processing proteins, and looking at where they came from showed they should not simply be written up - they come from a survey in which the main protein was picked up by half of all the experiments, which may say more about its size and abundance than about any real partnership.

The same thing, for a scientist

**Statement.** Cycle 93 recorded a work order — the **62** discussed-but-unwritten interactome partners before the **34** never-read ones — on the ground that the 62 are *cheaper*. That entry's own caveats called the word an inference: *"no one has timed promoting a discussed partner against reading a new paper"*, and *"is 'discussed' a useful proxy for 'the reading is done'? It is a word-boundary token match."* Since the decision was already being leaned on, the check runs. **Pre-registered before reading any mention.** A partner is **draftable** if all three hold: **D1** named in a prose-bearing field, not only a data block; **D2** at least one mention inside an evidence item carrying a resolvable `pmid`, `doi` or `url`; **D3** a bait symbol in the same field value — an explicit *proxy* for "says something about the relationship". Read against the **committed corpus at HEAD** (2,428 node files), so population and evidence share one snapshot. | clause | pass | |---|---:| | **D1** named in prose | **62 / 62** | | **D2** inside a cited evidence item | **44 / 62** | | **D3** bait co-occurs in the same field | **58 / 62** | | **all three — draftable** | **43 / 62 (69%)** | **VINDICATED.** 69% is a real cost basis, and the median draftable partner is discussed across **4** distinct entries against **2** for the non-draftable. **But my own uninformativeness clause predicted the wrong culprit.** `U1` said *"D3 is the weak clause and I expect it to carry the result"* — it passes **58 of 62**. **The clause that decided was D2, the citation requirement I treated as routine.** Eighteen partners are named in prose, next to a bait, and **never inside an evidence item that carries a resolvable identifier.** **And the failures are not scattered. 12 of the 19 are one block:** `CNOT1`, `DDX52`, `DDX55`, `EFTUD2`, `EXOSC5`, `FASTKD2`, `HNRNPA1`, `SF3B4`, `SRSF4`, `SRSF5`, `SRSF6`, `U2AF2` — RNA processing and splicing. **CHECKED BEFORE SHIPPING, because this entry's own counter-evidence said the opposite reading was live.** Seven of those twelve are named in `ana-kmt2a-proximity-promiscuity-25-of-50`, which reports the K562 50-bait BioID2 map. Its evidence items **are** cited — IntAct PSICQUIC query URLs — but the individual partners appear in its **sections and data, never inside a cited evidence item**, which is exactly why `D2` failed for them. **And that entry gives a reason not to write them up individually.** It reports **KMT2A recovered by 25 of 50 baits, the 97.6th percentile**, against MEN1's 3, and states two competing readings it deliberately does not separate: either KMT2A is an abundant, very large chromatin protein that any nuclear bait's labelling radius reaches, or it genuinely shares a compartment with RNA processing. **Calling this block "a specific, actionable defect" was wrong** — it is a census whose interpretation is openly unresolved, and the promiscuity confound argues against writing 12 individual interaction entries off it. **The revision to cycle 93's work order:** the RNA-processing block should not simply be promoted. Either the promiscuity confound is settled first, or those partners are adjudicated as a class — which is what the census entry already does.

Go deeper
A contamination I created, found and measured: a gap list that names its gaps makes them look discussed

Eight of the 19 non-draftable partners appear in ana-the-interactome-read-queue-is-empty…, four in ana-depmap-co-dependency-cannot-rank… and two in ana-the-interactome-tail-is-orderable…. Those are my own coordinator entries from cycles 91, 92 and 93, and they name the missing partners explicitly — in a KMT2A_notable list, in a never-discussed roster, in prose about what is absent.

So the instrument that measures "discussed anywhere in the corpus" counts my own enumeration of what is missing as discussion of it. Writing down a gap converts it, by the next measurement, into something the corpus has apparently addressed. That is the inverse of the failure this project recorded as a gap list cannot see what was decided — here a gap list makes undecided things look decided.

Measured rather than assumed, because last cycle my explanation for a number was refuted by the same kind of check. Of the 62, exactly 3 — MAPT, SRSF4, SRSF5 — are discussed ONLY inside coordinator gap-list entries. The other 59 have at least one substantive mention elsewhere, and all 43 draftable partners do. The contamination is real, the mechanism is real, and the size is 3 of 62. It does not move the verdict, and it will grow every cycle a coordinator writes another gap list, so the measure needs the exclusion built in before it matters.

Facts and figures
questiondoes 'discussed in the corpus' mean an entry could actually be written from it?
verdictVINDICATED - 43 of 62 (69%) draftable, so cycle 93's ordering has a real cost basis
my uninformativeness clause predicted the wrong culpritU1 said 'D3 is the weak clause and I expect it to carry the result'. D3 passes 58/62. The clause that decided was D2, the citation requirement I treated as routine - 18 partners are named in prose beside a bait and never inside an evidence item with a resolvable identifier.
absence classmeasured-not-analysed
absence class noteEvery mention counted here was already committed and readable; what had never been analysed is whether a mention is sufficient to write from, which is the question the cycle-93 work order silently assumed an answer to.
per clause
all_three_draftable43.0all_three_draftable: 43D2_in_a_cited_evidence_it…44.0D2_in_a_cited_evidence_item: 44D3_bait_cooccurs_same_fie…58.0D3_bait_cooccurs_same_field: 58D1_named_in_prose62.0D1_named_in_prose: 62n62.0n: 62
per clause — the numbers behind the chart
{
  "D1_named_in_prose": 62,
  "D2_in_a_cited_evidence_item": 44,
  "D3_bait_cooccurs_same_field": 58,
  "all_three_draftable": 43,
  "n": 62
}
median entries per partner
{
  "draftable": 4,
  "not_draftable": 2
}
the failures are one block
{
  "n_rna_processing_of_19": 12,
  "genes": [
    "CNOT1",
    "DDX52",
    "DDX55",
    "EFTUD2",
    "EXOSC5",
    "FASTKD2",
    "HNRNPA1",
    "SF3B4",
    "SRSF4",
    "SRSF5",
    "SRSF6",
    "U2AF2"
  ],
  "reading": "this corpus discusses its splicing partners repeatedly and has never attached a source to any of them - invisible while 'discussed' was the measure",
  "competing_reading_not_ruled_out": "ana-kmt2a-proximity-promiscuity-25-of-50 is named in 8 of the 19, so these may be a deliberately unadjudicated census rather than a gap. Not checked.",
  "CHECKED": {
    "source_entry": "ana-kmt2a-proximity-promiscuity-25-of-50 (K562 50-bait BioID2 map)",
    "its_citations_ARE_resolvable": "IntAct PSICQUIC query URLs on E1 and E2; PMID 15199122 + DOI on E3",
    "why_D2_still_failed": "the individual partners are named in that entry's sections and data, never inside a cited evidence item",
    "what_it_reports": "KMT2A recovered by 25 of 50 baits (97.6th percentile) against MEN1's 3, with two competing structural readings it deliberately does not separate",
    "so_my_framing_was_wrong": "'a specific, actionable defect' overstated it. This is a census with an openly unresolved interpretation, and the promiscuity confound argues AGAINST writing 12 individual interaction entries off it.",
    "revision_to_the_cycle_93_work_order": "the RNA-processing block should not simply be promoted; either settle the promiscuity confound first or adjudicate the block as a class, which the census entry already does"
  }
}
contamination i created
{
  "mechanism": "coordinator gap-list entries NAME the missing partners, so the 'discussed anywhere' measure counts my own enumeration of what is absent as discussion of it. Writing down a gap converts it, by the next measurement, into something the corpus has apparently addressed.",
  "measured_size": 3,
  "genes_discussed_only_in_gap_lists": [
    "MAPT",
    "SRSF4",
    "SRSF5"
  ],
  "genes_with_substantive_discussion": 59,
  "does_it_move_the_verdict": "no - all 43 draftable partners have substantive discussion elsewhere",
  "but": "it grows every cycle a coordinator writes another gap list, so the exclusion belongs in the instrument before it matters"
}
Still open
  1. ANSWERED IN THIS ENTRY: it is a census with an unresolved interpretation, not a gap, and the promiscuity confound argues against promoting its partners individually. The question that replaces it: can the confound be settled? A baits-per-prey null over the same 50-bait map would say whether KMT2A's 25 is more than size and abundance predict, and that single number decides whether ANY of its proximity partners deserve individual entries.
  2. Build the gap-list exclusion into core-complex-interactome-gap/run.py before the contamination grows. Coordinator entries that enumerate absent partners should not count as discussing them; 3 of 62 today, more every cycle.
  3. Measure TIME, not feasibility. 'Cheaper' asserts an economy this analysis does not test - drafting from prose still requires verifying the prose and re-fetching the citation. Timing three promotions against three readings would settle whether the ordering is worth sequencing work around.
  4. 18 partners sit in prose beside a bait with no resolvable citation anywhere near them. Is that a curation gap in this corpus, or are they cases where the source genuinely is only an IntAct row? The answer changes whether they belong with the 43 or with the 34.
Every source behind this entry (1)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    Across the 2,428 node files committed at HEAD, of the 62 IntAct-curated partners that are discussed without having their own entry: 62 are named in a prose-bearing field, 44 appear inside an evidence item carrying a resolvable pmid, doi or url, 58 have a bait symbol in the same field value, and 43 satisfy all three. The median draftable partner appears in 4 distinct entries against 2 for the non-draftable. Of the 19 that fail, 12 are RNA-processing or splicing factors: CNOT1, DDX52, DDX55, EFTUD2, EXOSC5, FASTKD2, HNRNPA1, SF3B4, SRSF4, SRSF5, SRSF6, U2AF2. Exactly 3 of the 62 - MAPT, SRSF4, SRSF5 - are mentioned only inside coordinator gap-list entries.

    What this evidence cannot show

    DRAFTABLE IS NOT CORRECT, and this was declared before running. It measures whether material EXISTS, never whether that material is right - an entry drafted from existing prose inherits every error in it, and this project has already found a committed entry that misattributed a measured channel to the wrong beat. D3 is a co-occurrence proxy and will false-negative when prose names the partner in one sentence and the bait in the next, and false-positive in a long section about something else; its 58/62 pass rate should be read as an upper bound on 'says something about the relationship'. D2 counts a citation as resolvable if the field is present and non-empty, NOT by fetching it - verify-citations.py checks resolution corpus-wide but this run did not re-check per partner. The 62 itself comes from a token match and carries the gap-list contamination measured above.

    Evidence that argues against this (3)
    • 69% draftable is compatible with the ordering being nearly worthless in practice. If drafting from existing prose takes most of the time that reading a paper takes - because the prose must still be verified, the citation re-fetched and the claim restated - then 43 cheaper-by-a-little partners do not justify sequencing work around them. Nothing here measures TIME, which is what 'cheaper' actually asserts, and the honest position is that this establishes feasibility and not economy.
    • I wrote that the uncited RNA-processing block was 'a specific actionable defect' and flagged that I had not opened the entry it comes from. OPENED BEFORE SHIPPING and the counter-reading won: ana-kmt2a-proximity-promiscuity-25-of-50 is a K562 50-bait BioID2 census whose evidence items carry resolvable IntAct query URLs, whose individual partners sit in sections and data rather than in cited evidence items, and which reports KMT2A recovered by 25 of 50 baits while explicitly declining to separate 'abundant large protein any bait reaches' from 'genuinely shares a compartment with RNA processing'. Promoting those twelve to individual entries would write up a confound the source entry already names. What survives is narrower: 18 partners lack a citation adjacent to them, and for the 7 from this census that is a deliberate editorial position rather than a gap.
    • This analysis validates a decision I made last cycle, and I designed both the criterion and the threshold that judges it. The 60% bar for VINDICATED was set in the same file, on the same day, by the person whose decision was on trial. A 43/62 result clears it comfortably, but a pre-registration written by the interested party is weaker than it looks, and an outside reader should re-derive the bar before treating this as independent confirmation.
Connected to