AFF3 (P51826) — review notes

Reviewer journal for the PAINT + affinage campaign. Third and last member of the AFF
family to be reviewed here, after AFF4 (PR #2349) and AFF1 (PR #2348), both of which merged
to main while this review was being written.

What the gene is

AFF3 / LAF4 is a 1226-residue, almost entirely intrinsically disordered nuclear protein
of the AF4/FMR2 (ALF) family — PTHR10528, subfamily PTHR10528:SF16
[file:human/AFF3/AFF3-uniprot.txt "DR PANTHER; PTHR10528:SF16; AF4_FMR2 FAMILY MEMBER 3; 1."].
It is the scaffolding subunit of SEC-L3, one of the four paralogous super elongation
complexes: PMID:22547686 reports
"biochemical isolation of SEC-like 2 (SEC-L2) and SEC-like 3 (SEC-L3) containing AFF2 and
AFF3 in association with P-TEFb, ENL/MLLT1, and AF9/MLLT3", and
"The SEC family members demonstrate high levels of polymerase II (Pol II) C-terminal domain
kinase activity; however, only SEC is required for the proper induction of the HSP70 gene
upon stress". Note the negative in that same sentence — AFF3/SEC-L3 is not
interchangeable with AFF4/SEC.

Where AFF3 differs from AFF1/AFF4 is in how it reaches chromatin. It is recruited by
sequence-specific zinc-finger proteins: PMID:28180295 shows
"ZFP281 recruits AFF3 to the Meg3 enhancer within the imprinted Dlk1-Dio3 locus, thus
regulating the allele-specific expression of the Meg3 polycistron" (mouse ES cells) and
concludes "different zinc finger proteins can recruit AFF3 to different regulatory elements
and differentially regulate the function of AFF3 in a context-dependent manner" (ZFP57 at
the IG-DMR). In human cells it binds a methylated CpG island directly:
PMID:30535390 (HEK293T + IMR-90)
"Quantitative PCR after chromatin immunoprecipitation (ChIP-qPCR) demonstrates that AFF3 is
associated with the XIST DMR in both IMR-90 and HEK293T", and
"Knockdown of AFF3 leads to de-repression of the inactive allele of XIST in terminally
differentiated cells".

Its own DNA binding is not sequence-specific: PMID:8555498
"In vitro-translated LAF-4 was able to bind strongly to double-stranded DNA cellulose",
and the same paper localises it — "Antibodies to LAF-4 showed it to be a nuclear protein
that showed an uneven, granular immunofluorescence pattern".

The single most useful measurement: WITH/FROM resolution

AFF3-bioinformatics/resolve_withfrom.py (output withfrom_resolution.json) resolves every
token on all five IBA rows and then asks what evidence each donor itself carries for the
propagated term. Results:

token resolves to reviewed own experimental evidence for the propagated term
MGI:MGI:106927 P51827 mouse Aff3 — the true orthologue Swiss-Prot IDA + IMP GO:0006355 (PMID:25162227)
MGI:MGI:1100819 O88573 mouse Aff1 (paralogue) Swiss-Prot IDA GO:0045893 (positive)
MGI:MGI:1202294 O55112 mouse Aff2 (paralogue) Swiss-Prot IMP GO:0007611 (PMID:11923441)
FB:FBgn0041111 Q9VQI9 Drosophila lilli (single fly AFF co-orthologue) Swiss-Prot IMP GO:0003712; IMP GO:0007611; IPI GO:0032783; IMP/IGI GO:0006355
UniProtKB:P51825 P51826's paralogue, human AFF1 Swiss-Prot EXP GO:0006354 (PMID:22547686)
PANTHER:PTN000829417 PANTHER tree node — not a protein n/a unqueryable

All MGI/FB tokens returned multiple UniProt hits (10, 3, 3); the script reports every
candidate and queries the Swiss-Prot one, and it prints reviewed status using
entryType.startswith("UniProtKB reviewed") because "reviewed" in entryType also matches
unreviewed.

The brief's question 3, resolved. AFF1 and AFF4 both carry a byte-identical
PANTHER:PTN000829417|UniProtKB:P51825 WITH/FROM on GO:0006354, self-referential on AFF1
and paralogue-derived on AFF4. AFF3 has the same field, and on AFF3 it is
paralogue-derived — P51825 is AFF1, P51826 is AFF3. So the same six bytes now carry
three different evidential meanings across three reviews. Recorded for the family write-up.

The finding worth reporting upstream: PMID:22547686 annotated the wrong paralogue

reference_projection.py queries QuickGO by reference and counts distinct gene products
(entities, not annotations; pagination asserted against len(results)).

PMID:22547686 — the paper whose headline result is the isolation of the AFF3-containing
SEC-L3 — has produced exactly one annotation in the whole of GOA: GO:0006354 EXP on
UniProtKB:P51825 (AFF1), the paralogue the paper explicitly contrasts with SEC-L2/L3.
AFF3 got nothing directly and then received the term back second-hand, as an IBA whose
WITH/FROM points at AFF1.

This is the same shape as AFF4's sharpest datum (PMID:20159561, titled for AFF4, produced one
annotation and it was on AFF1). Two independent instances in one family, both with AFF1 as the
recipient. Filed as a suggested_question.

Same script, negative results worth recording:

GO:0034612 response to tumor necrosis factor — REMOVE

The one clear error, and it is exactly the failure mode a disease gene invites: a clinical
drug-response association read as a cellular response-to-stimulus.

PMID:20444755 is a pharmacogenetic study.
"Eighteen single nucleotide polymorphisms (SNPs) mapping to 11 genetic loci were genotyped in
1012 patients with RA receiving treatment with etanercept, infliximab or adalimumab", with the
outcome "the absolute change in 28 joint count disease activity score (DAS28) between baseline
and 6-month follow-up". The result is that
"SNPs mapping to AFF3 and CD226 had a statistically significant association with the response
to anti-TNF treatment".

Three independent grounds:

  1. Definitional. GO:0034612 requires "a change in state or activity of a cell or an
    organism ... as a result of a tumor necrosis factor stimulus". The study applies TNF
    blockade to patients. The direction of the perturbation is the opposite of the term's,
    and there is no TNF stimulus anywhere in the paper.
  2. Evidence code. IMP requires a mutant phenotype. A common tag SNP associated with a
    6-month DAS28 change is neither a mutant nor a phenotype of the gene product.
  3. No assay. The paper contains no cellular or biochemical experiment on AFF3 at all;
    it is genotyping plus multivariate linear regression.

It has propagated: mouse Aff3 (P51827) holds GO:0034612 only by ISO GO_REF:0000119
with UniProtKB:P51826 — i.e. from this human row. Removing the human row retracts it from
both species, which is why it is worth acting on rather than leaving.

Erratum status, per the campaign's correction rule: PubMed's record for PMID:20444755 carries
an ErratumIn reference to Ann Rheum Dis. 2011 Aug;70(8):1519 whose PubMed id is null
(see corrections.json, produced by corrections_check.py). It is not resolvable: Crossref
returns empty relation/update-to/updated-by for 10.1136/ard.2009.118406, and Europe PMC
returns 0 hits both for that journal/year/page and for a citing Published Erratum. So its scope
is unestablished. The verdict does not rest on the erratum — it rests on the study design
stated in the paper's own full text — and I have deliberately not let an unresolvable correction
manufacture a hedge (the mistake AFF4 made in the other direction).

That check ran over all 31 PMIDs cited anywhere in the review YAML, these notes, or the
affinage record (corrections.json records the list): 1 of 31 flagged, and no
retractions and no expressions of concern. The denominator has been 17, then 29, and is now
31 as these notes accumulated the donor, IntAct and review-round references; every number in
the committed artifacts is read from corrections.json rather than remembered, which is the
only reason it has stayed correct through three changes.

GO:0035116 embryonic hindlimb morphogenesis — well founded, kept as non-core

My first instinct was that a single case report of a 500-kb deletion could not support a
gene-specific IMP. That was wrong, and an independent paper says so:
PMID:25162227 states "A human microdeletion of 500 kb on chromosome 2q11.1 encompassing only
the LAF4 gene has been detected by array comparative genomic hybridization on peripheral
lymphocytes". So it is not a contiguous-gene deletion.

The evidence then triangulates, and it is specifically the lower limb:

GO:0035116's definition is "the anatomical structures of the hindlimbs ... the posterior
limbs of an animal" — no taxon problem for a human lower limb, and the term is more apt than
the parent because both the patient and the knockin mice are lower-limb.

The dosage point, and why it does not change the verdict. AFF3's limb phenotypes arrive
from opposite directions — haploinsufficiency (deletion; promoter GCC silencing) and
gain-of-function degron stabilisation (KINSSHIP). PMID:38811945 frames this as
"minute changes in AFF3 function are deleterious"; both homozygous LoF and homozygous KINSSHIP
isogenic lines perturb "more than a third of the AFF3 bound loci". A gene whose increase and
decrease both break limb development is participating in limb development either way, so
involved_in is right. What the evidence does not support is any claim about the KINSSHIP
protein's activity being AFF3's normal activity — KINSSHIP is a degradation-resistance defect,
not a new function — and nothing in this review annotates from the mutant protein's behaviour.
The same discipline applies to the MLL–AFF3 and RUNX1–AFF3 leukaemic fusions
(PMID:12203795, PMID:12743608, PMID:17968322): a fusion oncoprotein's behaviour is not
the wild-type protein's function, and no GO row in this review derives from a fusion.

GO:0050877 nervous system process — the IBA is less precise than its donors, and that is correct here

resolve_withfrom.py shows both protein donors carry the same, lower experimental term:
mouse Aff2 GO:0007611 IMP (PMID:11923441, Fmr2-knockout conditioned fear) and fly lilli
GO:0007611 IMP (PMID:18310460). term_relations.py confirms GO:0007611 is a
descendant of GO:0050877, so the propagation landed two levels above two agreeing donors —
the ACRV1 shape.

But refining it downward would be wrong. AFF3 has no learning-or-memory data of its own, and
its documented nervous-system involvement is developmental: cortical migration in mouse
(PMID:25162227 — "we discovered that Laf4 is required for cortical cell migration"), ID and
seizures in KINSSHIP, and a GCC promoter expansion associated with
"a 2.4-fold reduced probability of completing secondary education" (PMID:39313615).
term_relations.py also verifies that GO:0001764 is NOT a descendant of GO:0050877 —
the developmental branch is not reachable from this term at all. So the honest reading is:
the row's propagation route is weak, the term itself stands on AFF3's own human genetics, and
the specificity AFF3 actually deserves lives in a different branch, proposed separately as
GO:0021795. Verdict KEEP_AS_NON_CORE, agreeing with both AFF1 and AFF4 but for a measured
reason rather than by convention.

Sibling comparison: where I agree with AFF1/AFF4 and where I do not

Checked against origin/main:genes/human/AFF4/... and, initially,
origin/paint/AFF1:genes/human/AFF1/.... AFF1 merged (commit 92eb534fe) partway through, and
the merged file is byte-identical to the branch head the comparison below was built from, so
nothing had to be redone.

row AFF1 AFF4 AFF3 (this review)
GO:0006355 IBA ACCEPT ACCEPT ACCEPT
GO:0003712 IBA ACCEPT ACCEPT ACCEPT
GO:0032783 IBA ACCEPT ACCEPT ACCEPT
GO:0005634 IEA GO_REF:0000120 ACCEPT ACCEPT ACCEPT
GO:0050877 IBA KEEP_AS_NON_CORE KEEP_AS_NON_CORE KEEP_AS_NON_CORE
GO:0006354 IBA MODIFY → GO:0006368 ACCEPT MODIFY → GO:0006368
GO:0010468 IEA GO_REF:0000002 MODIFY → GO:0006355 ACCEPT ACCEPT

The two siblings disagree on exactly the two rows where I had to choose, and they disagree in
opposite directions, so "follow the family" gives no answer. The rule I used instead, applied
to both:

MODIFY a broad term only when the more specific term the donor evidence supports is not
already annotated on this gene. If it is, the broad row is a redundant ancestor and MODIFY
buys nothing.

So the divergences are not stylistic; each follows from a stated rule, and the rule reproduces
AFF1 on one row and AFF4 on the other.

Where I diverge on core_functions shape, and why. AFF1 and AFF4 independently put the
scaffolding in molecular_function as GO:0030674 and the complex's catalytic contribution in
contributes_to_molecular_function as GO:0003711. I keep the second and place GO:0003712
alongside GO:0030674 as the first core function's MF, because AFF3's own best-evidenced
molecular role is the coregulator one, and GO:0003712's usage comment describes it exactly:
"Most transcription coregulators do not bind DNA. Those that do usually bind DNA either in a
non-specific or non-direct manner." AFF3 binds dsDNA non-specifically in vitro, is recruited
by ZFP281/ZFP57, and has no sequence motif of its own. AFF1/AFF4 had bare GO:0005515 IPI rows
that needed an informative replacement and mapped interfaces to justify it; AFF3 has zero
GO:0005515 rows in GOA, so nothing pushed me toward GO:0030674 from that direction —
I propose it on independent grounds (below) instead of inheriting it.

term_relations.py caught me getting this wrong on the first pass: I assumed GO:0030674 was
a descendant of GO:0005515. It is not — its only ancestors are GO:0003674 and GO:0060090.
Proposing it is therefore a move into the molecular-adaptor branch, not a refinement of protein
binding, and the guard is what told me so.

Two annotations that exist in UniProt but not in GOA — same route, opposite verdicts

AFF3's UniProt entry carries two GO cross-references that GOA does not have (QuickGO returns
exactly 11 annotations for P51826, matching the GOA TSV):

DR   GO; GO:0003700; F:DNA-binding transcription factor activity; IEA:Ensembl.
DR   GO; GO:0003690; F:double-stranded DNA binding; IEA:Ensembl.

Both are Ensembl-Compara projections from mouse Aff3, and they should be handled oppositely:

A related stale projection, noted while querying the donor: mouse Aff3 holds
GO:0016604 nuclear body and GO:0005829 cytosol by ISO GO_REF:0000119 from human
P51826, while human AFF3 currently carries neither. The mouse rows are reflections of human
annotations that no longer exist.

Automated-route provenance (checked, and mostly clean)

IntAct: 14 records, and the reason there is nothing to adjudicate

AFF3 has no GO:0005515 rows in GOA at all, so there are no per-partner verdicts to make.
Queried IntAct anyway (findInteractions/P51826, 14 records, all returned):

Two of these — CDK9 and MLLT1 — are the two SEC modules AFF3 is supposed to bridge, replicated
across independent studies and methods, and GOA has curated none of them. That is the
under-curation diagnosis AFF4 established, reproduced here from a different instrument.

Deep research provider

AFF3-deep-research-affinage.md: gates_passed: True, 15 citations, all numeric PMIDs, no
PMID:bio_* preprint ids. Recall was good on this gene — unusually so, given the campaign's
experience — but two things it did not supply were decisive and both had to be found elsewhere:

  1. PMID:22547686 (the SEC-L3 isolation paper) is absent from the affinage citation list.
    It surfaced only from QuickGO, as the reference behind human AFF1's GO:0006354 EXP — i.e.
    by following the WITH/FROM one level deeper, which is the brief's own recipe.
  2. PMID:18616733 and PMID:20444755, the two papers behind AFF3's only two
    experimental BP rows, are both absent from the affinage record. The provider returned the
    gene's notable literature; the annotation-relevant literature was a different set. Same
    characterisation as ADIPOQ.

Also worth recording: the affinage narrative says AFF3 "binds double-stranded DNA and carries a
domain that strongly activates transcription", citing PMID:8555498. That is faithful. But the
GAL4-fusion result it alludes to ("both LAF-4 and AF-4 had domains that activated transcription
strongly when fused to the GAL4 DNA-binding domain") is a heterologous transactivation assay,
which is why this review annotates coregulator activity rather than transcription factor
activity from it.

Process

The number that refused to add up, and the second retraction

After the PR was opened I re-derived a figure I had counted by eye. The review said IntAct
records "CDK9 in five records across four distinct publications and four distinct
methods with a MI score of 0.73". intact_partners.py computes 6 records, 5 publications,
3 methods, MI in {0.35, 0.73}. Every one of the four numbers was wrong — the publication
count too low, the method count too high, and the MI presented as uniform when one record
(an isoform-2 pairing) sits at 0.35.

Two things worth recording about how it happened and how it was fixed:

  1. The wrong version was produced by an ad-hoc query that deduplicated on a composite key,
    which silently dropped one CDK9 record, and then by counting the printed lines by eye. The
    corrected figures come from a committed script that asserts the subject is not in its own
    partner set and that every record IntAct reports is retrieved.
  2. The repair script's first anchor did not match, because the notes use × and — where
    I had typed x and -. The anchor assertion caught it rather than the edit silently
    landing in 4 of 5 places — which is exactly the hyphen/en-dash trap the campaign brief
    warns about for quotes, arriving here in a find-and-replace instead.

The guard against the recurrence is structural, not a phrase pin: audit_claims.py parses
any prose stating CDK9's counts and compares them against intact_partners.json, so a
rewording cannot evade it. It is break-tested against git show HEAD:... — the version that
actually shipped — and has a vacuity direction that fires if the sentence disappears
altogether. The correction direction is favourable: 5 independent publications is a stronger
replication claim than 4.

A third retraction, and the same lesson: the guard caught a claim I had already shipped

The GO:0006355 row's original reason argued that the row should stay at the unsigned parent
"because the donors disagree in sign — mouse Aff1's descendant is positive regulation while
human AFF1's is in the negative-regulation-of-elongation branch". I had read GO:0032786 as
negative by its proximity to GO:0032785. Adding it to term_relations.py refuted it
immediately: GO:0032786 is positive regulation of DNA-templated transcription, elongation,
a verified descendant of GO:0045893. So every signed donor on that row points the same way,
and GO:0045893 is itself a descendant of GO:0006355, meaning a positive child was available
and unused. The AEBP2 donor-disagreement test does not apply at all.

The verdict did not change, but the reason had to. What actually forbids refining the row is the
recipient, not the donors: AFF3 represses XIST from the silent allele in two human lines
while establishing a permissive state at the Meg3 enhancer with ZFP281 and raising 84% of the
transcripts it changes on over-expression in mouse cortical cells. A positive-only term would be
false for the repressive half — which is why the specific negative instance is a separate
GO:0045892 row rather than a modification of this one.

Three retractions on this gene, and the common shape is worth naming: each was an inference
from an identifier's or a label's neighbourhood rather than from a fetched fact.
GO:0030674 was assumed to be under GO:0005515 because adaptor activity sounds like a kind
of protein binding; GO:0032786 was assumed negative because it sits one integer from
GO:0032785; and the IntAct counts were read off a printed list instead of computed. In all
three cases the fix was to write the check, and in all three the check fired.

Review round 1 (PR #2351, ai4c-reviewer): approved, six suggestions

Verdict was approve with no critical or important issues. Each suggestion's premise was
checked before conceding, and one was declined with evidence rather than accepted.

Accepted, and each a real defect.

  1. GO:0030674 was an ill-formed IPI. For IPI the WITH/FROM field takes the interactor, and
    the row had none. Added CDK9 (UniProtKB:P50750) and cyclin T1 (UniProtKB:O60563), both
    confirmed reviewed human entries by asserting primaryAccession on the fetch. The reviewer's
    alternative — recode as ISS on AFF1/AFF4, whose interfaces are mapped — was declined with a
    stated reason: the assay is a co-IP of AFF3 itself in human cells, so IPI is the correct code
    and what was missing was the entity, not the code. The unmapped-interface caveat belongs in
    the reason and in the knowledge gap, which is where it already was.
  2. GO:0003711's reason asserted a sibling relation on one leg only and read a kinase
    measurement as an elongation-factor claim without saying so. Both fixed: the inference is now
    named as an inference, and term_relations.py checks GO:0003711 under GO:0140110 as well
    as GO:0003712.
  3. GO:0001822's second human quote is "urogenital tract malformations", which is broader
    than the kidney. The reason now says the mouse null is doing the work and the human quote
    corroborates the organ system rather than the organ.
  4. Two follow-ups the review was positioned to file but had not. The GO:0034612 removal's
    downstream consumer (mouse Aff3's ISO row) is folded into the MGI question so the retraction
    is actionable; and AFF4's merged review asks whether GO:0032783 should reach AFF3 at all on
    the premise that AFF3 "has not been shown to be a subunit of that complex" — I verified that
    text at AFF4-ai-review.yaml before asserting it — which AFF3's own human biochemistry
    refutes, so that correction is now filed too.

Half accepted, and the half that was wrong is the interesting part. The reviewer observed
that the GO:0050877 reason grounded the row in AFF3's human genetics while the same review
argues that AFF3's developmental nervous-system role is off-branch — so the grounding cited
evidence the review says belongs elsewhere. Checking it split the claim: GO:0050890 cognition
IS a descendant of GO:0050877, and GO:0007611 sits one step under cognition, so the
intellectual disability, seizures and the education association are legitimately inside the
term; GO:0021795 is not a descendant, so only the migration evidence is off-branch. The
reason had lumped the two halves together. It now rests primarily on the two donor IMPs, cites
the cognitive phenotypes as on-branch corroboration, and explicitly assigns the migration
evidence to the separate proposed row.

Declined, with evidence. The reviewer suggested GO:0030674 or GO:0003712 might capture
the class-switch core function better than GO:0003690. Both would over-claim, and the paper
says so itself: GO:0030674 would assert a bridge between switch-region DNA and AID, but
PMID:36001653 states "While the mechanism by which AID is recruited to switch regions is still
unclear, the following mechanisms have been proposed" and offers cohesin and P-TEFb as
alternatives — "AFF3 may regulate CSR by facilitating the interaction of AID with cohesin
factors". And GO:0003712 belongs to transcription regulation whereas class switch recombination
is a DNA recombination reaction. What was measured is switch-region occupancy: "We detected
significant increases in the signals near the switch regions of IgM and IgG1, indicating that
AFF3 can bind to these regions". Those three quotes are now in the row's supported_by, so the
refusal is itself CI-checked rather than asserted in prose.

A measurement the review round added. Asked how large the GO:0005515 coverage gap is,
one QuickGO call over all three paralogues answers it with its own positive controls: AFF4 15
rows, AFF1 3, AFF3 zero — so the endpoint works, the term is alive for the siblings in the
same request, and AFF3's zero is a real absence rather than a rejected query. Given that AFF3's
CDK9 contact is the most replicated of the three in IntAct, that is a curation asymmetry rather
than a biological one.

Review round 2: five carried-over items were all the same failure

The round-2 review requested changes for one new issue plus five carried over, and the five
share a single cause worth naming: round 1 fixed each claim on the surface the reviewer had
named and left it standing on a low-salience one. The surviving instances were a
reference_review note, a source_entities comment, a script docstring, a notes bullet and the
top-level description — precisely the surfaces nobody re-reads.

Worse, my own guard was structurally unable to catch two of them. fix_sign_claim.py's
narration exemption was file-scoped: it exempted AFF3-ai-review.yaml wholesale so that the
file could narrate its own retraction, which made the re-grep blind to every unnarrated
instance inside that same file. Two survived there. The exemption is now per-occurrence — a
sentence asserting the retracted premise passes only if the surrounding window marks it as
retracted — and the check has moved out of the one-shot script into audit_claims.py, so it runs
on every audit rather than once. It is break-tested against git show 2bf0d3d5e:..., the commit
that actually shipped the two survivors, and has a companion direction asserting it stays silent
on the narrated retractions that remain by design.

That is guard-defeat mode number ten for this campaign, and it has a name of its own:
an exemption coarser than the thing it exempts. The narration exemption was correct in
intent and one scope-level too wide in implementation, which made the guard report success over
exactly the file it was written for.

The one new issue, partly declined. Round 1's GO:0001822 commit — whose stated purpose was
to stop over-reading the human evidence — changed "named partly for horseshoe kidney" to
"horseshoe or hypoplastic kidney". The reviewer is right that PMID:33961779 spells the
acronym "KI for horseshoe kidney" and never says hypoplastic. But the phrase is not unsourced:
UniProt's DISEASE line for KINSSHIP reads
[file:human/AFF3/AFF3-uniprot.txt "facial features, horseshoe or hypoplastic kidney, and failure to"].
So the defect was attribution, not fabrication — the phrase sat next to a PMID quote and read
as the paper's. It is now attributed to UniProt explicitly on both surfaces. The lesson is still
the reviewer's, and sharper for being narrower: a commit that tightens a claim can loosen it in
the same breath, and the added words came from a different source than the quote beside them.

The other four: :414's "all verified" was vacuous for GO:0045893 (verified under itself) and
untrue of GO:0032968, whose ancestry claim was simply missing from the guard — now present, 20
claims; the speckle P-TEFb co-concentration is qualified on both surfaces, since the
redistribution of CDK9 and cyclin T1 to AFF3 sites required strong over-production and only
AFF3's own speckle localisation is a baseline observation; intact_partners.py's docstring still
said "5 records" a full round after the prose had been corrected to 6; and PIP4K2A was listed as
a singleton when it has 2 records across 2 publications.

Review round 3: approved, three tooling suggestions taken, and a stopping criterion

Round 3 approved on the head and raised three non-blocking suggestions. All three were taken,
because each closed a real hole rather than rewording prose:

  1. The new UniProt file: quote lived only in AFF3-notes.md, which verify_file_quotes.py
    does not walk — so the one quote added to answer a fabrication-surface criticism was itself
    outside the gate. It is now in the GO:0001822 row's supported_by, inside both quote
    checkers (15 file: quotes verified byte-exact, 75 total).
  2. The two prose guards were scoped to the review YAML while the folder-wide sweep lived
    inside a one-shot repair script — which is precisely how the intact_partners.py docstring
    and the PIP4K2A bullet survived a round. audit_sibling_surfaces() now runs both guards over
    every .md, .yaml and .py in the gene folder on every audit, and it fails loudly if it
    scans zero files.
  3. The GO:0003711/GO:0003712 reciprocal non-containment pair was unchecked, unlike the
    equivalent GO:0003700/GO:0003712 pair. Both directions now asserted — 22 claims.

Widening the sweep immediately produced ten false positives, and that is the interesting
part. Repair scripts must contain the retracted strings as their find anchors, and the guard
must contain them as patterns — so a naive folder-wide sweep forbids the guard from coexisting
with its own implementation. The exemption vocabulary was widened to cover anchors, patterns and
fixtures, kept strictly per-occurrence (never per-file, the mode that failed in round 2), two
find-anchors were labelled as such in fix_intact_counts.py, and both directions are
break-tested: the sweep fires on an unnarrated sign claim injected into the notes, and stays
silent on the anchors and narrated history that remain by design.

The synthetic IntAct-count break-test then failed for the right reason and had to be repaired:
the per-occurrence exemption legitimately suppressed the mutation, because the first of the two
CDK9 count sentences sits inside a window explaining the retraction. The test now selects an
un-narrated occurrence explicitly, asserts one exists, and uses a raw anchor unique to it — the
raw and whitespace-normalised forms differ because the phrase is line-wrapped differently at the
two sites. A break-test that fails because the exemption worked is not evidence the guard is
broken, and distinguishing the two took reading the window rather than the result.

Stopping criterion. Rounds 2 and 3 changed no GO term, no action, no evidence code, no
quote and no reported number in the curation itself — every item was about where a claim was
written, which surface a guard scanned, or which ancestry pair was pinned. The curation content
has been stable since 2bf0d3d5e. So: I will still fix anything that misstates a number,
misattributes a source, or lets a guard report coverage it does not have. I will not keep
refining the guards' own prose, and I am not looking for further tooling symmetry for its own
sake. If a later round finds a factual defect in the annotations, that is a different matter and
gets fixed.

Review round 4: the fix for round 3's false positives created a coverage hole

Round 4 requested changes for one item, introduced by the round-3 commit and squarely inside
that commit's own stated stopping criterion ("lets a guard report coverage it does not have").
The reviewer is right, and the diagnosis is worth recording because it is the same named mode a
third time, one level over each time.

What happened. Widening check_intact_counts' narration exemption fixed the ten false
positives on the repair scripts — but the function is called from both audit() (the review
YAML) and audit_sibling_surfaces() (the scripts and notes), and only the second needs the
exemption. The consequence is concrete: the YAML states the CDK9 counts about 170 characters
before "the hand-counted version of them was wrong", so that occurrence became exempt from the
numeric check — on the one file the guard exists for, at the sentence that announces the
correction.

Demonstrated against the shipped code, not argued. Extracting audit_claims.py and
AFF3-ai-review.yaml from 710a315df, reverting the counts to the wrong 5/4 at that narrated
occurrence, and running the shipped guard reports 0 problems. The current guard fires
there, and a new break-test direction pins it.

The fix. allow_narrated now defaults to False and is switched on only by the sibling
sweep. Verified before restoring the default that neither of the YAML's two count statements
quotes the retracted numbers, so nothing on that surface needs the exemption. And the vocabulary
is split by surface: the prose set is back to the original narrow retraction words, while the
code-shaped tokens (anchor, pattern, fixture, guard, premise, startswith, finditer,
re.compile) apply only on .py files — the reviewer's secondary point, that guard and
premise are ordinary words in this document and would have exempted a future unnarrated sign
claim written near either.

Three instances of one mode, at three scopes. Round 2: an exemption coarser in file
scope. Round 3: coarser in surface scope. Round 3 again, secondarily: coarser in
vocabulary. Each time the exemption was correct in intent and one level too wide in
implementation, and each time the symptom was a guard reporting success over exactly what it was
written to protect. The generalisable form: when you widen an exemption to stop a false
positive, ask which callers need it and give it to those callers only — the fix for a false
positive is almost always narrower in scope than the false positive's cause.