paperlint 4.1.0 → 4.2.0

package/skills/paper-pipeline/grade-paper-writing-ablation.eval.mjs

@@ -56,7 +56,7 @@
  *
  *   A0  baseline          verbatim. The control.
  *   A1  short             keep sentence 1 ("Use when asking …") and sentence 2 ("Grades WRITING CRAFT
- *                         only — nine dimensions … the fix for each."). Delete everything after.
+ *                         only — ten dimensions … the fix for each."). Delete everything after.
  *                         1174 -> ~530 chars. The blunt arm: does ANY deletion revive it?
  *   A2  no-subordination  delete ONLY the two clauses announcing that this skill is a SUBROUTINE of
  *                         other skills: "whose stall inventory … is the readability gate other skills
@@ -67,7 +67,7 @@
  *                         deferring — declining to call a thing that presents itself as something
  *                         another skill calls. No other skill in the roster says this about itself.
  *   A3  no-negation       delete ONLY the sentence carrying the "NOT …" delegation clauses.
- *   A4  no-machinery      delete ONLY the opaque internals and proper nouns: the nine-dimension
+ *   A4  no-machinery      delete ONLY the opaque internals and proper nouns: the ten-dimension
  *                         parenthetical, the "(Peyton Jones, McEnerney, Gopen & Swan; exemplars …)"
  *                         authorities, the PERSONA sentence and the blind-panel/claims-diff sentence.
  *                         Keeps sentence 1 and the NOT-clauses.
@@ -76,7 +76,7 @@
  *                         "grade-paper-writing" reads as a niche in-house verb the selector skips.
  *                         The only arm that changes nothing about the description text.
  *   A6  persona-only      ADDED AFTER A0-A5 RAN, on request, and it is A1's MIRROR. The skill has two
- *                         jobs — a nine-dimension rubric and a persona cold-read stall pass — and A1
+ *                         jobs — a ten-dimension rubric and a persona cold-read stall pass — and A1
  *                         as specified keeps the rubric sentence and deletes the persona one, so the
  *                         six-arm design had a rubric-only description and no persona-only twin. A6 is
  *                         that twin: the opening sentence (minus the `"grade the writing"` trigger,
@@ -518,9 +518,9 @@ const D_NEGATION = [
   " NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT venue-bar content strength (study-accepted-papers).",
 ];
 const D_MACHINERY = [
-  " (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking)",
+  " (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue)",
   " (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake)",
-  " Runs the PERSONA cold-read stall pass (a committed non-academic persona subagent, per-section) whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper).",
+  " Runs the cold-read stall pass with TWO committed personas (a non-academic engineer per section, and a workshop reviewer skimming abstract, introduction and conclusion) whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper).",
   " Defaults to a blind multi-grader panel; after fixes, mandates the claims-preservation diff (row `claims`).",
 ];
 // NOTE ON "FOUR NOT-CLAUSES": the description carries THREE uppercase "NOT …" delegation clauses
@@ -546,7 +546,7 @@ const applyCuts = (s, cuts) => {
 // its subordination tail removed — see the header for why that tail goes.
 const D_PERSONA_ONLY = [
   ' / "grade the writing"',
-  " Grades WRITING CRAFT only — nine dimensions 1–5 (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking) against how the best-written papers read (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake), naming the offending sentence and the fix for each.",
+  " Grades WRITING CRAFT only — ten dimensions 1–5 (Title, Abstract, Intro architecture, Structure, Prose clarity, Jargon discipline, Landing-the-point, Figure economy, Honesty-without-hedge-stacking, Register distance to the target venue) against how the best-written papers read (Peyton Jones, McEnerney, Gopen & Swan; exemplars Trusting Trust, Carlini, Greshake) AND against 2–3 accepted papers of the target venue, naming the offending sentence and the fix for each.",
   " whose stall inventory — not the rubric number — is the readability gate other skills consume (pc-panel-review, paper-adversarial-review, harden-paper)",
   " Defaults to a blind multi-grader panel; after fixes, mandates the claims-preservation diff (row `claims`).",
   " NOT a content/defect review (paper-adversarial-review / pc-panel-review), NOT a structural cut plan (tighten-paper — run that first on a bloated draft), NOT venue-bar content strength (study-accepted-papers).",

package/skills/paper-pipeline/references/writing-craft.md

@@ -79,10 +79,13 @@ attributed to Kent Beck — treat as widely-cited, not primary-verified.)
 - **MapReduce** — "Simplified" in the title = the reader's _benefit_ as the promise; clean topic sentences;
   motivating example before mechanism.
 
-## The gradeable rubric — 9 dimensions, score 1–5 (anchors), most-severe first
+## The gradeable rubric — 10 dimensions, score 1–5 (anchors), most-severe first
 
-For the "wall of jargon" failure mode, **weight #5 Prose, #6 Jargon, #7 Landing-the-point ×2.** Score each,
-name the offending sentence/section, write the fix.
+For the "wall of jargon" failure mode, **weight #5 Prose, #6 Jargon, #7 Landing-the-point ×2**; for the
+"blog post" failure mode, **weight #10 Register ×2** as well. Score each, name the offending
+sentence/section, write the fix. Dimensions #1–#9 push toward plain and memorable; #10 is the floor
+under them — a fix that raises one of them by leaving the venue's register lowers #10, and the
+grade says so.
 
 1. **Title — value + memorability.** 1 = generic topic label, forgettable. 5 = states the finding or the
    reader's benefit and is quotable; repeatable after one read.
@@ -91,7 +94,7 @@ name the offending sentence/section, write the fix.
 3. **Intro architecture — problem/why-care/contribution/evidence, fast.** 1 = background/lit-tour, point
    arrives on page 2–3. 5 = SPJ arc in the first half-page + a bulleted, forward-referenced contributions list.
 4. **Structure & signposting (incl. visual density).** 1 = arbitrary order, paragraphs open mid-thought,
-   boilerplate roadmap, OR a wall — one giant unbroken paragraph the reader's eye slides off. 5 = each
+   no signposting between claims (the reader must infer how one claim relates to the next), OR a wall — one giant unbroken paragraph the reader's eye slides off. 5 = each
    section one job, every paragraph's first sentence is its claim, forward refs not a roadmap, AND no wall:
    long related-work / discussion blocks are broken into scannable chunks with bold/italic run-in sub-heads
    (see the wall-of-text check below).
@@ -115,6 +118,14 @@ validity`, `null`, `coarse`, coined handles) used with no plain-words gloss —
    self-contained. 5 = one "money figure" a skim-reader understands alone; self-contained captions; tables show the one comparison that matters.
 9. **Honesty without hedge-stacking.** 1 = overclaims OR drowns in hedges + defensive early related-work
    wall. 5 = explicit "what we do NOT claim," one clean threats-to-validity move, related work at the end as context.
+10. **Register distance to the target venue — graded against 2–3 accepted papers of that venue (or its
+    parent conference), never against the exemplars above.** 1 = reads as another genre: aphorisms and
+    punchline closings, fragments, colloquial asides, claims with no stated owner, relations between
+    claims left for the reader to infer. 5 = a reviewer at the venue could not tell it from the
+    reference papers by register: results owned (_we measured_, a citation, a named population),
+    relations signposted (_therefore_, _in contrast_, _this section_), sentences complete. Every flag
+    quotes the reference sentence it was compared against. No reference papers → not graded, and the
+    report says so; this dimension is never scored from memory.
 
 ## Scoring calibration — how to avoid an inflated grade (read before scoring)
 
@@ -175,6 +186,25 @@ bucket-1 field terms and trips on buckets 2–3. Optionally also run it on a **d
 lacks the vocabulary for real (not by role-play) and stalls even more like a human. The persona's inventory feeds
 straight into the fix pass.
 
+**Run the second persona too — the workshop reviewer. Both, never one instead of the other.** Sam cannot see two
+defects by construction: he does not know what a paper at this venue sounds like, and a sentence with no stated
+owner reads to him as confident rather than ambiguous. Those are exactly the two a real reviewer wrote down
+("difficult at first to distinguish among properties of LLMs, statistical observations from repositories, and
+experimental findings produced by the authors"; "blogpost"). Spawn a second fresh subagent:
+
+> You are a program-committee member at <venue>. You review for it every year, you have eleven papers this
+> round, and you have 40 minutes for this one. Your confidence will be 3: you know the area, not this exact
+> topic. Read the abstract, then the introduction, then the conclusion — nothing else. For EVERY sentence ask:
+> whose result does this state — the authors' own measurement, a cited work, or a general property of the
+> world? If you cannot tell, quote it and write WHOSE?. Where two sentences sit side by side and you have to
+> work out yourself how they relate, quote the pair and write RELATION?. Where a sentence would not appear in
+> a paper at <venue> — an aphorism, a punchline, a fragment, a colloquial turn — quote it and write REGISTER.
+> Then give the overall recommendation you would write after these 40 minutes, and the one sentence of your
+> review a chair would read first. Output the list and the recommendation only.
+
+Its WHOSE? / RELATION? / REGISTER lines go into the stall inventory as classes 7 and 8 of `grade-paper-writing`.
+Give it the three sections as text, and no framing: it is a reviewer who has not talked to the authors.
+
 **Run the persona PER SECTION for a dense or long (>~6pp) paper — never one whole-paper pass.** A single pass over
 a full paper skims the middle: this session a whole-paper sweep scored 44/60 and missed the walls that per-section
 persona reads (intro, method, results) each caught immediately. Chunk it to 2–3pp / one section per subagent, so
@@ -233,16 +263,18 @@ of `rm -rf` collapses to one"), a triple-nested-em-dash sentence you parse twice
 (the author's rule: don't open on the spec sheet.) When unsure which bucket a term is in, ask: _would a reviewer at THIS
 venue use it, or is it imported from stats/philosophy/our own coinage?_
 
-**De-jargon is RE-VOICING the sentence, not swapping the word.** Replacing a coined term with a plainer noun inside
-an otherwise stiff academic sentence leaves the sentence academic — the reader still stalls on the _register_, not
-just the word (GateBench 2026-07-25: a term-level de-jargon pass cleared the flagged words but the author still hit
-sections "way too academicy for 0 reason"). The fix: rewrite the whole SENTENCE the way you'd say it to a colleague
-at a whiteboard — subject–verb–object, active voice, one clause, contractions fine, no nominalizations ("performs
-an evaluation of" → "checks"). **Read each rewritten sentence ALOUD: if you'd never say it that way to a person,
-it's still academic.** E.g. "We therefore position the guard as a defense against trigger~A and a measurement
-instrument, not an injection defense" → "So the guard is there to catch accidents and to measure the problem — it
-is not built to stop an attacker." A word-swap that leaves the academic sentence shape is a HALF-fix; the
-deliverable is plain-spoken sentences, and a de-jargon/tighten pass is not done until the prose reads spoken.
+**De-jargon is RE-VOICING the sentence, not swapping the word — and the new voice is the venue's, not
+a conversation's.** Replacing a coined term with a plainer noun inside a sentence built around nominalizations
+leaves the stall in place (GateBench 2026-07-25: a term-level de-jargon pass cleared the flagged words but the
+author still hit sections "way too academicy for 0 reason"). The fix is the sentence shape: subject–verb–object,
+active voice, the actor named, no nominalizations ("performs an evaluation of" → "checks"). **The fix stops at the
+register of the venue's accepted papers:** no contractions, no fragments, no colloquial asides, and the signposting
+stays (_therefore_, _in contrast_, _we show_). E.g. "We therefore position the guard as a defense against
+trigger~A and a measurement instrument, not an injection defense" → "The guard is therefore a defense against
+accidental triggers and an instrument for measuring them; it is not designed to stop an attacker." — plainer,
+and still a sentence the venue prints. "So the guard is there to catch accidents — it is not built to stop an
+attacker" goes one step too far: it is what a reviewer quotes back as informal. Test each rewrite against a
+sentence from the reference papers of dimension #10, not against how it sounds aloud.
 
 **Step 0 — mechanical pre-scan (BEFORE reading, so the pass can't rubber-stamp).** Run the coined-compound
 grep from the avoid-list against the source and add EVERY hit to the stall inventory as a candidate; grep the
@@ -376,9 +408,11 @@ axis that catches "the conclusion isn't strong enough."
 
 ## The avoid-list — words that read as slop or stall a human (flag every hit)
 
-The north star: **a paper should be engaging and human-readable — write for a smart non-specialist who will
-read it willingly, not to sound academic.** The test (the author's 10-second rule): would a strong engineer who is
-NOT a researcher get this sentence in 10 seconds? If not, it's too dense. The jargon dimension (#6) and the
+The north star: **a paper should be readable by a smart non-specialist AND read as a paper at its venue —
+plain where plainness costs nothing, academic where the register expects it.** Two tests, both apply: would a
+strong engineer who is NOT a researcher get this sentence in 10 seconds? If not, it's too dense. Would it sit
+unremarked in one of the venue's accepted papers (dimension #10)? If not, it's too loose — a replacement that
+passes the first test by failing the second is not a fix. The jargon dimension (#6) and the
 stall pass are the judgment; the list below is the _checkable_ part (prose isn't policy — compile what's
 mechanical). These are **scrutinize-words, not absolute bans** (some are legit in context — "significant" for
 a real stat, "method"); on every hit, ask "does this earn its place, or is it slop / hype / a reader-stall?"
@@ -435,7 +469,8 @@ number, so the paper does not have to print them all — divide the labor:
 
 1. **DO lead with the concrete crazy instance, not the aggregate stat** (Carlini's UUID).
 2. **DO manufacture the "wait, what?" pivot** — surface the held assumption, then break it (Greshake).
-3. **DO write the intro as the SPJ arc + a bulleted, forward-referenced contributions list**; kill the "organized as follows" roadmap.
+3. **DO write the intro as the SPJ arc + a bulleted, forward-referenced contributions list**; keep an
+   "organized as follows" paragraph only when the venue's accepted papers carry one.
 4. **DON'T imitate Attention-style dense contribution-less prose** unless the audience already cares.
 5. **DON'T stack jargon or hedges** — gloss every term plainly at first use; replace hedge-piles with one
    precise scope sentence; land one repeatable moral (Trusting Trust).

package/skills/pc-panel-review/SKILL.md

@@ -1,10 +1,10 @@
 ---
 name: pc-panel-review
-description: Use when asking "what would the program committee decide?" / "simulate the reviewers" / "what's this paper's accept probability?" on a drafted paper + artifact. Spawns N independent reviewers with DISTINCT lenses (one actually RUNS the artifact), then synthesizes a PC-chair meta-review into an accept/reject decision, a calibrated probability, and consensus must-fixes. Requires grade-paper-writing's persona stall inventory + tighten-paper's structural verdict as inputs (blocked until they exist — the panel cannot feel reader fatigue on its own). Also has a lightweight single-reviewer VENUE-FIT MODE for a quick CFP-fit spot-check. NOT a single fast defect hunt (paper-adversarial-review), a writing grade (grade-paper-writing), or the full pre-submit gate (harden-paper, which calls this) — this models the PC decision itself.
+description: Use when asking "what would the program committee decide?" / "simulate the reviewers" / "what's this paper's accept probability?" on a drafted paper + artifact. Spawns N independent reviewers with DISTINCT lenses (one actually RUNS the artifact, one is a skimming genre/register reviewer), then synthesizes a PC-chair meta-review into an accept/reject decision reported as min/median/max, a calibrated probability, and consensus must-fixes. Requires grade-paper-writing's persona stall inventory + tighten-paper's structural verdict as inputs (blocked until they exist — the panel cannot feel reader fatigue on its own). Also has a lightweight single-reviewer VENUE-FIT MODE for a quick CFP-fit spot-check. NOT a single fast defect hunt (paper-adversarial-review), a writing grade (grade-paper-writing), or the full pre-submit gate (harden-paper, which calls this) — this models the PC decision itself.
 allowed-tools: [Read, Write, Grep, Glob, Bash, WebSearch, WebFetch, Agent, Skill]
 ---
 
-<!-- vigiles:sha256:b9feb099347291ff compiled from skills/pc-panel-review/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:7e5fb04d1cfa0548 compiled from skills/pc-panel-review/SKILL.md.spec.ts -->
 
 # pc-panel-review — model the whole PC, not one reviewer
 
@@ -51,7 +51,15 @@ prose-only read can.
    panel's only legitimate source for the readability/structure signal.
    *(Scorecard: this gate is row **`panel`**, and it declares **`structure`** (tighten-paper) + **`writing`**
    (grade-paper-writing) as its required inputs.)*
-3. **Pick 3 (or 4) reviewer lenses** that a real PC for THIS venue would field. Menu — choose by fit:
+3. **Pick 3 (or 4) reviewer lenses** that a real PC for THIS venue would field. One is fixed; the rest
+   come from the menu, by fit:
+   - **Genre/register reviewer — MANDATORY on every panel, at every venue.** Confidence 3 (knows the
+     area, not this topic), reads ONLY the abstract, the introduction and the conclusion, and judges
+     whether this reads as a paper at this venue: whose result each sentence states, whether the
+     relations between claims are signposted or left to infer, and whether anything is an aphorism,
+     a fragment or a colloquial turn. This is the reviewer who writes one short review, scores low,
+     and calls the paper a blog post — the one reviewer a panel of careful full readers never
+     predicts. Give it the three sections as text and the venue's name, nothing else.
    - **Methods/stats skeptic** — sample size, construct validity, multiple comparisons, overclaims,
      whether the CIs support the claims.
    - **Domain practitioner** — relevance to the CFP, novelty vs the nearest prior work, does the
@@ -67,7 +75,13 @@ prose-only read can.
 4. **Spawn them in PARALLEL, each on a separate model instance** (e.g. Fable via subagent), each blind
    to the others — independence is the whole point. Give each: the paper (`.tex`/PDF), the artifact
    path, the quoted rubric, its lens, **and the two step-2 artifacts (persona stall inventory +
-   tighten-paper structural verdict)**. Force a filled review form (below).
+   tighten-paper structural verdict)** — the genre/register reviewer gets only its three sections.
+   Force a filled review form (below).
+
+   🔴 **No reviewer is shown a previous round** — not its scores, not its reviews, not "the panel
+   gave 3.25 last time", not a list of what changed since. A reviewer who is told the last score
+   anchors on it (*"up from the 3.25 the earlier rounds reported"*), and the rounds then converge on
+   each other instead of on the paper. Comparing rounds is the chair's job, after the reviews land.
 5. **Synthesize the meta-review yourself** (PC chair) once all land — see output.
 
 ## Each reviewer returns a review form
@@ -90,10 +104,19 @@ prose-only read can.
 - One-line meta: does this belong at this venue?
 
 ## The meta-review (PC-chair synthesis) — the actual deliverable
-- **Score table:** each reviewer × each criterion + overall + confidence, at a glance.
-- **Decision:** the rec the PC discussion converges on (weight by confidence; a lone low-confidence
-  outlier doesn't sink two confident accepts), + a **calibrated accept probability** for this venue
-  and edition.
+- **Score table:** each reviewer × each criterion + overall + confidence, at a glance, and under it
+  the overall score's **min / median / max** across reviewers — never a mean alone, which hides the
+  one low review that decides a workshop.
+- **Decision:** for a **workshop**, the decision follows the **MIN**: a workshop PC with three
+  reviews and a short discussion does not overrule a reject, so one Reject or Strong Reject sinks the
+  paper however confident the others are. For a conference, the rec the PC discussion converges on,
+  weighted by confidence, with the min named beside it. Either way, a **calibrated accept probability**
+  for this venue and edition.
+- 🔴 **One reviewer is not a panel.** A run with fewer than three reviews and a meta-review — a
+  Venue-fit spot-check, a single re-review, a panel whose other seats did not return — is recorded as
+  what it is (`venue-fit`, `single review`), never in the scorecard's `panel` row. A single
+  review recorded as a panel verdict is how a paper reaches submission on an 88% that three real
+  reviewers turned into a reject.
 - **Consensus must-fixes:** issues **≥2 reviewers independently raised** — these are the real ones;
   apply before submitting.
 - **Single-reviewer flags:** noted, triaged (fix if cheap, else defer to camera-ready).
@@ -199,9 +222,11 @@ reject-risk without rubber-stamping. For the real decision, escalate to the full
 ## The reproduce loop (why this beats a prose review)
 This is "double-blind review as a real venue does it, WITH reproduction." The artifact-runner reviewer
 is the difference-maker on both runs so far: it doesn't judge the paper's numbers, it re-derives them.
-Run the panel, apply the consensus fixes, then **re-run the panel** (same lenses, told what changed) to
-confirm the fixes landed and nothing regressed — the artifact-runner re-executes every harness each
-round. Track the accept-probability progression in the review ledger. Stop when the panel converges to
+Run the panel, apply the consensus fixes, then **re-run the panel** — same lenses, FRESH reviewers,
+told nothing about the previous round — to confirm the fixes landed and nothing regressed; the
+artifact-runner re-executes every harness each round. The chair, not the reviewers, sets the new
+reviews beside the old ones and checks that each earlier must-fix is gone. Track the min / median / max
+progression in the review ledger. Stop when the panel converges to
 accept and the artifact reproduces clean; don't loop past one confirming round.
 
 ## Provenance (two papers, battle-tested)

package/skills/pc-panel-review/SKILL.md.spec.ts

@@ -9,7 +9,7 @@ import { experimental_skill } from "vigiles/spec";
 export default experimental_skill({
   name: "pc-panel-review",
   description:
-    'Use when asking "what would the program committee decide?" / "simulate the reviewers" / "what\'s this paper\'s accept probability?" on a drafted paper + artifact. Spawns N independent reviewers with DISTINCT lenses (one actually RUNS the artifact), then synthesizes a PC-chair meta-review into an accept/reject decision, a calibrated probability, and consensus must-fixes. Requires grade-paper-writing\'s persona stall inventory + tighten-paper\'s structural verdict as inputs (blocked until they exist — the panel cannot feel reader fatigue on its own). Also has a lightweight single-reviewer VENUE-FIT MODE for a quick CFP-fit spot-check. NOT a single fast defect hunt (paper-adversarial-review), a writing grade (grade-paper-writing), or the full pre-submit gate (harden-paper, which calls this) — this models the PC decision itself.',
+    'Use when asking "what would the program committee decide?" / "simulate the reviewers" / "what\'s this paper\'s accept probability?" on a drafted paper + artifact. Spawns N independent reviewers with DISTINCT lenses (one actually RUNS the artifact, one is a skimming genre/register reviewer), then synthesizes a PC-chair meta-review into an accept/reject decision reported as min/median/max, a calibrated probability, and consensus must-fixes. Requires grade-paper-writing\'s persona stall inventory + tighten-paper\'s structural verdict as inputs (blocked until they exist — the panel cannot feel reader fatigue on its own). Also has a lightweight single-reviewer VENUE-FIT MODE for a quick CFP-fit spot-check. NOT a single fast defect hunt (paper-adversarial-review), a writing grade (grade-paper-writing), or the full pre-submit gate (harden-paper, which calls this) — this models the PC decision itself.',
   tools: [
     "Read",
     "Write",
@@ -67,7 +67,15 @@ prose-only read can.
    panel's only legitimate source for the readability/structure signal.
    *(Scorecard: this gate is row **\`panel\`**, and it declares **\`structure\`** (tighten-paper) + **\`writing\`**
    (grade-paper-writing) as its required inputs.)*
-3. **Pick 3 (or 4) reviewer lenses** that a real PC for THIS venue would field. Menu — choose by fit:
+3. **Pick 3 (or 4) reviewer lenses** that a real PC for THIS venue would field. One is fixed; the rest
+   come from the menu, by fit:
+   - **Genre/register reviewer — MANDATORY on every panel, at every venue.** Confidence 3 (knows the
+     area, not this topic), reads ONLY the abstract, the introduction and the conclusion, and judges
+     whether this reads as a paper at this venue: whose result each sentence states, whether the
+     relations between claims are signposted or left to infer, and whether anything is an aphorism,
+     a fragment or a colloquial turn. This is the reviewer who writes one short review, scores low,
+     and calls the paper a blog post — the one reviewer a panel of careful full readers never
+     predicts. Give it the three sections as text and the venue's name, nothing else.
    - **Methods/stats skeptic** — sample size, construct validity, multiple comparisons, overclaims,
      whether the CIs support the claims.
    - **Domain practitioner** — relevance to the CFP, novelty vs the nearest prior work, does the
@@ -83,7 +91,13 @@ prose-only read can.
 4. **Spawn them in PARALLEL, each on a separate model instance** (e.g. Fable via subagent), each blind
    to the others — independence is the whole point. Give each: the paper (\`.tex\`/PDF), the artifact
    path, the quoted rubric, its lens, **and the two step-2 artifacts (persona stall inventory +
-   tighten-paper structural verdict)**. Force a filled review form (below).
+   tighten-paper structural verdict)** — the genre/register reviewer gets only its three sections.
+   Force a filled review form (below).
+
+   🔴 **No reviewer is shown a previous round** — not its scores, not its reviews, not "the panel
+   gave 3.25 last time", not a list of what changed since. A reviewer who is told the last score
+   anchors on it (*"up from the 3.25 the earlier rounds reported"*), and the rounds then converge on
+   each other instead of on the paper. Comparing rounds is the chair's job, after the reviews land.
 5. **Synthesize the meta-review yourself** (PC chair) once all land — see output.
 
 ## Each reviewer returns a review form
@@ -106,10 +120,19 @@ prose-only read can.
 - One-line meta: does this belong at this venue?
 
 ## The meta-review (PC-chair synthesis) — the actual deliverable
-- **Score table:** each reviewer × each criterion + overall + confidence, at a glance.
-- **Decision:** the rec the PC discussion converges on (weight by confidence; a lone low-confidence
-  outlier doesn't sink two confident accepts), + a **calibrated accept probability** for this venue
-  and edition.
+- **Score table:** each reviewer × each criterion + overall + confidence, at a glance, and under it
+  the overall score's **min / median / max** across reviewers — never a mean alone, which hides the
+  one low review that decides a workshop.
+- **Decision:** for a **workshop**, the decision follows the **MIN**: a workshop PC with three
+  reviews and a short discussion does not overrule a reject, so one Reject or Strong Reject sinks the
+  paper however confident the others are. For a conference, the rec the PC discussion converges on,
+  weighted by confidence, with the min named beside it. Either way, a **calibrated accept probability**
+  for this venue and edition.
+- 🔴 **One reviewer is not a panel.** A run with fewer than three reviews and a meta-review — a
+  Venue-fit spot-check, a single re-review, a panel whose other seats did not return — is recorded as
+  what it is (\`venue-fit\`, \`single review\`), never in the scorecard's \`panel\` row. A single
+  review recorded as a panel verdict is how a paper reaches submission on an 88% that three real
+  reviewers turned into a reject.
 - **Consensus must-fixes:** issues **≥2 reviewers independently raised** — these are the real ones;
   apply before submitting.
 - **Single-reviewer flags:** noted, triaged (fix if cheap, else defer to camera-ready).
@@ -215,9 +238,11 @@ reject-risk without rubber-stamping. For the real decision, escalate to the full
 ## The reproduce loop (why this beats a prose review)
 This is "double-blind review as a real venue does it, WITH reproduction." The artifact-runner reviewer
 is the difference-maker on both runs so far: it doesn't judge the paper's numbers, it re-derives them.
-Run the panel, apply the consensus fixes, then **re-run the panel** (same lenses, told what changed) to
-confirm the fixes landed and nothing regressed — the artifact-runner re-executes every harness each
-round. Track the accept-probability progression in the review ledger. Stop when the panel converges to
+Run the panel, apply the consensus fixes, then **re-run the panel** — same lenses, FRESH reviewers,
+told nothing about the previous round — to confirm the fixes landed and nothing regressed; the
+artifact-runner re-executes every harness each round. The chair, not the reviewers, sets the new
+reviews beside the old ones and checks that each earlier must-fix is gone. Track the min / median / max
+progression in the review ledger. Stop when the panel converges to
 accept and the artifact reproduces clean; don't loop past one confirming round.
 
 ## Provenance (two papers, battle-tested)

package/src/adapters/latex/document-class.ts

@@ -2,6 +2,7 @@
 import {
   documentClassLine,
   venueClass,
+  type ClassCandidate,
   type ClassLine,
   type DocumentClass,
 } from "../../domain/tex-document.ts";
@@ -52,11 +53,12 @@ const optionText = (src: string, part: readonly Node[]): string =>
       .join(""),
   );
 
-const classNode = (t: ParsedTex): Macro | undefined =>
-  t.root.content.find(isDocumentClass);
+/** Every `\documentclass` at the top of the tree, in source order. */
+const classNodes = (t: ParsedTex): readonly Macro[] =>
+  t.root.content.filter(isDocumentClass);
 
 /** A `\documentclass` node as a class line: empty, or the class with its options as read. */
-function classLine(src: string, node: Macro): ClassLine {
+function classLine(src: string, node: Macro): ClassCandidate {
   const place = macroPlace(node);
   const cls = collapse(textOf(mandatory(node)[0]?.content));
   if (cls === "") return { kind: "empty", place };
@@ -67,13 +69,18 @@ function classLine(src: string, node: Macro): ClassLine {
 }
 
 /**
- * The first `\documentclass` at the top of the tree: missing; empty (`\documentclass{}`); or the
- * class with each of its options as written (`foo={a,b}` stays one option, braces kept), whitespace
- * collapsed. Its place spans the macro and its arguments.
+ * The `\documentclass` at the top of the tree: missing; empty (`\documentclass{}`); the class with
+ * each of its options as written (`foo={a,b}` stays one option, braces kept), whitespace collapsed,
+ * its place spanning the macro and its arguments; or, when there are several (behind a TeX switch),
+ * each of them as a candidate.
  */
 export function documentClassOf(t: ParsedTex): ClassLine {
-  const node = classNode(t);
-  return node === undefined ? { kind: "missing" } : classLine(t.src, node);
+  const lines = classNodes(t).map((n) => classLine(t.src, n));
+  const [first, second, ...rest] = lines;
+  if (first === undefined) return { kind: "missing" };
+  return second === undefined
+    ? first
+    : { kind: "ambiguous", candidates: [first, second, ...rest] };
 }
 
 /**
@@ -118,15 +125,16 @@ function addOptions(
 /**
  * The source with its `\documentclass` line changed to satisfy `want` (see `venueClass`): another
  * class is replaced by the template's line; the same class gains the options it lacks in place. A
- * source whose class already satisfies `want`, or that has no class line the parser could place, is
- * returned as it is.
+ * source whose class already satisfies `want`, that has no class line the parser could place, or
+ * that has several, is returned as it is.
  */
 export function replaceDocumentClass(
   t: ParsedTex,
   want: DocumentClass,
 ): string {
-  const node = classNode(t);
-  if (node === undefined) return t.src;
+  // Several lines behind a switch: which one builds is not known, so none is rewritten.
+  const [node, ...others] = classNodes(t);
+  if (node === undefined || others.length > 0) return t.src;
   const line = classLine(t.src, node);
   if (line.kind !== "class" || line.place.kind === "unplaced") return t.src;
   const { start, end } = line.place.span;

package/src/adapters/latex/includes.ts

@@ -0,0 +1,66 @@
+/** The files a source pulls in, and the body of a source that is a document of its own. */
+import type * as Ast from "@unified-latex/unified-latex-types";
+import type { Include } from "../../domain/paper-source.ts";
+import type { Span } from "../../domain/tex-document.ts";
+import { isDefinition } from "./hidden.ts";
+import {
+  inPlace,
+  isNode,
+  macroPlace,
+  mandatory,
+  placeOf,
+  visited,
+  type Macro,
+  type Node,
+} from "./nodes.ts";
+import type { ParsedTex } from "./parse.ts";
+
+const INCLUDE_MACROS: ReadonlyMap<string, Include["macro"]> = new Map([
+  ["input", "input"],
+  ["include", "include"],
+  ["subfile", "subfile"],
+]);
+
+/**
+ * The path an include names, as written: its braced argument, or — TeX's own form, `\input name` —
+ * the name up to the next space. unified-latex gives the unbraced form an argument of the first
+ * token only (`sections` of `sections/a`), so the source is read for it.
+ */
+function argumentText(src: string, m: Macro): string | null {
+  const start = mandatory(m)[0]?.content[0]?.position?.start.offset;
+  if (start === undefined) return null;
+  const rest = src.slice(start);
+  const stop = rest.search(src[start - 1] === "{" ? /\}/u : /[\s{}%]/u);
+  return stop < 0 ? rest : rest.slice(0, stop);
+}
+
+/** One include macro as an `Include`, or nothing when it names no file or has no place. */
+function includeOf(src: string, m: Macro): readonly Include[] {
+  const macro = INCLUDE_MACROS.get(m.content);
+  const target = argumentText(src, m);
+  if (macro === undefined || target === null || target.trim() === "") return [];
+  return inPlace(macroPlace(m), (span) => [{ macro, target, span }]);
+}
+
+/**
+ * Every `\input`, `\include` and `\subfile` of the source, in source order — not in a comment (a
+ * comment is a node of its own) and not in a macro definition's body, which is read where the macro
+ * is used, and nothing here expands macros.
+ */
+export const includesOf = (t: ParsedTex): readonly Include[] =>
+  visited(t.root, isDefinition)
+    .filter(isNode)
+    .flatMap((n) => (n.type === "macro" ? includeOf(t.src, n) : []));
+
+const isDocument = (n: Node): n is Readonly<Ast.Environment> =>
+  n.type === "environment" && n.env === "document";
+
+/** The span of the `document` environment's body, or null when the source has none. */
+export const documentBodyOf = (t: ParsedTex): Span | null =>
+  (t.root.content.find(isDocument)?.content ?? [])
+    .flatMap((n) => inPlace(placeOf(n), (span) => [span]))
+    .reduce<Span | null>(
+      (body, span) =>
+        body === null ? span : { start: body.start, end: span.end },
+      null,
+    );