AFF1 (P51825) — review notes

Identity, verified before anything else

Row count reconciles — the stub did not collapse anything

genes/human/AFF1/AFF1-goa.tsv : 21 lines = 20 data rows, 20 distinct
AFF1-ai-review.yaml stub      : 20 `- term:` entries, in TSV order

Verdict tally

action rows
ACCEPT 11
MODIFY 7
KEEP_AS_NON_CORE 2
NEW 1
total entries 21 (20 GOA rows + 1 NEW)

Recorded here deliberately. The tally was first written by hand into the PR body and
a commit message as "11 ACCEPT · 6 MODIFY · 2 KEEP_AS_NON_CORE · 1 NEW", which sums
to 20 rather than 21 and undercounts MODIFY by one — the third instance in this
review of a count disagreeing with its own enumeration. Stating it in a file the
audit can read makes it checkable: clause 4 of the counted-claims guard computes the
tally from the document and compares it against any tally-shaped string here, so this
table cannot drift from the YAML. The PR body and git commit messages are not
lintable from inside the repository, which is why the numbers there should be copied
from the guard's COMPUTED_VERDICT_TALLY output rather than counted by hand.

This gene is a negative result for the known fetch-gene under-seeding defect
(GOAValidator.seed_missing_annotations keys on term+evidence+reference+negated+qualifier and omits
WITH/FROM). AFF1's three GO:0005515 rows differ in reference as well as partner
(PMID:21729782/P42568, PMID:22190034/P04608, PMID:23260655/P42568), so the key separated them
and each partner got its own entry. Checked and reported rather than assumed.

The worklist name is stale here too

AFF1 sits on human-no-IBA-simple.csv and has five IBA rows (GO:0006355, GO:0003712,
GO:0006354, GO:0050877, GO:0032783), all from PANTHER:PTN000829417. Confirmed against
QuickGO with a working-endpoint control, not read off the file name.

What UniProt does and does not say

The one direct molecular measurement on human AFF1

PMID:23260655 is titled for AF9, and it contains the only quantitative binding measurement on
AFF1. What was assayed matters: not full-length AFF1 but a 14-residue AFF1 peptide —

PMID:23260655

and the affinity is extraordinary:

PMID:23260655

with AFF1's residues becoming ordered only on binding:

PMID:23260655

This is a coupled folding-and-binding motif, so a 14-mer is the biologically meaningful unit rather
than a truncation artefact — unlike the ADNP-style hazard where a synthetic peptide stands in for a
gene product it cannot represent. AFF1 residues form part of the AF9 hydrophobic core, and the same
paper measures the competing ligands (Dot1L 1.6 nM, BCoR 32 nM, hPC3 > 0.9 µM), so AFF1 is the
tightest of AF9's known partners by an order of magnitude.

The distinct peptide hazard is present in this gene's literature, just not in its GOA rows:
PMID:15269783 characterises PFWT, a synthetic peptide modelled on AFF1's AF9-binding domain,
and its cellular effects (apoptosis in t(4;11) cells) are properties of the inhibitor. No GOA row
rests on it, and none should.

The paralogue-transfer hazard runs in both directions and neither is in GOA

AFF1 vs AFF4: they are not interchangeable, and one paper measures the difference

PMID:28955517 is indexed as "AFF1 and AFF4 differentially regulate the osteogenic
differentiation of human MSCs" — a title the affinage record does not reflect, summarising only the
AFF1/DKK1 half. In human mesenchymal stromal cells:

Both directions were tested, so this is requirement and sufficiency, plus an in-vivo ectopic
bone-formation arm. This is the best-characterised gene-specific human function AFF1 has, and GOA
carries no osteogenesis annotation for it at all — hence one NEW row proposing
GO:0045668 negative regulation of osteoblast differentiation.

Two caveats recorded rather than smoothed over:
1. The paper carries a 2020 Correction (PMID:32257529) whose content is not stated anywhere
retrievable — its abstract is only "[This corrects the article DOI: 10.1038/boneres.2017.44.]",
and the PMC record adds nothing. Neither affinage nor GOA flags it.
2. One sentence in the overexpression section reads "the expression of osteogenic-related genes …
was significantly repressed in AFF1-depleted cells" where the surrounding figure and the rest
of the section describe AFF1-overexpressing cells. The claims above deliberately rest only on the
unambiguous sentences.

The paper's discussion also states the mutual-exclusivity model:
PMID:28955517
That is contested by measurement: IntAct records an AFF1–AFF4 physical association across 8
records and 2 distinct PMIDs (20153263, 21729782) at MI 0.6, and PMID:20159561's AFF4
purification recovers AFF1. Filed as a knowledge gap, not resolved.

The DNA-damage block: one recent paper, seven rows, and it is not a projection

PMID:41062835 (Nat Chem Biol 2026) supplies 7 of the 20 GOA rows. The cache is
full_text_available: false, so nothing here characterises its figures. The abstract supports:

PMID:41062835

and

PMID:41062835.

The reference-projection test on this PMID (fully paginated, entities as a distinct id set)
returns 8 annotations over 2 entities: six terms on AFF1 alone, and GO:0090734 site of DNA damage shared with PARP1 (P09874). The phenotype terms do not spread. This is the benign shape,
not the ACTR8-style complex-projection defect — and the check was run precisely because AFF1 is a
complex subunit. The two IMP rows and one shared location row are attributed to the gene whose
depletion was performed.

The same PARP1/Siah1 axis is corroborated independently and from the opposite direction by
PMID:31611376, which has full text:

PMID:31611376

PMID:31611376

That last experiment is the cleanest demonstration that the bridging is AFF1's own activity: a
point change in AFF1 alone dissolves SEC and costs Pol II CTD phosphorylation. It is why this review
places GO:0030674 protein-macromolecule adaptor activity in molecular_function and the
elongation-factor and coregulator terms in contributes_to_molecular_function. The same paper states
the paralogue slot the same way UniProt does:
PMID:31611376.

Together the two papers describe one process — transcription shutdown then restart around DNA damage
— that GO cannot name; see the ontology gap below.

AFF1 is a constitutive P-TEFb partner

PMID:24367103

(abstract only; nothing below is asserted about its figures). IntAct agrees and quantifies it: CDK9
(P50750) is AFF1's best-supported partner anywhere — 16 records over 7 distinct PMIDs and 5
distinct detection methods, MI 0.9 — with cyclin T1 (O60563) at 8 records across six different
cell lines from one publication.

WITH/FROM resolution, and the donor evidence question asked properly

All five IBA rows come from PANTHER:PTN000829417. The protein donors resolve as:

token resolves to relationship to AFF1
FB:FBgn0041111 Q9VQI9 lilli, D. melanogaster, Swiss-Prot the single fly family member; co-orthologue of all four vertebrate AFFs
MGI:MGI:1100819 O88573 Aff1, mouse, Swiss-Prot the 1:1 ortholog
MGI:MGI:106927 P51827 Aff3, mouse, Swiss-Prot paralog
MGI:MGI:1202294 O55112 Aff2, mouse, Swiss-Prot paralog
UniProtKB:P51825 AFF1 itself self-reference (rows for GO:0006355 and GO:0006354)

MGI tokens arrive as MGI:MGI:1100819 and UniProt's xref:mgi- index needs the bare number; a query
containing the inner colon returns HTTP 400. Every lookup returned multiple candidates (3–10) and the
script reports all of them rather than taking the first, then requires exactly one Swiss-Prot entry.

Every one of the nine (donor, term) pairs holds its own experimental evidence in that term's
subtree. So SOURCE_WEAK_OR_INFERRED would be contradicted by the measurement; the correct
root cause on every accepted IBA row here is NO_FAILURE_CORE (or NO_FAILURE_NON_CORE), and the
self-referential rows record a PAINT curator judging the function core rather than a circularity.

Two findings that only appear when you ask which term the donor holds:

  1. GO:0050877 nervous system process lands two levels above its donors' evidence. Both
    experimental donors — fly lilli (PMID:18310460) and mouse Aff2 (PMID:11923441) — carry
    GO:0007611 learning or memory IMP, a descendant. Propagating GO:0007611 itself to human
    AFF1 would be unsupported, so the generalisation is deliberate and conservative: good PAINT
    practice, the mirror image of the ACRV1 defect where a propagation landed above a donor that
    did hold the specific term.
  2. The ortholog is absent from four of the five donor sets. Mouse Aff1 is cited only for
    GO:0006355. For GO:0050877 in particular it contributes nothing — and yet the AFF1-specific
    nervous-system literature is a mouse phenotype: the robotic mutant stabilises Af4 and causes
    Purkinje-cell loss and ataxia (PMID:12629167), with Af4 directly regulating Igf-1 in Purkinje
    cells (PMID:20007461). MGI has annotated no nervous-system term to Aff1. So the nervous
    system claim reaches human AFF1 through its paralogues while the ortholog's own relevant
    phenotype is uncaptured — a curation gap worth reporting upstream, and the reason this row is
    KEEP_AS_NON_CORE rather than core.

Node reach, both halves of the question

PTN000829417 carries 395 IBA annotations over 79 recipient gene products — five terms × 79, so
the block is applied uniformly. Its human reach is exactly AFF1, AFF2, AFF3, AFF4, and all four
receive the identical five terms. Verified against QuickGO per paralogue.

The reciprocal question — which node's reach is exactly my gene set, and what did it give them —
comes back clean here. There is no node whose human reach is a proper subset of the AFF family doing
something odd to it: GO:0032783 reaches AFF2 twice (also via PTN002575678), which is redundant
rather than wrong. And the family-level assignment is well founded on its face: GO:0032783's own
definition names the family —

"At minimum, the complex contains a transcription factor of the ELL family, an EAF protein, and an
AFF family protein or distant relative"

— and PMID:22547686 reports that AFF2 and AFF3 occupy the SEC-L2 and SEC-L3 variants, so all four
paralogues genuinely belong to SEC-family complexes. This is the inverse of the ACTL8/GO:0035267
case: reading the term definition here supports the propagation.

GO:0006355 staying at the unsigned parent is likewise correct rather than lazy. The node's members
are heterogeneous in direction: AFF1 activates DKK1 (PMID:28955517) while mouse Aff2 carries
GO:0010629 negative regulation of gene expression by IMP. When donors disagree on sign, the unsigned
parent is the LCA and there is no granularity defect to fix.

The finding: one reference, one clade got the specific term, the other did not

PMID:22195968 annotates both GO:0008023 transcription elongation factor complex and its child
GO:0032783 super elongation complex. Resolving every recipient's organism:

Twenty-two gene products receive one of the two terms, 11 human and 11 Drosophila. Not one of the
11 human recipients got the specific term, and 10 of the 11 fly ones did. Stated precisely because
the exception matters: the split is not purely clade-based — Drosophila Ell is the twelfth
parent-only recipient — so this is a per-annotation gap rather than a rule about clades. (An earlier
draft of this section said "12 human", counting Ell among them; the enumerated list has 11 names and
that mismatch is what surfaced the error. A count disagreeing with its own enumeration is the bug
report, not a rounding detail.)

And the consequence for AFF1 is a three-step detour: fly lilli's GO:0032783 IPI from this very paper is the donor of human AFF1's
GO:0032783 IBA, while AFF1's own direct annotation from the same paper sits one level up. The
specific term reached the gene by phylogenetic inference from a fly protein annotated in the same
experiment that annotated the human protein.

UniProt's own reference table for that paper reads
[file:human/AFF1/AFF1-uniprot.txt "RP IDENTIFICATION IN THE SEC COMPLEX."], i.e. the curator's
reading was SEC-specific. Hence MODIFY GO:0008023 → GO:0032783, and a suggested_question naming
all eleven affected human gene products once rather than repeating it per gene.

The other granularity call: polymerase specificity

The node asserts, of the same 79 gene products, both GO:0032783 — a term whose definition is
explicitly "increases the overall rate of RNA polymerase II transcription elongation" — and
GO:0006354 DNA-templated transcription elongation, which is polymerase-agnostic. Those two cannot
both be maximally precise. Every characterised SEC substrate is Pol II:
PMID:22547686,
and the p300 paper measures CTD phosphorylation of Pol II specifically. AFF1 already holds the
Pol II-specific regulatory term GO:0032968 by IMP. GO:0006368 transcription elongation by RNA polymerase II was verified to be an is_a/part_of descendant of GO:0006354 before proposing it,
so both GO:0006354 rows are MODIFY → GO:0006368.

Deliberately not claimed: that GO:0032968 is under GO:0006368. It is not — GO links them by
positively_regulates and keeps regulation out of the is_a hierarchy. Both closures were fetched.

Redundancy the ontology can see

Fetched, not inferred (all seven relation claims are asserted in results.json):

relation verified consequence for this gene
GO:0032968 ⊂ GO:0032786 yes the two IMP rows from PMID:41062835 are a parent/child pair from one reference; the parent is redundant → MODIFY → GO:0032968
GO:0032783 ⊂ GO:0008023 yes the two complex IDAs differ only in precision
GO:0006355 ⊂ GO:0010468 yes the InterPro2GO row is the least specific regulation row on the gene
GO:0006368 ⊂ GO:0006354 yes licenses the polymerase-specificity refinement
GO:0003711 ⊄ GO:0003712 correct the two MF rows are different claims, not a general/specific pair, so both stand
GO:0000785 ⊄ GO:0005634 correct chromatin and nucleus are independent location claims
GO:0090734 ⊄ GO:0000785 correct the two damage-associated locations are not a pair

GO:0003711 has no children, so it is already maximal; there is no Pol II-specific
elongation-factor-activity term to refine to, and ACCEPT is the only available action.

The GO:0010468 row comes from InterPro2GO via IPR007797, the AF4/FMR2 family signature. PAINT
gives the same family GO:0006355 — one level more specific — so two automatic pipelines assign
different granularities to the same family from the same evidence base, and the coarser one is
strictly redundant on this gene. MODIFY → GO:0006355.

Interaction rows, decided per partner

IntAct expansion (all 104 records accounted for; the run fails if any is unassigned):

partner records distinct PMIDs distinct methods max MI
CDK9 P50750 16 7 5 0.90
MLLT1/ENL Q03111 9 2 3 0.60
AFF4 Q9UHB7 8 2 3 0.60
CCNT1 O60563 8 1 (six cell lines) 3 0.53
MLLT3/AF9 P42568 7 2 5 0.73
EAF1 Q96JC9 5 3 1 0.64
ELL3 Q9HB65 4 3 1 0.64
HIV-1 Tat P04608 3 1 2 0.56

Ontology gap

There is no GO term for transcriptional restart / recovery of RNA synthesis after DNA damage.
Confirmed on a working endpoint (control: a search for "super elongation complex" returns
GO:0032783 first) across five phrasings, and by enumerating all 96 is_a/part_of descendants
of GO:0006974 — none has "recover", "restart" or "resum" in its name. Consequently the one process
that PMID:41062835 and PMID:31611376 jointly describe has to be expressed as an unrelated pair,
GO:0006974 + GO:0032968, which loses the fact that AFF1 is the switch for both the shutdown and
the restart. Filed under proposed_new_terms.

Considered and not filed: child terms for AFF1-SEC versus AFF4-SEC. GO:0032783 has no children
and ComplexPortal's seven AFF1-containing entries all list AFF1 and AFF4 together, so no external
resource models the distinction either — but the underlying biology is contested (mutual exclusivity
per PMID:28955517's discussion versus a measured AFF1–AFF4 association in IntAct). Proposing terms
for a distinction the evidence does not settle would be an over-annotation of the opposite sign; it is
a knowledge gap instead.

affinage assessment

gates_passed: True, faith_pct: 100.0, 23 citations, all well-formed numeric PMIDs, no
PMID:bio_* preprint ids. Precision looks fine. Recall on the reference set that actually decides
this gene's annotations was 0 of 7 — affinage returned none of PMID:41062835, 20159561,
22195968, 22547686, 21729782, 22190034, 23260655, including the 2026 paper behind 7 of the
20 rows. Computed, not estimated, in RESULTS.md §G. This matches the ADIPOQ characterisation: on a
well-studied gene the provider returns the textbook history rather than the annotation-relevant
literature.

Four provider defects beyond recall:

  1. Two cited papers carry unflagged corrections. PMID:28955517 → Correction PMID:32257529
    (2020); PMID:17135274 → Erratum PMID:37777189 (2023). Both found by reading
    CommentsCorrections/RefType on each cited article's own record. Crossref was also checked for
    null-PMID corrections on all seven GOA references: none (update-to and updated-by empty
    throughout; two carry only a has-review relation).
  2. A mechanistic claim contradicted by a cached paper: the DOT1L-recruitment sentence, above.
  3. A hypothesis reported as a demonstration. affinage's row for PMID:12629167 reads "A missense
    mutation in the highly conserved region of mouse Af4 causes autosomal dominant cerebellar
    ataxia". The paper is titled "A mutation in Af4 is predicted to cause cerebellar ataxia and
    cataracts in the robotic mouse" and says
    PMID:12629167.
    The provider's sentence upgraded a stated hypothesis into a causal claim. Caught because the
    builder verifies every quote before writing and refused the paraphrase — which is the whole point
    of making the check a precondition rather than a review step.
  4. A complex-level result attributed to the single gene. affinage's row for PMID:20007461 reads
    "Af4 directly regulates transcription of the Igf-1 gene". The paper attributes it one level up:
    PMID:20007461.
    Same class of error as the one this gene is most exposed to, arriving from the provider rather
    than from GOA.

The last two together are why the reason field for GO:0050877 states the mouse evidence at exactly
the strength the papers state it. Note that neither error would have been caught by quote hygiene
alone had I paraphrased: the ADPRS lesson is that a provider's framing contaminates even when none of
its text is quoted, and here two of its framings were wrong in the same direction — overstating
strength and overstating attribution to the single gene.

No affinage sentence is used as supporting_text anywhere in this review, and no number is taken
from it — the ADPRS lesson is that a provider's arithmetic contaminates even when none of its text
is quoted, so every figure here is re-derived from the primary source or computed by the script.

Disagreement with the concurrent AFF4 review — MEASURED, not predicted

paint/AFF4 (PR #2349) opened while this review was in its sixth round, so the comparison the
brief asks for could be run rather than left conditional. Ten rows are byte-identical between the
two genes (same term, evidence code, reference and WITH/FROM). Seven agree; three diverge.

shared row AFF1 AFF4
GO:0008023 IDA PMID:22195968 MODIFY → GO:0032783 ACCEPT
GO:0006354 IBA GO_REF:0000033 MODIFY → GO:0006368 ACCEPT + separate NEW GO:0006368
GO:0010468 IEA GO_REF:0000002 MODIFY → GO:0006355 ACCEPT
GO:0003712, GO:0005634, GO:0006355, GO:0032783, GO:0050877, and both shared GO:0005515 rows — agree

All three are the same assessment implemented differently — this is a correction to an
earlier draft of this section, which called the GO:0008023 divergence substantive because it had
not read far enough into that row's reason. It still matters, because two reviews recommend
different things for identical rows, but neither review is wrong about the biology.

GO:0006354 — we agree on the biology. AFF4's review independently reaches GO:0006368 as
the right term on its own human evidence, and verified the same ancestor closure. It implements
that as a NEW row while keeping the IBA at ACCEPT; this review implements it as a MODIFY of
the IBA. Convergent conclusion, divergent action. Since the WITH/FROM is byte-identical
(PANTHER:PTN000829417 + UniProtKB:P51825), one node-level recommendation should cover both.

GO:0010468 — both reviews call the term correct-but-uninformative and both note it is a
verified ancestor of terms the gene already holds. AFF4 accepts on that basis; this review
MODIFYs on the upstream argument that IPR007797 supports the more precise term for every
protein it matches. Again a mechanism difference on a shared assessment.

GO:0008023. AFF4's review ran the projection test (does the phenotype spread across the
subunits?) and reports a clean negative, which this review's own data confirms: the paper's
functional term GO:0042795 sits on 7 entities and reaches neither AFF4 nor AFF1. That is sound.

And it also asked the granularity question — an earlier draft of this section said it had not,
which was wrong and is the reason this paragraph exists. Its reason ends "GO:0008023 is a verified
ancestor of GO:0032783, so core_functions records the specific complex and not this parent", and
its core_functions.in_complex is GO:0032783, identical to this review's. So both reviews reached
the same place and differ only in whether the parent row is modified or left standing beside a
specific core_functions entry.

What is genuinely unique to this review is therefore not either gene's verdict but the
measurement behind the upstream recommendation: resolving every recipient's organism shows 10 of
the 11 Drosophila recipients got the child term and none of the 11 human ones did, which is an
argument about the annotation pipeline rather than about AFF1 or AFF4. The projection and
granularity tests are orthogonal, and a negative on the first does not bear on the second.

One checkable correction for that review: it states this term reaches 16 entities and
enumerates 16. The fully-paginated count is 17 — the enumeration omits human ICE2
(Q659A1), which is in the recipient set. Its other two figures for this reference (61
annotations, 26 entities) match this review's computation exactly, so the discrepancy is confined
to that one list, and it does not affect its "no bystander in the list" conclusion, since ICE2 is
a genuine elongation-complex subunit. The likely cause is a mental filter to SEC members: ICE2 is
a little elongation complex subunit.

Two further asymmetries where the reviews agree and should: AFF4 carries GO:0034976 response to ER stress and nucleoplasm/nuclear-body localisations that AFF1 does not, and AFF1 carries the
entire seven-row DNA-damage block from PMID:41062835 that AFF4 does not. Those are real
gene-level differences, not propagation defects.

And the sign trap this family sets, restated because it is the one that would do real damage:
AFF1 and AFF4 have opposite effects on osteogenesis (PMID:28955517), so any term moved
between them in either direction inverts the biology. Neither review does this.

Two findings from merging after AFF4 landed

A byte-identical WITH/FROM field can mean different things on two genes — AFF4's observation, and
it sharpens this review's own. The shared GO:0006354 IBA row carries
PANTHER:PTN000829417|UniProtKB:P51825 on both genes. Verified here rather than taken on report:
the two GOA fields are byte-identical, and P51825 is AFF1 — so on AFF1's row that token is the
target itself (a self-referential IBD, a PAINT curator judging the function core) while on AFF4's
row the same token is a paralogue. Same field, different evidential status. To be precise about
what was and was not already here: this review does identify the seed as self-referential — the
GO:0006355 and GO:0006354 rows label it "human AFF1 itself (self-referential IBD seed)" and the
GO:0006354 summary calls the row self-referential. What it did not draw is the cross-gene
half: that the identical field is paralogue-derived on AFF4, so the same bytes carry different
evidential weight on the two genes. That half is AFF4's, and it does not change the verdict here: the
polymerase-specificity argument for GO:0006368 rests on the node asserting a Pol II-specific
complex term of the same 79 recipients, not on who the seed is. Both positions are on the record in
the two merged reviews, which is the right outcome.

Reviewing two members of one family concurrently duplicates the shared term cache by
construction. Predicted by the campaign brief and observed exactly. AFF4 needed the same two GO
terms this gene needed — GO:0003711 and GO:0032783 — and appended them independently. When AFF4
merged first, git merge origin/main produced a line-based union that kept both copies:
GO:0003711 ×2 and GO:0032783 ×2, caught by cache_lint exiting non-zero and by a multiset
(not set) comparison. A set comparison would have sailed straight through, since both curies were
present.

The resolution was to rebuild rather than merge: take main's file verbatim, append only rows
whose curie main lacks, and assert the four union properties before writing. Here the append set was
empty — main already carried both curies — so cache/go/terms.csv became byte-identical to main
and dropped out of this PR's diff entirely. Note this is not the forbidden
git checkout origin/main -- cache/go/terms.csv clobber: that is unsafe when the branch holds rows
main lacks, and the precondition set(mine) - set(main) == {} was checked before relying on it.

The same-folder PANTHER artefacts collided the same way, harmlessly: both agents fetched
PTHR10528 about 13 milliseconds apart (…51.732839 vs …51.745794, a 12 955 µs gap — an
earlier draft of this bullet said "microseconds", off by a factor of a thousand), so
PTHR10528-metadata.yaml conflicted add/add on its fetched_date line alone — verified as the only difference by diffing the two blobs before
accepting main's copy, with PTHR10528-entries.csv byte-identical. Five publications AFF4 also
fetched were byte-identical too, so no full-text downgrade was possible on any of them.

Where the review effort actually went

For whoever reviews AFF2, AFF3 or AFF4 next. This section has been rewritten twice and the file's
retrospective claims corrected three times: one draft claimed no curation field was ever challenged
(false); one stated counts it could not enumerate, in a section whose leading defect class is "a count
disagreeing with its own enumeration"; and the AFF4 comparison above mis-stated what the sibling
review had done (also false, and corrected there rather than here — the distinction matters, because
an earlier version of this paragraph counted that one as a correction to this section, which it was
not). So this version enumerates and does not count: every list below is the list, which removes
the failure mode rather than correcting its latest instance.

What is true, and it is the load-bearing claim. Diffing the review YAML from the commit that
first created it (83165cac) to HEAD, exactly two structured-field lines change anywhere in the
document: the id and label of GO:0006355, removed from
core_functions[0].directly_involved_in. No action, evidence_type, term, qualifier,
original_reference_id or supporting_text moved anywhere, at any point, across the whole PR. Every
proposal — the seven MODIFY replacement terms and the one NEW row, eight in all — stands exactly
as first written, as does every quote and every biological claim.

The one finding that touched a curation field: core_functions[0].directly_involved_in carried
GO:0032968 together with its ancestor GO:0006355; the ancestor was removed in 72c74fa. An
earlier draft of this section wrote that out of the record to tell a tidier story, which is the
failure the rest of the section is about.

Everything else was a claim about the checks, not about the biology. Enumerated:

How they were found, which is the part worth carrying forward. Deriving an expected number
independently and comparing found: the recipient-set count, the disorder figure, the
publicationIdentifiers impossibility ("1 record, 5 publications"), and the 5-of-104 IntAct records
silently dropped by an accession equality test. Two more surfaced only because a guard's own
break-test failed — a case-sensitive number match that let "ALL ELEVEN" through, and an enumeration
regex that required the correct gene list, so substituting a symbol made the check go silent rather
than fire. Those two are the entry to keep: a passing self-test proves the guards you thought of
fire, and coverage stays a reading question. The rest came from the PR reviewer.

One process failure of mine, recorded because it is the kind that repeats. After posting a
"stopping here" comment I pushed a further change without reading the review that had landed in
between, so one correction sat unaddressed for a whole round. A stopping criterion is not a reason
to stop reading the thread.

Process

Every claim about provenance in the review YAML is computed by
AFF1-bioinformatics/analyze_aff1_annotations.py (nine analyses, guards break-tested with
--self-test). Two real defects were found by numbers that refused to add up while writing it:
publicationIdentifiers produced "1 record, 5 publications" because its entries are
"32296183 (pubmed)" strings mixed with DOI/IMEx ids; and an equality test against the bare
accession silently discarded 5 of 104 IntAct records because the subject also appears as isoform
P51825-3 and as its own Ensembl transcripts. Both are now assertions rather than comments.