paperlint 2.0.0

package/skills/paper-pipeline/references/body-vs-appendix.md

@@ -0,0 +1,97 @@
+# Body, appendix, or artifact — the one rule that decides
+
+Shared reference. Read by `draft-paper` (where do I put this as I write?), `tighten-paper` (what
+moves out when the body overruns?) and `harden-paper` (is the body self-contained?). Do not copy it
+into those skills — link it, the way they link `review-ratchet.md`.
+
+## Why this file exists
+
+There was no rule. There was taste, and taste loses to a page limit every time. Author, 2026-08-06
+(translated from Russian):
+
+> _"I don't get it — the rule about the compiler-and-rule-synthesis paper, does that go mostly in
+> the body or in the appendix. I figured mostly body, with some details in the appendix, but IDK."_
+
+Without a rule the page limit does the deciding, and it decides badly: it evicts whatever was
+written last, which is usually the honest qualification nobody had room for. On `compile-rules-2026`
+that is exactly what happened, and it is traceable to a single commit. The day the contribution was
+re-typed, an admission that **eight of the 23 build-stopping findings sat inside globs covering
+tests and vendored code** was written and parked in the appendix, because the body was full. It
+qualified a headline number and no pointer in the body reached it. A reader of the abstract had no
+way to learn it existed.
+
+## The rule
+
+> **The body carries everything a reviewer's decision to accept depends on.
+> The appendix carries what a reader who already believes you wants next.
+> The artifact carries what nobody reads and everybody may need to re-run.**
+
+This is not our invention; it is the ACL/ARR rule stated positively. What the venues actually
+require, quoted:
+
+- ACL formatting guidelines: appendices _"are not required to fit within these page limits"_, they
+  come _"after the references"_, and _"it is optional for reviewers to look at appendices"_.
+- ARR author guidelines: _"the main text of the paper must be self-contained, and reviewers are not
+  expected to read the supplementary materials."_
+- ARR reviewer guidelines: _"any details important for understanding the key aspects of the work
+  should be in the paper rather than in appendices."_
+- ARR, and this one is a desk-reject: putting _"substantive parts of related work"_ only in an
+  appendix _"will be viewed as an attempt to circumvent the content page count."_
+
+**There is no size rule anywhere, and no tool checks appendix length** — `aclpubcheck` structurally
+stops at the appendix boundary. So "the appendix is too long" is never the finding. The finding is
+always _"a decision-relevant thing is not in the body"_ or _"nothing in the body sends the reader
+to this."_
+
+## Sorting test, applied per block
+
+Ask one question: **if a reviewer read only the body, would they be missing something they need to
+decide?**
+
+| goes in the BODY                                                    | goes in the APPENDIX                                | goes in the ARTIFACT                 |
+| ------------------------------------------------------------------- | --------------------------------------------------- | ------------------------------------ |
+| what the thing does, on what input, what it outputs                 | the full protocol behind a body result              | run logs, fixtures, raw per-row data |
+| the claim, and the evidence the claim rests on                      | per-item tables a reader might audit                | scoring code and its self-tests      |
+| **the price** — what the construction costs, where it stops working | comparisons you ran but do not lean on              | prompts, seeds, model versions       |
+| **every bound that narrows a headline number**                      | your own tools' failures, at length                 | anything regenerable from the above  |
+| the nearest competitor and your delta over it                       | the long form of a bound stated briefly in the body |                                      |
+
+🔴 **The asymmetric one, and the reason for this file: a qualification that narrows a number in the
+body belongs in the body, however tight the page count.** A number stated without the bound that
+narrows it is an overclaim, and moving the bound to an appendix does not fix it — it hides it and
+adds "and they buried it" to the reviewer's objection. If the page count will not take the full
+bound, state the bound in one clause in the body and put its working in the appendix. Never the
+other way round.
+
+## When the body overruns
+
+In order. Stop at the first one that fits.
+
+1. **Cut a whole block that duplicates another block.** Free.
+2. **Cut scaffolding**: rhetorical questions staging results the next sentence delivers, signpost
+   sentences restating subsection headings, a lead-in that restates its own next clause.
+3. **Move a whole block to the appendix**, leaving a one-clause statement of what it concluded.
+4. **Move a measurement to the artifact** and cite it.
+5. **Drop a claim.** An honest reduction in scope, recorded as a decision.
+
+🔴 **Never pay by trimming the subordinate clause that EXPLAINS a sentence.** It carries no fact, so
+its loss is invisible to every fact-checker and every threshold, and what remains is short, true,
+and comprehensible only to someone who already knows the idea. This is the single most common way a
+paper gets worse while every metric improves.
+
+## When the appendix is the thing that looks wrong
+
+Measure reachability, not size. Resolve every pointer from the body and Limitations into the block
+it actually answers, then split the appendix into reached and unreached. On `compile-rules-2026`
+that gave 19 of 35 blocks reached — and **two thirds of the unreached mass was not surplus**, it was
+material the body half-promised and never sent anyone to find. The fix for those is a pointer, not
+a deletion; when the body has no room for pointers, build an index at the head of the appendix.
+
+Only what stays unreached _and_ unwanted is surplus, and its destination is the artifact.
+
+## Provenance
+
+Written 2026-08-06 for `compile-rules-2026`, from a fetch of the ACL/ARR guidelines (all quotes
+above verified against the source pages) plus a reachability audit of one 3,480-word appendix. The
+prompting failure: the body sat at exactly its limit for two weeks, every cut in that time came out
+of the body, and nobody had written down what the body was _for_.

package/skills/paper-pipeline/references/credit-criteria.md

@@ -0,0 +1,69 @@
+# What makes a paper COUNT — the grading this pipeline applies (single source of truth)
+
+Why an independent researcher writes papers at all: each one is **evidence in a review of a body of
+work** — a credential file that someone else will read, score criterion by criterion, and either credit
+or refuse. Every stage of the pipeline is graded against this, not against abstract "impact." Cite this
+file from the other skills instead of re-deriving it.
+
+⚠️ **Two things this file is NOT.** It is not legal advice — where a credential review is run by a
+government or a professional body, the final call belongs to a qualified adviser in that system, and the
+statutes, policy manuals and decided cases that ground the rules below belong in the author's own private
+notes, not in a shared pipeline. And it is not a promise: what follows is how reviewers of a body of work
+have been observed to score, which is a prior, not a guarantee.
+
+## The criteria a paper feeds (a paper is a HUB — one effort, several criteria)
+
+- **Authorship of scholarly articles** — the paper itself. The test is a _scholarly article_ in a
+  _professional / major-trade / major-media_ outlet. **Peer-review + indexing is the operative bar, NOT
+  venue prestige**; peer-reviewed conference and workshop _proceedings_ qualify. Publishing
+  "**regardless of the caliber**" satisfies the plain criterion at the first step. There is **no
+  approved-venue list** — "a venue with proven acceptance" is not a thing, and anyone who offers you one
+  is selling something.
+- **Original contribution of major significance** — driven by the paper's **citations / adoption**, not
+  GitHub stars. Being funded, patented or published "will not necessarily establish… major significance";
+  what is looked for is work "highly cited relative to others". This is the riskiest criterion for an
+  independent researcher, and the paper's downstream impact — not its existence — is what feeds it.
+- **Judging the work of others** — reviewing / PC service (the organizer **review-offer email** that
+  arrives after submission). It must be _genuine_ peer review, and you need proof you _completed_ a
+  review, not merely that you were invited.
+- **Published material about you** — press or coverage _about_ the work. Distinct from authorship, and it
+  needs its own evidence (a front-page aggregator thread, a trade-press write-up).
+
+## Where venue quality actually bites (and where it does not)
+
+- **The plain criterion (first step) is venue-agnostic on prestige and newness.** A first-edition workshop
+  with indexed proceedings clears it. In decided cases this criterion is granted in a sentence, with
+  journals and conference proceedings counted interchangeably.
+- **The merits stage (second step — "sustained standing at the top of the field") is where prestige,
+  citations and recency matter.** A fresh venue does not _subtract_; it simply does not _add_ standing
+  yet — cured by citations, expert letters, and a **body of work**. Recency cuts both ways: a burst of
+  publications immediately before a filing undercuts "sustained" (a real refusal ground — papers all
+  dated to the year of filing), and stale work with declining citations fails from the other side. Target
+  a **steady cadence**.
+
+## Decisions this forces on the pipeline
+
+1. **Body of work > one trophy.** Several peer-reviewed papers in one coherent lane beat one paper at a
+   fancy venue. Keep the corpus connected.
+2. **Idea validation** (`research-ideate`): favor ideas that yield a _reusable method or benchmark_ (it
+   survives review) over a one-off "tool X is bad" result (it reads as a blog post), and that plausibly
+   earn downstream citations.
+3. **Venue choice** (`find-venue`): a peer-reviewed, **indexed** venue is the whole game; workshop newness
+   is fine; **non-archival** workshops (no indexed proceedings) rank LOW. The prestige upgrade is
+   `extend-paper` to a stronger venue later, not holding out now.
+4. **Timeline** (`plan-paper-timeline`): **space submissions out**; do not cluster them into the month
+   before a planned external filing. This is the FILING-SPACING rule, and this paragraph is its reason.
+5. **Name consistency**: publish under ONE canonical (passport) name across every paper, and use an
+   **ORCID** to tie all output to one identity regardless of how a venue spells the name. Where a form
+   asks for a country, it usually means **residence**, not citizenship — the two differ for many authors,
+   and the wrong one creates payment and sanctions friction that has nothing to do with the science.
+6. **Judging**: after each submission, offer to review for the venue; keep both the invitation and the
+   proof that you completed the review.
+7. **The personal half stays private.** Which filing, on what schedule, under whose advice — that is the
+   author's own business and belongs in the author's own notes, never in a shared or public repository.
+
+## Provenance
+
+Two papers were run on exactly this logic: **AgenticDev 2026 @ ASE** ("Measuring the Wrong Number") and
+**AISec 2026 @ ACM CCS** ("Safety Theater") — one coherent lane, both peer-reviewed and indexed, both with
+an extension path for prestige.

package/skills/paper-pipeline/references/occupancy-2026-08-06-probe/README.md

@@ -0,0 +1,35 @@
+---
+title: The probe that measured whether any shipped prose linter catches our actual defect
+created: 2026-08-06
+tags: [paper-pipeline, prose, occupancy, repro]
+---
+
+# What this is
+
+Six real sentences from `compile-rules-2026` that the author rejected on sight — *"wtf"*,
+*"VAGUE AF"*, *"confusing af"* — run through every shipped deterministic prose checker we could
+install. `sentences.txt` holds the test set; `test_sentences.py` runs proselint / write-good /
+textstat; `run_retext.mjs` runs retext-readability at several thresholds.
+
+**These scripts are the evidence, not the write-up.** The finding — that nothing flags the defect —
+is a negative result, and a negative result nobody can re-run is an opinion. Rescued out of the
+session scratchpad, which is erased when the session ends.
+
+Result and full tool-by-tool detail: `../occupancy-2026-08-06-prose-checkers.md`.
+
+## Re-running it
+
+```
+pip install proselint textstat && npm i -g write-good retext retext-readability
+python3 test_sentences.py && node run_retext.mjs
+```
+
+Both read `sentences.txt`, so adding a newly-rejected sentence to that file is how the test set
+grows. Do add them: the set is the only ground truth we have about what "unreadable but passing"
+looks like, and it is what any future check must be measured against.
+
+## One caveat, recorded rather than smoothed over
+
+Sentence 6 is a **reconstruction**. The original's exact text was not supplied to the probe, so a
+same-shape stand-in was built (six factual claims joined by dashes and semicolons). It is marked as
+such here and in the report. Replace it with the real sentence when convenient.

package/skills/paper-pipeline/references/occupancy-2026-08-06-probe/run_retext.mjs

@@ -0,0 +1,24 @@
+import {unified} from 'unified';
+import retextEnglish from 'retext-english';
+import retextReadability from 'retext-readability';
+import retextStringify from 'retext-stringify';
+import {reporter} from 'vfile-reporter';
+import fs from 'node:fs';
+
+const sentences = fs.readFileSync('sentences.txt', 'utf8')
+  .split(/\n\n+/).filter(Boolean).map(s => s.replace(/^\d+\.\s*/, '').trim());
+
+for (const [label, opts] of [
+  ['DEFAULT (age=16, minWords=5, threshold=4/7)', {}],
+  ['STRICT (age=6, minWords=1, threshold=1/7)', {age: 6, minWords: 1, threshold: 1/7}],
+]) {
+  console.log(`\n======= ${label} =======`);
+  const processor = unified().use(retextEnglish).use(retextReadability, opts).use(retextStringify);
+  for (let i = 0; i < sentences.length; i++) {
+    const file = await processor.process(sentences[i]);
+    console.log(`--- sentence ${i+1} ---`);
+    console.log(sentences[i]);
+    console.log(reporter(file));
+    console.log();
+  }
+}

package/skills/paper-pipeline/references/occupancy-2026-08-06-probe/sentences.txt

@@ -0,0 +1,11 @@
+1. The state to remove is therefore not the rule but its claim about itself.
+
+2. These files are filled with constructions nobody built.
+
+3. Three answers, all of them after the fact.
+
+4. Section 4.4 invites the suspicion that the resolver guesses.
+
+5. The only part that reads English is the one not allowed to decide anything. It proposes; the other two dispose.
+
+6. The audit found six problems: the config was stale, the hook silently no-ops on error, the guard reads a variable the harness never sets, the fallback path was never tested, the log rotates before anyone reads it, and the alert fires into a channel nobody watches.

package/skills/paper-pipeline/references/occupancy-2026-08-06-probe/test_sentences.py

@@ -0,0 +1,25 @@
+import textstat
+
+sentences = {
+    1: "The state to remove is therefore not the rule but its claim about itself.",
+    2: "These files are filled with constructions nobody built.",
+    3: "Three answers, all of them after the fact.",
+    4: "Section 4.4 invites the suspicion that the resolver guesses.",
+    5: "The only part that reads English is the one not allowed to decide anything. It proposes; the other two dispose.",
+    6: "The audit found six problems: the config was stale, the hook silently no-ops on error, the guard reads a variable the harness never sets, the fallback path was never tested, the log rotates before anyone reads it, and the alert fires into a channel nobody watches.",
+}
+
+for k, s in sentences.items():
+    print(f"--- Sentence {k} ---")
+    print(s)
+    print("words:", textstat.lexicon_count(s))
+    print("sentences:", textstat.sentence_count(s))
+    print("Flesch reading ease:", textstat.flesch_reading_ease(s))
+    print("Flesch-Kincaid grade:", textstat.flesch_kincaid_grade(s))
+    print("Gunning fog:", textstat.gunning_fog(s))
+    print("SMOG:", textstat.smog_index(s))
+    print("Coleman-Liau:", textstat.coleman_liau_index(s))
+    print("ARI:", textstat.automated_readability_index(s))
+    print("Dale-Chall:", textstat.dale_chall_readability_score(s))
+    print("difficult words:", textstat.difficult_words(s))
+    print()