v0.37.1 — an absence is deleted before it is rewritten

An absence is deleted before it is rewritten

v0.37.0 named the absence residue and told the writer to say what the
object does instead. Applied to the author's two manuscripts, that action
had a failure mode of its own: a positive rewrite that only restates what
the neighbouring sentences already carry (downstream of detection reached
five occurrences in one paper; a statistic "reads position in catalog
coordinates" in an appendix whose method section had said so). The ruling
now in the rule's action, the paper skill's self-check, the de-ai skill's
residue bullet, and the standard's L4 bullet: judge the sentence in context
and delete the clause first; rewrite only what the page does not yet say;
keep a physical fact, a scope limit, or a conceded limitation as it stands.
The second sweep over the two manuscripts cut 17 such clauses or sentences
that the first sweep had rewritten, and the Pass-3 gate of rewrite_reward
then did what §6 says it does: the dropped negation marker is a missing
invariant, reported on 4 of 7 changed Letter paragraphs and
38 of 54 pipeline-paper paragraphs, every one the absence clause
itself or its companion (\ref, because, a math token inside the deleted
clause), and one tokenizer artifact (CDM- read as an acronym). The action
text says so, and the disposition is recorded beside the finding rather than
the negation restored (§23.4a). No detector changed; one test reads the new
action. 523 tests in 27 files.

v0.37.0 — a sentence that says what its object never does is a residue

A sentence that says what its object never does is a residue

The negative-label rule of v0.36.0 read headings and captions; the same
defect in body prose had no detector and no name in the standard. The
author's name for it is a menu line reading "tomato and egg (no braised
pork)": the head never participates in the detection decision, the
reference stratum carries no quoted number, no support threshold is applied, because …. Each tells the reader what the thing is not, in the
place where the thing should stand, and each is what a revision leaves when
an ingredient was taken out and its absence written in. residue-absence
(deai_residue.py, rule 6) reads body sentences for the families, exempts a
sentence that cites (a contrast with published work is a baseline, not a
tombstone), and skips a hyphenated compound (never-touched controls is a
name). Strengths were tiered on the held-out refereed full texts, 442 files
and 1,899,092 prose words: never and the nothing is / none sees / no … is applied forms occur 0.008 times per 1,000 words there (15 never
sentences in 13 files, all physics, bounds or procedure) and are strong;
carries no, is not applied and does not participate occur 0.02–0.05 per
1,000 words, in 6–15% of files, mostly as procedure, and are ordinary; is not a and with no were measured and left out, being the hedges and
definitions-by-contrast §6 protects (§23.4a). On the author's pipeline paper
before its sweep the rule found 15 strong and 17 ordinary sentences, never
at 118 times the refereed rate; after it, 0 strong and 2 ordinary physics
statements kept under a disposition. The families live once in the tool,
mirrored between absence-family markers in skills/paper/SKILL.md beside
the history families, and validator_check now proves both mirrors. The
standard's L4 voice bullet says what the rule enforces: a sentence says what
an object does, not what it never does; the paper skill's forward-narrative
self-check gains the item, with the physics-fact and scope-limit exemptions
stated beside it. Seven tests; 523 in 27 files.

v0.36.2

An external review of v0.36.0/v0.36.1 (Codex gpt-6-astra at maximum
reasoning), every finding verified first-party before it was acted on; the
record is docs/audits/codex-review-2026-09-04.md.
Fifteen defects, six root causes, and one measurement the fixes reopened.

The manuscript side never let a subsection inherit its section

deai_metrics carried a second section parser beside extract_sections: no
whitespace before the brace, no optional argument, and a subsection whose title
matched no bucket fell to unknown and was silently unmeasured — 344 of the
540 paragraphs of one held-out Planck paper — while the corpus side has
inherited since v0.28.0. section_units now derives from
extract_sections.RE_SECTION with the same inheritance; a document with no
heading is one unit, the preamble is none. Measured on the same 203 held-out
papers in one process: salience 2,753 → 4,473 findings, 0.2776 → 0.4542
per scorable paragraph against the 0.2710 design rate (§17.5 records the
excess as open; the reference banks were bucketed the same way); cohesion
10.87% → 10.81% and hedging in intro 7.89% → 7.80%, unchanged; the mentor
population's salience 0.2781 → 0.3984.

One paragraph sweep, one line-preserving blanker

Four tools projected the manuscript line by line where the corpus side
projects a passage in one pass. extract_sections.blank_preserving is the one
blanker (a match becomes same-length spaces, newlines kept) and
deai_reference.paragraphs the one paragraph sweep; structure, oracle, voice,
register, residue, metrics and the labeller read them, so a heading, a float
or a citation split across lines is blanked on both sides.

Rates per body prose word, and no AUC from two documents

eval_findings divided by text.split() over raw source — bibliography and
preamble included, which a single-file machine draft has none of — and an
axis whose novelty was NaN on all but two documents still reported an AUC.
The denominator is prose_words(body_only(text)), placeholders dropped; fewer
than 20 scorable documents is unmeasured. The same findings re-read:
register 57 findings 0.0247 → 0.0371 per 1,000 (AUC 0.392 → 0.391),
zero-hit 2.212 → 3.374 (AUC 0.246 → 0.174), collocation AUC 0.691 →
0.704, salience AUC 0.774 → 0.572 (with the bucketing fix above).

Assembly is one module, and the gate counts prose

tex_assembly.py: an \input mid-line keeps the words around it, a second
call splices again (only a child in the current include stack is a cycle), and
a --git-ref baseline is the document assembled at the ref — length_gate
had compared an assembled draft against a bare root. The gate counts prose
words (no [math]/[FIGURE-OR-TABLE] placeholders) and reads
--require-shrink as a percentage, a fraction or a word count, rejecting
100%, 0%, inf and 30%% and rounding a fraction up to one word.

The removal map counts each unit once

A sentence that was both a restatement and zero-gain, or that sat inside a
removed paragraph, was budgeted twice; a canonical home may no longer drop a
negation, a comparative or a number its copy carries; an opener is a
whole-sentence removal only when at most three of its own content words would
go. Held-out default target median 3.08% → 1.59% of prose (§23.5).

Residue: marks across lines, and the procedural we have added

Edit-meta marks are scanned over the visible body as one text, so a phrase
wrapped at a line break is one mark; once visible, we have added fired
fourteen times on refereed prose, every one a procedure (we have added uniform Gaussian noise) or another paper's history, so it is a mark only with
a document object. The diff rule compares whole negated objects and skips a
negation the old version already carried. Strong residue 34 → 27 of 203
papers (16.7% → 13.3%; §23.4).

Register, collocation, labelling, validator

Citations are blanked across lines and macro definitions matched at line
start; a defined term's scope is its sentence. Collocation pairs break at /,
. and digits, the weights are named for what they are (expected_copresent_passages,
p_copresence_absent), scope and calibration unit are sentence, and the
bank is rebuilt (541,309 → 530,677 pairs). The labeller's control sample
excludes any passage a finding touches by line span. validate_plugin gains an
eleventh check: no tracked file past the 750-line budget. Standard and skill
text corrected where they had drifted from the code (zero-hit exemptions,
the paper-agent example, the modifier-stack definition, the restatement
rule). 491 tests in 24 files; §23, §17.5, §19.4, §21 and both README
limitation tables re-taken.

v0.36.1 — floats are the fourth line seam

The first run of v0.36.0 on a manuscript with AASTeX tables: three tool
defects, none of them in the manuscript.

The fourth line seam: floats

deai_register projected the manuscript one line at a time, so a
\begin{table*}…\end{table*} spanning lines was never blanked the way the
corpus side blanks it in one pass over a passage. A tabular* column
specification (@{\extracolsep{\fill}}lll), \tabletypesize{\scriptsize},
\tablenotetext{a}{Tied…} and every caption's words counted on the manuscript
side only: 23 of 90 zero-hit terms on one paper were filllll, tabcolsep,
aTied, crimson, isosurfaces. body_only now blanks floats and
length/table-note commands across lines beside the math spans of §23.1; the
shared float pattern names deluxetable*, longtable and a bare tabular*
as tables on both sides; a token with fewer than three letters (a--c, a
panel range) is not a word. Re-measured on the same populations: held-out
register-zero 2.658 → 2.212 per 1,000 words, still on 100% of refereed
papers, rank AUC 0.221 → 0.246; the thresholded rule 81 → 57 findings,
0.0351 → 0.0247 per 1,000, 30.0% → 22.2% of documents, AUC 0.352 → 0.392,
own-membership 98.2% of 57. Salience and collocation lose the passages that
were table cells (held-out 1.2025 → 1.1925 and 2.044 → 2.031 per 1,000, AUCs
0.770 → 0.774 and 0.688 → 0.691); every machine row and the Letter figure
§23.1 quotes (14 words, 11 strong) are unchanged.

A negated object stops at the sentence end

residue-negative-label captured "a mass reconstruction. The solid spheres
mark…" as one object and reported solid absent from the body. The capture
now ends at ., ! or ? as well as at a clause mark. Two regression tests
(465 in 23 files); manifests and headers at 0.36.1; §23.1 table and both
README limitation rows re-taken.

v0.36.0 — vocabulary the field never wrote, the removal map, and residue

Five tracks from one plan, each answering a specific complaint about what the
plugin could not see: words the field never wrote, sentences a mentor marked
as "jargon, what does this mean", a condense pass that scraped a few percent
by hand, and the trace an edit leaves behind. Three new tools, three new
structure families, one new exit contract, and the standard at v3.8 inside its
750-line budget. Evidence: §23.

The zero-hit audit, and a third projection asymmetry

deai_register kept a 15-use floor so that it would not flag a refereed
paper more often than not. The owner's rule replaces the knob with an
exhaustive question: which body words does the manuscript use that no
passage of the field's corpus carries? Every one is listed under
register-zero:<term>, strong unless it is a mechanical formation of an
attested stem, and the only other exemptions are the author's — the paper
defines the word, or cites the method it names. Measured on 203 held-out
refereed papers against 173 machine documents it is not a detector (2.66 words
per 1,000 on every refereed paper, rank AUC 0.221) and ships as advice.

Building it exposed the third instance of the projection asymmetry recorded
in §17.4 and §18.1: section headings sat in the manuscript's body projection
but never in a corpus passage, so Validation fused with the sentence under
it and read as a word the field never wrote. extract_sections.RE_HEADING_COMMAND
is now the one owner of the heading pattern and deai_reference.units,
deai_register.body_only and length_gate consume it. The thresholded rule
moved with it on the same 203 papers: 196 findings → 81, 0.0858 → 0.0351 per
1,000 words.

L2.collocation: words the field never joins

physical cells is two ordinary words and a pair no passage of 41,710 has
written. tools/deai_collocation.py judges each sentence by the fraction of
its distinct adjacent common-word pairs the bank does not attest, against a
leave-one-out reference per bucket at sentence unit — at calibration a pair
seen in exactly one passage is that passage's own. Only common words are
judged as partners (11,286 of them), pairs break at punctuation, placeholders
and dashes, and each flagged pair carries its expected co-occurrence and
e^−λ. On the private Letter it flags five of the six mentor-marked phrases
still present; on held-out papers the document novel-pair fraction separates
machine text at AUC 0.688.

Three structure families from the mentor's margin

Paper-as-agent subjects ("This paper presents"), wh-cleft openers ("What
matters is") and modifier stacks (three-plus tokens before a head noun with
two hyphenated compounds) join deai_structure's auxiliary class: named on
the sentence, never in template_score, with per-bucket human fractions in
the recalibrated baseline (stacks 2.5–15.9%, paper-agent 0.13–1.00%, wh-cleft
0.00–0.23%).

Condensation that is measured against a map, not a feeling

tools/condense_map.py enumerates every removable entry — restatements with
their canonical home, zero-gain sentences, dead figures/tables/labels/macros/
acronyms, verbose constructions, repeated glosses, duplicated paragraphs —
with the words each frees, and totals a default target. length_gate.py --require-shrink turns that target into an exit code (length-shrink-short,
strong, exit 1). Held-out refereed papers carry a median default target of
3.1% of prose; the old condense skill removed less than that by reading. The
skill is rewritten around the map: one disposition per entry, closed by the
gate.

L4.residue: the trace an edit leaves

tools/deai_residue.py: first-person drafting history (we initially, no longer), edit-meta text (TODO, see previous version), a heading or
caption whose object the body never names, and with --before/--git-ref
the label an edit added and does not earn. Exit 1 on a strong finding is the
third narrow exit contract, so its strengths were set on 203 held-out refereed
papers rather than assumed: the first families put a strong finding in 154 of
them; reading body prose only (a \newcommand{\TODO} in a preamble and
"Planck Collaboration XXX" in a bibliography are not residue), dropping used to (174 instrumental hits) and demoting initially/originally/at first
bring that to 34, and the static negative-label rule is ordinary while the diff
rule gates. The history families live once in the tool,
mirrored between markers in skills/paper/SKILL.md, and validate_plugin's
tenth check calls the tool's own validator_check to prove the mirror and to
scan shipped documentation for the edit-meta literals.

Everything that had to move with it

ai_ism_lint runs both new axes by default (--no-collocation,
--no-residue); eval_findings and label_findings cover collocation; the
examples/ table gains a collocation-novel row (3 → 6, and why); the
standard is v3.8 at 749 lines. 38 tools, 463 tests in 23 files, 10 validator
checks. The README latency table is re-taken whole: the model-free row rose
from 458 ms to 1.07 s, most of it the 541,309-pair bank load, which
--no-collocation drops. v0.30.0–v0.31.0 moved to CHANGELOG-ARCHIVE-RECENT.md;
this file had reached 790 lines.

v0.35.1 — the record audited against what it describes

A post-release audit of the documentation, not of the code. An escape the
writer lost had truncated the tools table in both READMEs, the one published
figure with no artifact behind it was wrong, and several figures and dates in
the record had drifted from what they describe.

A lost backslash ended the tools table four rows early

The tools/tex_macros.py row wrote \newcommand without its backslash, and
the consumed newline split the row, so the table stopped rendering there —
retrieve_exemplars, fetch_arxiv_abstracts, train_ai_ism_classifier and
extract_md_negatives fell out of it, in both READMEs. The same damage sat in
the v0.32.0 changelog entry at \nocite.

The worked example published a count the linter does not report

examples/README.md said 19 total advisories where ai_ism_lint reports 18.
It was the only published-figure document with no artifact behind it, so
nothing read it. tests/test_published_figures.py now renders both of its
tables — the before/after summary and the per-rule counts — by running the
linter on the two shipped manuscripts and looking for the result, the same way
every other pinned figure is rendered from the artifact it was read from. A
cell the linter does not produce fails the case, so a document that agrees with
a stale run and a document nobody updated fail identically.

Figures and dates the record had drifted from

• 203, not 200, held-out refereed papers, in the two DISPOSITIONS.md rows
  that still carried the count from before §18 re-measured it.
• The v0.30.0–v0.33.0 changelog headings were dated 2026-08-27; their tagger
  dates are 2026-08-26.
• The release-gate label in EVALUATION.md §12 stamped a version on a suite
  size measured after it. It now reads "as of 2026-09-03; last tagged release
  v0.35.1", so the version names the tag and the date names the measurement.
• Both README latency tables paired 394 tests with 81.4 s, a wall time taken on
  the 393-test suite. The suite row is re-taken: 73.0 s, median of 3, the
  three runs spanning 70.0–86.9 s. Every other row still stands from the
  2026-08-27 take and the preamble now says which rows carry which date.
• docs/README.md called the evidence record five files where it is eight,
  gave three of them section maps missing the sections added since, counted
  four kinds of document where six live there, and described the validator's
  version and suite-size checks narrower than they are.
• The unit-pattern anchor in both READMEs pointed at rewrite_reward.py:41;
  the pattern is at :56.
• Both READMEs said CI runs on every push. ci.yml filters branches: [main],
  so it runs on every push to main and every pull request.

Also

• .gitignore covers models/hf-cache/, .ce/ and .ccm/ — relocated model
  cache and machine-local tool state, never repository content.
• Suite: 394 tests across 20 files; validate_plugin.py 9/9.

v0.35.0 — a worked example that is safe to ship, and one that was not

The axis that counts numbers could not see numbers written as macros, a public
repository was carrying an unpublished manuscript's method summary, and the
worked example that demonstrates any of this now runs on a synthetic paper
instead of a real one.

Numbers held in macros were invisible, in both directions

A manuscript that writes \newcommand{\Nfields}{63} in its preamble and
\Nfields{} in its results put a measured quantity where neither named text
projection could read it. RE_TEX_SIMPLE_CMD reduces a command to its
argument, so the use site contributed nothing while the definition site
contributed the digits once, in the preamble, attributed to no reported
section. Two errors running in opposite directions, which is why the net stayed
small enough to go unnoticed: on the manuscript that surfaced it, expanding the
uses adds 650 digits and dropping the definitions removes 493.

Found by running the tools on a real manuscript, not by a test. The salience
axis had been reporting recital percentiles for that paper computed on 91% of
its digits, and correcting it moved that paper from 16 recital findings to 26.

tools/tex_macros.py expands numeric-literal macros once, on the assembled
document root, because that is the only scope holding both a preamble
definition and a body use — the same reason read_tex_document folds
\include in the first place. Only a bare numeric literal expands, so
\newcommand{\Msun}{M_\odot} and every macro taking an argument are untouched.

Both salience baselines were rebuilt rather than left to score expanded
manuscripts against an unexpanded reference. They move by zero to four
decimals at the p90 and p95 gates, across all three features and all seven
buckets, in both wgl and wgl-letter: 88.2% of 390 corpus documents never
use the construction. That is what separates a correction from a rescaling, and
it means no published salience figure in EVALUATION.md changes (§22).

An unpublished manuscript's method summary was in a public repository

This repository is public. Content from the author's unpublished manuscripts
had accumulated in the evaluation record since 2026-07-12, in increments that
each looked locally harmless — a codename in a test fixture, a macro and its
value in a docstring, one fidelity-preserving rewrite quoted in full.

The aggregate was a complete method summary: enough of the method for a competitor to reconstruct the contribution. Public for 46 days.

The quoted rewrite was the largest single exposure, and it was labelled a
proposal rather than a quotation, which is exactly why it survived review: the
rewrite gate protects numbers, citations, macros, entities and claim/evidence
relations, so a fidelity-preserving rewrite discloses what the original
discloses. Being a rewrite is not a defence.

Removed: the quoted block, the method-component list, manuscript codenames
throughout, one real measured value, and manuscript-derived test fixtures,
which now use invented macros. Kept: every measurement. Finding counts,
percentiles, rule names, before/after tables and the panel results describe how
the tools behave, not what the paper found, and they are why the record exists.

Git history was rewritten to remove the same content from every earlier commit.
That does not undo publication — the content may persist in existing clones,
forks and upstream caches.

A worked example that is safe to ship

examples/ now carries a synthetic manuscript and the same paper after acting
on the findings. The topic is a textbook one and every value is invented, so
the demonstration no longer depends on anyone's unpublished work.

It is also a more honest demonstration than a clean sweep would be. L0 targets
go to zero and discourse-cohesion from 3 findings to 1, while
salience-recital rises from 4 to 6 — because carrying a noun forward to
link two sentences pulls the subject into sentences that also carry a numeral.
In a number-dense passage cohesion and recital want opposite things and no
rewrite satisfies both. Both findings are true, and which to act on is the
author's judgement, which is why neither is a blocker and there is no score.

Also

• The re-export contract test named its excluded module imports one at a time
  (re, defaultdict, Path), so it failed the moment a new import appeared.
  It now excludes module objects by type.
• Suite: 393 tests across 20 files; validate_plugin.py 9/9.

v0.34.0 — review skills split into measurement primitives and composites

The review skills separate into measurement primitives and the composites that
call them, the plugin stops depending on a skill it does not ship, a recorded
count is now checked wherever it is written rather than only where someone
thought to look, the labelling harness gains the two axes it was missing, and
building a population to point it at turned up two silent corpus bugs.

The one skill the plugin orchestrated but did not ship

final-review launched modern-physics-review as a worktree agent and
paper-review re-read it after its own disposition-complete state, but the file
lived in the user's global skill directory. On any other machine both skills
reached for something that was not there. Its content is now in the repository —
as skills/physics/SKILL.md and a paper-review mode, for the reasons in the next
section — and no skill references anything the plugin does not ship.

Its content could not be taken verbatim: that would have put a second consequence
vocabulary in the repository. The grading was 🔴/🟡/🟢 and the termination
condition read "0 个 🔴 / 0 个 🟡 … 必须收敛到 0 才算完" — a universal
zero-advisory gate, which is the thing validate_plugin.py's
STALE_REVIEW_MARKERS exists to refuse. That table matches English strings
(zero-issue convergence, 0 issue across all), so a Chinese spelling of the
same rule walked straight past it. Grading maps onto sci-paper.feedback.v1
instead — 🔴 to integrity_blocker, 🟡 to a strong advisory that must carry a
disposition, 🟢 to an ordinary advisory — and termination is a stable
disposition-complete state, not an empty report. The physics itself is unchanged.

M9 was deleted rather than translated. It ran a hard-coded keyword grep, and
its em-dash pattern was byte-identical to the one at tools/ai_ism_lint.py:30,
which already files those hits as kind="l0_target". Its boilerplate word list
asks the question tools/deai_register.py answers from the field's own document
frequency, and skills/paper-review/SKILL.md already carried an override saying
corpus evidence beats an upstream word list. Keeping M9 would have given one
question two contradicting answers. physics delegates to those axes instead.

calibrate joins NORMATIVE_SKILLS along with the new primitives:
docs/architecture/RESPONSIBILITIES.md claimed every skill but brainstorm was
enforced there, and calibrate was not — it referenced the standard anyway, so
the claim was true in fact and unenforced in code.

A count was only checked where someone thought to look

check_registry_counts opened README.md and matched ## Skills (N). The
repository's shape numbers are written in four other places — the headline line,
the ASCII tree, the release paragraph and the latency table — and in two
languages. Nothing read the translation at all, so README.zh-CN.md advertised
9 skills, 25 product tools, 15 test files and 252 tests to a repository of
12, 34, 19 and 381. The English latency table carried a stale figure of its own,
360 passing against a 381-test suite, written in a spelling
check_recorded_test_counts had no pattern for, inside a file it never opened.

A fifth regex would have been the fourth patch for one defect. SHAPE_CLAIMS
instead lists every spelling either README uses, each paired with the quantity it
claims, and both files are scanned — a shape number is checked wherever it is
written, in whichever language. check_recorded_test_counts reads the two
READMEs alongside the evidence record for the same reason. The check tuple moved
to module level as CHECKS, so the published "9 contract checks" is verified
against the tuple that actually runs instead of against a hand count. Twelve
claims across both READMEs are verified per run, and the number of checks is
still nine — which is what keeps that particular claim true.

Both stale sets are corrected. Each fix was confirmed the same way: the old value
was left in place first, to watch the validator name the file and the quantity.

The latency table, re-taken whole

The interpreter floor moved 84 → 98 ms, so under the table's own rule no v0.33.0
row compares and every row was re-measured together on the same 5,084-word
assembled paper: L0 274 → 317 ms, all model-free axes 409 → 458 ms, the length
gate 194 → 258 ms, the validator 2.3 → 2.9 s, the suite 51.0 → 60.1 s — the last
one now covering 381 tests where the published figure covered 360, which is the
other half of why a partial update was refused.

The two model-backed rows went the other way: --oracle 48.7 → 33.8 s and
--voice 59.1 → 37.2 s, against every standard-library row rising with the
floor. The table cannot explain that and does not try. Their cost relative to a
full model-free pass falls from 120×–145× to 74×–81×, and the stamp's existing
instruction stands — read those two rows as one machine on one day, not a trend.

The physics checklist existed three times

paper-review dimension K listed eight modern-physics checks; modern-physics-review
listed the same eight as M1-M8; and paper-review also launched that skill as an
isolated agent after reaching its own disposition-complete state. Dimension B carried a
fourth partial copy — conservation, symmetry, parity and asymptotics again. Dimensions
E, C and G had the same shape against the narrative-spine protocol, claim-evidence
discipline and figure inspection.

The rule that was missing is now written down beside the dimension list: a dimension
either calls a measurement primitive or owns its content — never restates one. Three
primitives were extracted to give the calls somewhere to land:

• skills/physics/SKILL.md — P1-P8, the single copy of the first-principles checklist.
• skills/mainline/SKILL.md — the purpose record, contribution graph and cold reader's
  seven questions, out of dimension E.
• skills/logic/SKILL.md — claim graph, empirical statistics and the review side of
  claim-evidence discipline, out of dimension C.

Each is measurement-only: it emits findings and never edits. Repair stays with the action
primitives, /sci-paper:de-ai Pass 3 and /sci-paper:condense, which is what dimension I
already did and what the other dimensions now do too.

modern-physics-review therefore never becomes a skill of its own. Its checklist is
physics; its fix loop and report were already paper-review §4 and §6; and the part worth
keeping — an isolated agent that cannot inherit this review's "already verified" conclusions
— is now a mode of paper-review rather than a second skill. --no-isolated-mpr named one
skill, and three are orchestrator-owned now, so the flag is --orchestrated;
--skip-final-mpr is --skip-final-physics.

final-review launches the three primitives at parent level instead of the single MPR
agent. That is not extra process: it is what makes their cold read independent of
paper-review's context instead of nested inside it.

What did not change: paper-review dimension D still runs /sci-paper:de-ai --audit-only every round, and that is correct under this model — a composite calling a
primitive's measurement half is composition, not nesting. The defect was never the call.

One skill table instead of two, neither of which was complete

README.md listed the skills twice — as "the eight functions" and again under
## Skills — and both lists predated calibrate, so the setup step every
measured axis depends on appeared in neither. They are now one table of ten
ordered by stage (setup, write, revise, review, explore), with the existing
per-skill descriptions carried over unchanged; the ## Skills section points at
it. The same consolidation is in README.zh-CN.md.

The labelling harness covers every axis that emits findings

tools/label_findings.py sampled two axes while four emit findings. It now
samples all four — L0.register, L2.salience_hierarchy, L2.cohesion,
L2.hedging. The discourse entry forced the emitter table's shape:
discourse_findings returns both of its axes from one call, because they are
measured over different spans, so an emitter now declares the axes it owns plus
how to tell which one a finding belongs to. deai_discourse.axis_name() gives
the L2.<feature> spelling one owner, so the sheet cannot file findings under a
name no report prints.

Recall is now reported pooled over the axes, and only pooled. A control row
asks whether a passage should have been flagged; a labeller who answers yes has
not said by which axis. Dividing every axis by the same miss count — what this
did while it carried two — charges each axis with every other axis's misses, and
does so four times over at four axes.

Populations are named rather than fixed: --population NAME=DIR, repeatable,
each sampled and scored separately. --drafts DIR remains as shorthand for
--population draft=DIR. Whichever population is named must sit outside the
calibration banks, and the size of that effect is now measured on this corpus:
the register axis yields 1 finding across three in-calibration papers and
8 across fifteen of the same author's papers held out.

Two silent bugs in the arXiv corpus builder

All ten AUTHOR_QUERIES were dead. Surname_Initial is the arXiv listing
URL's format (/a/hoekstra_h_1), not the search API's au: syntax, so every one
returned exactly 0 records — the broad query set was silently ten queries short
while its comment claimed those records broadened the curated-field class. A dead
query leaves no trace in the output, so AuthorQueryFormTest now guards the form.
Re-run with the quoted form, the ten contributed 18, 37, 6, 24, 31, 21, 4, 2, 18
and 1 new records: 13,642 → 13,804.

And a surname is not a person, which took two goes to get right. The
corrected queries are scoped to cat:astro-ph* because unsc...

v0.33.0 — discourse axes ship; citation placement refuted

The roadmap's last engineering item ships and its strongest measurement is
refuted. Both outcomes came from the same discipline: measure the thing against
a population it was not built on, and believe the result.

Two discourse axes, at two different units

tools/deai_discourse.py adds L2.cohesion (given/new linkage — how much of a
sentence's content vocabulary the sentence before it already used) and
L2.hedging (epistemic markers per 1,000 words). Both fire below the field's
tenth percentile, the opposite direction from every existing axis, because here
the defect is absence rather than excess. Roadmap rank 6, Deferred since
v0.26.1.

They measure at different units, and that is forced rather than chosen. Hedging
has no paragraph-scale lower tail at all: on the 27,917-paragraph wgl bank
its tenth percentile is exactly 0.000 in every one of the seven section buckets,
because a 40-word paragraph that hedges nowhere is entirely ordinary. A gate
there is one no passage can fall below, and the axis would have reported a
confident zero findings forever. Regrouped so one section is one unit, six of
seven buckets separate (p10 from 1.055 in data to 3.350 in discussion);
abstract stays flat and abstains, since an abstract is one passage. So
cohesion calibrates and detects per paragraph, hedging per section, and each
artifact records its own unit so the two can never be read against each other.

Hedging ships restricted to intro, and two independent measurements put the
restriction in the same place. Held-out transfer at the p10 gate on 203 refereed
papers: 7.89% in intro against 15.48%–26.77% everywhere else. Worst-of-six
generation regimes: 0.613 in intro against a 0.460 human-vs-human null, while
conclusion runs below chance at 0.376. Cohesion needs no restriction — it
transfers at 10.87% overall against a 10% design point, in every bucket, and
separates all six regimes in intro at worst-of-six 0.676 (null 0.515).

Both floors were measured, not picked. The cohesion sentence floor is 3 rather
than 4 because at 4 the 20-document ai tier offers 15 measurable introduction
paragraphs in total and at 3 it offers 62, with the separation unchanged (0.676
against 0.674). The hedging word floor is 150 because it is the first floor at
which every non-abstract bucket resolves, and 250 buys nothing while costing the
abstract bucket 86% of its sections.

Evidence: EVALUATION §19.

Citation placement: refuted on its own pre-registered condition

v0.32.0 recorded citation placement as the strongest model-free discriminator in
the whole record — section-matched rank AUC 0.866 in method, surviving the
section, length and human-vs-human controls — and declined to ship it for one
stated reason: all 173 machine documents came from a single generation process,
so one bank could not separate "AI cites more" from "these prompts made it cite
more". The condition for shipping was a second, independently produced bank.

Two banks of 20 documents each, same 20 topics, a different model (Codex
gpt-5.6-terra), differing in one line of prompt. With no citation instruction
the same statistic scores 0.053 — not the absence of separation, but
separation of nearly equal strength in the opposite direction. With one it
scores 0.734. Density swings 12.5× (1.00 to 12.55 citations per 1,000 words)
around a human median of 6.20, so the two machine extremes bracket the human
distribution instead of sitting on one side of it.

The signal is real and it is not about authorship: it is about which model,
prompted how. Refuted, and the bar it sets for anything after it — hold your sign
across independently produced banks — is stronger than the three controls
v0.32.0 applied. EVALUATION §20.

One (feature, unit) reference for every per-bucket axis

tools/deai_reference.py (roadmap rank 2, Deferred (elegance debt)) now owns
the quantile grid, the plateau-top percentile reader, the 30-unit sample floor,
the paragraph and section sweeps, the artifact loader — which had five copies
across the suite — and the calibration loop. It holds no policy.

It stopped being elegance debt the moment a second per-bucket axis existed, and
earned its keep immediately: its one invariant is that calibration and detection
share a unit and a grid, and that check is what caught the hedging axis
calibrating where no lower tail exists. cli_common.axis_main absorbs the
--calibrate-or-read-one-file CLI that three tools carried near-byte-equivalent
copies of, and deai_feedback.reference_block the reference= payload six
detectors spelled out separately.

Also

• deai_discourse is wired into ai_ism_lint behind --discourse (default on),
  and reports one axis status per feature, never a joint one — a field can
  support cohesion and not hedging, and does.
• Both README demos are now pinned to the axis set that produced them
  (--no-discourse); demo 1 arm C would otherwise draw a second advisory it was
  never scored with.
• tests/test_published_figures.py pins the two new references, including each
  bucket's p10 gate, and was negative-tested 4/4.
• tests/_profilefixture.py gives the axis tests one throwaway-profile builder;
  the two copies had already drifted on whether a record carries a source,
  which decides whether a section-unit reference can be built at all.
• 34 tools, 360 tests (19 files). Dead argparse/json/deai_metrics imports
  removed from four tools after the CLI extraction.
• The latency table is re-taken whole, not row by row: the interpreter floor
  alone moved 56 → 84 ms since v0.32.0, so no row from that table was comparable.
  The validator row was the worst of them, published at 359 ms against a measured
  2.3 s. A complete model-free pass is now 409 ms with the two new axes on
  (~325 ms above the floor). The first re-run was discarded — it read a 295 ms
  interpreter floor because it was competing with the session that started it.

v0.32.0 — projection symmetry, and an operating point derived rather than estimated

Asked what was still open, the audit found two more places where the calibration
side and the detection side were not reading the same text — and one of them let
the held-out evaluation set be collected as calibration input.

A four-name allowlist against 46 citation commands

RE_TEX_CITE matched cite|citep|citet|citealt and required the brace to
follow the command name directly. The corpus carries 46 distinct cite-command
names over 75,566 uses, and an unmatched one falls through to
RE_TEX_SIMPLE_CMD, whose job is to substitute a command's argument as text.
So the bibliography key became a word: \citep[e.g.][]{Smith2020} → Smith2020.

8,835 leaking occurrences across 565 of 1,490 .tex files — 8,100
optional-argument forms, plus \citealp, \citeyear, \citeauthor,
\citenum, \citeyearpar.

Matched now by shape, not by name, in three behaviours, because three exist:
a citation renders a mark, a declaration ( ocite, \setcitestyle) renders
nothing, and \citetext wraps prose the author wrote. A name allowlist is the
wrong instrument — the tail is per-paper local macros (\citeg, \citejap,
\putcite, one paper each). Verified exhaustively: 46 names × 5 written forms
× both projections.

Seven per cent of the digits the salience axis read were citation years

A bibliography key ends in a year, and latex_to_numeral_text keeps numerals so
deai_salience can measure how a passage distributes its quantities. Over the
203 held-out papers, digits 396,814 → 369,056 (−27,758, −7.00%). This is a
correctness fix to a shipped, measured axis, larger in relative terms than
the register effect that led to it. Control re-run, not assumed: 0 of 2,759
salience findings on those papers start on a bibliography line.

The held-out set could be collected as calibration input

corpus_documents walks with rglob and filtered nothing, and both
deai_anchoring --calibrate and deai_docstructure --calibrate take a
--corpus-dir that the documentation points at the field root:

--corpus-dir	before	after
style-corpus/wgl	717	517
style-corpus/wgl/fulltext-arxiv	500	500
style-corpus/wgl/fulltext-heldout	200	200

The shipped baseline was 517 — built before the held-out set existed. Re-running
the documented command today would have absorbed all 200 evaluation papers, with
no error and nothing in the output but a count nobody was checking. The existing
held-out test pinned extract_style's source tuple, which this path never
consults; the guard now lives in the collector, so every caller inherits it.

The register operating point, derived rather than estimated

MIN_MANUSCRIPT_USES = 5 was an estimate. Swept 5 → 50 against the 203 held-out
papers and 173 machine documents, rank AUC stays below 0.5 at every setting:
the axis fires more on refereed prose than on machine prose everywhere on the
curve, and tightening silences the machine side faster. No setting makes it a
detector, so the roadmap item is answered by refuting its premise. Cut at 15,
the first point where a referee-grade paper is not flagged more often than not.

	v0.31.0	v0.32.0
held-out register findings	887	198
per 1,000 words	0.3842	0.0858
documents flagged	87.19%	44.83%
rank AUC vs machine text	0.1479	0.2856
paired leakage suppressed	86.25% of 887	94.44% of 198
salience gates		unchanged

machine:ALL register is byte-identical across the citation fix (63 findings,
0.0917, 0.2428): those documents carry no LaTeX citations, so the fix cannot
touch them. The sweep and the evaluator are separate programs and agree to four
decimals on the flag rate at the chosen point.

A format variant is not a domain

wgl-letter reported degraded since v0.30.1. Enlarging it is impractical
(ApJL is 24 papers in 5,364 arXiv records) and wrong: register measures domain
vocabulary and a letter is a format. On 36 letter-format documents the
706-passage letter bank produced 262 findings the field bank did not — sne,
bao, pantheon, posteriors, and letter itself — against 2 the other
way. A <field>-<variant> profile now judges against <field> and names the
borrowed bank in reference.borrowed_from. Its corpus grew 706 → 1,574 passages,
still 6.4× coarser than the gate, so the fallback stays active.

Citation placement: unblocked, measured, not shipped

The projection fix unblocked de-AI frontier rank 6. Whole-document rank AUC is
0.906 on cited-sentence fraction, but section-matched it is 0.616–0.866
(strongest in method: human 0.1652 n=155 vs machine 0.3671 n=149),
length-matched within that section 0.835, and the human-vs-human null is
0.553. It survives every control and is the strongest model-free discriminator
in the record — and it is not shipped, because all 173 machine documents come
from one generation process and one bank cannot separate "AI cites more" from
"these prompts made it cite more".

Also

• Broadening the citation pattern reclassified 1,028 of 79,904 held-out
  sentences (1.29%) from unanchored to anchored and 0 the other way; the
  anchoring baseline was rebuilt from the corrected 517-document population.
• corpus_cos ablation withdrawn rather than deferred: confound_audit bins on
  record metadata the feature cache does not carry, so a cache-only ablation
  computes a different statistic from the three recorded retrains, and its only
  consumer is a degraded audit-only classifier with no shipped operating point.
• Three register fixtures hard-coded 5 or 6 repetitions and would have gone
  silently green-to-red at the new floor; they now derive the count from
  MIN_MANUSCRIPT_USES.
• REFERENCE_DIR and the collector's held-out rule now share one literal.

328 tests across 17 files; validator 9/9.