TreeGrafter rejection re-review, 2026-09-24

This audit re-examines every TreeGrafter annotation (IEA / GO_REF:0000118)
whose gene review currently records action: REMOVE or
action: MARK_AS_OVER_ANNOTATED. The question for each row is whether that
rejection stands on stated biological grounds, or whether it should be relaxed
to ACCEPT, KEEP_AS_NON_CORE, MODIFY, or UNDECIDED.

Genes whose rejected rows were already re-assessed in the
2026-09-20 audit
(status: reviewed) are excluded here; those decisions stand as recorded
there.

Scope is the flagged rows only (scope: treegrafter_rejections), not a
full-gene re-review. Other annotations in the same gene are not re-adjudicated
unless a decision on a flagged row logically requires it (for example, a
MODIFY whose replacement term is already carried by another row).

Decision rules

Taken from CLAUDE.md and the 2026-09-20 audit README:

1. REMOVE needs positive contrary evidence. It is appropriate for an
   electronic propagation that is demonstrably wrong: reaction chemistry or
   substrate that the target cannot perform, a domain architecture the target
   lacks, lost catalytic residues, a pathway absent from the organism, or a
   direct experimental contradiction. "Family-level propagation" or "no
   target-specific assay" is not, on its own, grounds for REMOVE; the absence
   of a target experiment does not refute a supported phylogenetic inference.
2. Broad but true terms are not over-annotations. cytoplasm or cytosol
   on a soluble bacterial enzyme is correct; redundancy with a more specific
   sibling does not make it wrong. Such rows go to ACCEPT (or
   KEEP_AS_NON_CORE when the location is real but peripheral). A generic
   location that is incompatible with the protein (e.g. cytosol on an
   integral membrane transporter with no soluble pool) can stay rejected.
3. Too-coarse terms are MODIFY, not REMOVE. When the propagated term
   is a true ancestor of the gene's real activity, or a sibling that names the
   right chemistry with the wrong specificity, use MODIFY with a
   proposed_replacement_terms entry, or ACCEPT when the specific term is
   already carried by another row and the ancestor is simply true.
4. MARK_AS_OVER_ANNOTATED is for terms that are not false but attribute
   more than the evidence supports (e.g. a whole-pathway process term on a
   single enzyme whose product could feed that pathway). When even that is
   uncertain, prefer UNDECIDED over asserting a rejection.
5. Verify the term definition (QuickGO
   https://www.ebi.ac.uk/QuickGO/services/ontology/go/terms/GO:NNNNNNN) and
   the target's own evidence: *-uniprot.txt (catalytic activity, domains,
   InterPro/PANTHER cross-references), *-goa.tsv (other evidence for the same
   or related term), the notes and cached deep-research files, and cached
   publications. Do not invent evidence, PMIDs, or term ids.

Record format

One batch YAML per reviewer batch (batch-NN.yaml), top-level date, scope,
and genes. Each gene entry records gene, gene_file, status
(reviewed), scope: treegrafter_rejections, outcome, and an annotations
list with term_id, term_label, evidence_type, original_reference_id,
previous_action, action, outcome (retained when action unchanged,
changed otherwise), and a rationale written from the evidence considered.
When the review file was edited, the gene also gets an append-only history
record scaffolded with just new-history.

The two outcome fields mean different things, and the gene-level one is
defined here because the first pass left it implicit. Both describe
adjudication, not file edits:

• annotation-level outcome: retained when previous_action == action,
  changed otherwise.
• gene-level outcome: changed when any of the gene's annotations
  changed action, confirmed when none did.

So a gene whose actions all stand is confirmed even if its reason prose was
rewritten — strengthening a rationale is not a change of adjudication, and the
history record is what records the edit. The first pass applied this field
inconsistently (some entries tracked the file rather than the decision); PR

3165 settled on the definition above and brought every entry into line with

it. check_outcomes.py enforces it, so the next audit inherits a rule rather
than re-inferring one. summarize.py reads only the annotation-level field, so
the gene-level value is documentary.

Later reviews that overturn an adjudication

This audit's records are a snapshot of what it decided. A later gene review can
legitimately reach a different action on the same TreeGrafter row; the ProtNLM2
gene re-review campaign did so for three batch-01 rows. Do not rewrite the
recorded action or outcome for those rows: they describe this audit's
decision. Declare the later decision on the annotation instead:

    superseded_by:
      action: UNDECIDED          # the live review's action
      by: 'PR #5304'
      date: '2026-10-23'
      reason: 'Why the later review (PR #5304) differs.'

Quote every value in this block. In an unquoted YAML value a # preceded by a space
starts a comment, so reason: ... (PR #5304): ... silently loads as ... (PR. The
checker therefore requires by to name a PR (PR #NNNN) and reason to end with
sentence punctuation, so a truncated value fails the gate instead of passing it.

check_outcomes.py then checks the live row against superseded_by.action
rather than action, and lists each such row as superseded, so a further
drift still fails the gate. summarize.py is unaffected: it counts this
audit's own transitions.

What the verbatim check does and does not cover

This audit's rules lean on supporting_text being a verbatim substring of its
source, and round 3 caught a quote composed to fit a term exactly that way. The
check is not symmetric across reference types, and the asymmetry is worth
stating because it is easy to mistake a passing just validate for a verified
quote:

• PMID: / DOI: quotes are gated. validate_reference_finding_supporting_text
  runs on LITERATURE_PREFIXES only, and a non-verbatim snippet fails.
• file: quotes are not gated at all. conf/reference_validator_config.yaml
  lists file in skip_prefixes, so file: references are exempt from snippet
  checking, and a composed one passes validation silently. Every file: quote
  this audit added was checked by hand (grep -cF against the cited file);
  the ones it inherited were swept once, with the result recorded below rather
  than left implicit.

This is a known repo-wide blind spot, documented in
MISCITATIONS.md, SPKW.md and
FUNCTION_KNOWLEDGE_GAPS.md; the SPKW audit
measured ~48% of 550 file: quotes non-verbatim while passing validation.

What the sweep of this audit's changed set found

file: supporting_text lives in two different slot classes, and the tools
disagree about which one they read. Getting this wrong is how the first pass of
this sweep measured only half the set, so the slots come first:

So neither slot is gated for file:, and the two code paths cover opposite
halves. The first version of this sweep keyed on reference_id and
supporting_text appearing in the same mapping, which is true in supported_by
and false in findings (where the id sits on the parent - id:), so it silently
skipped all 261 — which is the slot the repo-local gate actually reads.
A future file: gate has to name its slot set, or it will measure the wrong half
exactly as this did.

Re-measured across both: 1206 quotes in 104 reviews, 15 non-verbatim, of which
3 were fixed and 12 were left.

All twelve are non-verbatim, all are byte-identical on main, and none sits on a
row this audit moved — so they were reported rather than fixed, to keep the
audit's diff from becoming a cleanup of inherited quote problems. Do not read
this section as "the changed set is verbatim": it is 12/1206 short, deliberately
and with the genes named. (merA and groES carry no_change history records
for this reason; their first records say "one", which was true of the
supported_by-only sweep that wrote them and is superseded by the figures here.)

Two details that a count of distinct strings hides: merA's quote occupies two
sites — supported_by and references[].findings[], the same text plus a
period — so occurrences exceed distinct strings; and the compositions are not
confined to notes files, which is the tempting generalisation. Of the five
findings misses, two are -notes.md-sourced (retS), two are
-deep-research.md-sourced (merA) and one comes from a .tsv
(SALSP/mcr-4, candidate_new_annotations.tsv). Deep-research-sourced quotes
are usually real copies, markdown emphasis included — spot-checks on flgG and
PP_1084 are verbatim — but merA shows the habit is not source-specific.

A naive substring test over-reports, and the way it does so differs per slot —
which matters because the breakdown below is the spec a future file: gate would
be built against. Raw flags, classified, with each denominator named:

• Wrap artifacts. The quote is present, split across CC continuation
  lines — secD's Part of the essential Sec protein translocation apparatus
  straddles secD-uniprot.txt:47–48. A gate must unwrap CC/DR continuations.
  The findings slot's two are PSEPK/infC:128 (the IF-3 30S-binding
  sentence, straddling infC-uniprot.txt:29–31) and PSEPK/mraY:191 (the
  phospho-MurNAc-pentapeptide transfer, straddling mraY-uniprot.txt:32–34 —
  the quote begins at transfers, the last word of :32). Both are under a
  - id: file:…-uniprot.txt parent in references[].findings[], and both return
  grep -cF 0 against their source while matching once the CC
  continuations are unwrapped. Named because they are the evidence for the rate
  difference below, and the one row in this table grep could not otherwise
  check.

(Earlier revisions of this line cited infC:182 and mraY:192 with a
two-line span. infC:182 was wrong twice over — it is a
core_functions[].supported_by quote, not a findings one, and its shorter text
is verbatim at count 1, so it is not an artifact at all. Both locators had been
derived by grepping for the quote text separately from the pass that classified
it, which matched a different occurrence; they are now read from the files.)
But do not size that work from the supported_by rate: it is 25/34 (74%)
there and 2/7 (29%) in the findings slot, because the two slots cut quotes
differently — findings-slot quotes are short and tend to stop at a wrap
boundary rather than run across one (flgG:106 ends exactly where
flgG-uniprot.txt:45 ends, which is why it passes). The same slot asymmetry
that made the first sweep measure half the corpus also makes its artifact rate
non-transferable.
- Marked elisions. zwf joins two fragments with a literal ..., visible
to a reader. Both are in supported_by; the findings slot has none. A gate has
to decide whether that is a supported convention.
- The slot set, per the table above: supported_by[] and
references[].findings[]. The two existing code paths each read one, so a gate
inheriting either one's scope measures half the corpus.
- Test the passes, not only the flags. Every round of PR #3165 audited the
quotes this sweep reported; a classifier checked only on its output can be
wrong in the direction nobody looks. Round 14 sampled two quotes it had
cleared (NICAT/NaPMT1.1, neither previously named) and both are verbatim at
count 1. A gate needs that check in its own tests, or a silent false-negative
rate stays invisible.

The difference between a marked and a silent elision is why only some of these
read as quotes at all: zwf shows the join; merA performs the same operation
invisibly, fusing two clauses three sentences apart and changing when to
When; groES splices across two source sentences. Silent splices are the
class that needs the gate; marked elisions are a convention question.

Two independent measurements of the same blind spot now exist — ~48% of 550
(SPKW) and 12/1206 ≈ 1.0% (here, both slots) — which differ by corpus, not by
method. Every rate in this section carries its denominator in the same sentence,
deliberately: three findings running on PR #3165 were "the measurement is sound,
the sentence around it is wider than the measurement" (166-vs-165, 7/945,
945-of-1206), and naming the scope at the point of use is the one habit that
would have caught all three.

One more thing a gate would have to settle: file: references use three
different base conventions, and under reference_base_dir: genes only one
resolves.

The second is what the PAINT citations here and dozens of existing reviews use;
the third appears in 50 reviews repo-wide (SALSP/mcr-4 uses it at ten sites,
alongside file:projects/… paths). So a gate cannot pick one base and treat the
others as the error case — it has to recognise all three. Nothing breaks today
only because file is skipped entirely, which means two of the three forms have
never been resolvable by any tool, only by a human with grep. This is
pre-existing convention drift, not something this audit introduced.

So: when reading a rationale here, treat a file: quote as a reviewer's
transcription rather than a gated fact.

What the counts count

192, 117 and 75 are counts of recorded rows, and the gene figures need
the same care, because one protein in this audit has two records. PSEPK/aroQ
and PSEPK/aroQ-III were two reviews of the same protein (Q88IJ6), which this
audit merged; the record for the merged-away folder stays (records are
append-only) and declares merged_into, so its single GO:0019631 row is
recorded twice but was adjudicated once. The live state is therefore:

summarize.py derives all four from merged_into rather than from a hand
adjustment, and check_outcomes.py checks a merged entry's action against the
surviving twin instead of skipping it. The transition table and the
117 stand / 75 relaxed split are row counts, so they include the aroQ pair
once each.

A gene entry's gene label mirrors its folder name under genes/<ORG>/,
accession suffix included where the folder carries one (PSEPK/dapF__Q88CF3,
PSEPK/dapA__Q88NH2). The suffix is the disambiguator the folder layout
already uses; two of these labels would otherwise collide.

summary.tsv is generated from the batch files by summarize.py.