paperlint 2.0.0

package/skills/submit-paper/references/venues/realm.md

@@ -0,0 +1,155 @@
+# REALM @ EMNLP — venue card
+
+**Family: ACL.** Everything ACL-wide (appendix rule, Limitations/Ethics not counting, anonymity,
+OpenReview mechanics) lives in `paper-pipeline/references/acl-venue-rules.md`. Only REALM-specific
+facts are here.
+
+Verified by direct fetch of `realm-workshop.github.io/call_for_papers` **2026-08-02 and 2026-08-03**
+(twice, independently), plus a live pull of the OpenReview invitation schema.
+🔴 **Re-verify every cycle** — dates, limits and tracks move year to year.
+
+## The 2026 edition (2nd edition)
+
+| field                      | value                                                                                                         |
+| -------------------------- | ------------------------------------------------------------------------------------------------------------- |
+| full name                  | REALM — workshop at EMNLP 2026                                                                                |
+| edition                    | **2nd** (a 1st-edition accepted corpus exists — mine it with `study-accepted-papers`, in SETUP)               |
+| direct submission deadline | **Wednesday 5 August 2026**, 23:59 **AoE**                                                                    |
+| → in the author's zone     | at UTC+5: **6 August 2026, 16:59** (AoE = UTC−12; 5 Aug 23:59 AoE = 6 Aug 11:59 UTC)                          |
+| ARR commitment deadline    | 31 August 2026, same AoE convention                                                                           |
+| long paper                 | **8 pages of content**                                                                                        |
+| short paper                | 4 pages of content                                                                                            |
+| refs + appendices          | **unlimited**                                                                                                 |
+| camera-ready               | **+1 page** (9 long / 5 short)                                                                                |
+| archival                   | optional at submission; archival → **ACL Anthology proceedings**                                              |
+| blind                      | **double-blind** — names and affiliations omitted                                                             |
+| preprints                  | explicitly allowed, any time, provided the PDF does not link to de-anonymised info                            |
+| portal                     | `openreview.net/group?id=EMNLP/2026/Workshop/REALM`                                                           |
+| format                     | **PDF only**                                                                                                  |
+| mandatory sections         | none beyond ACL's Limitations; _"authors are expected to attest to the ethical considerations in their work"_ |
+| venue mode                 | hybrid; held Thursday 29 October, 9:00–17:30, Room P1                                                         |
+
+## 🔴 The submission form has NO supplementary field
+
+Pulled live from `api2.openreview.net/invitations?id=EMNLP/2026/Workshop/REALM/-/Submission`. The
+form's **complete** field list:
+
+```
+title · authors · authorids · keywords · TLDR · abstract · pdf ·
+archival · cross_submission_to · serve_as_reviewer · venue · venueid
+```
+
+There is **no Responsible NLP Checklist and no supplementary upload**. Two consequences:
+
+1. **The artifact must be hosted externally and anonymously**, with the URL inside the PDF. Budget
+   time for this — it is a submission blocker, and an artifact reviewer will refuse the Available
+   badge without a resolving URL.
+2. Do not confuse this route with **ARR**, whose portal _does_ have the checklist and separate
+   software/data uploads. Different portal, different process. A previous version of this project's
+   checklist conflated them.
+3. **No checklist field means no LLM-disclosure gate here**, and therefore no required placement for
+   it — put the disclosure in Limitations or Ethical considerations, both outside the page limit,
+   and keep it out of the reader's first contact with a measurement. Full rule, with the policy text
+   and the failure it was written from: `paper-pipeline/references/acl-venue-rules.md`, §GenAI / LLM
+   disclosure.
+
+Note `keywords` is a form field, so the keyword line does **not** belong in the typeset PDF.
+
+## The submission form, field by field (filled 2026-08-06, ~15 min before deadline)
+
+Transcribed from the live form while filling it. The API schema above lists field _names_; this is
+what the form actually shows, in order, with what each one wants. A field's help text is quoted.
+
+| #   | field                   | required | type                                         | what goes in                                                                                                                                                                                                                                   |
+| --- | ----------------------- | -------- | -------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| 1   | **Title**               | ✱        | one line                                     | _"Add TeX formulas using $In-line$ or $$Block$$"_ — plain text otherwise. **Must match the PDF's title exactly**; a late title edit in the paper means re-uploading the PDF _and_ retyping this                                                |
+| 2   | **Authors**             | ✱        | profile search                               | pre-filled with the submitting author. _"All authors must have an OpenReview profile prior to submitting"_ — a co-author without an active profile blocks the whole submission                                                                 |
+| 3   | **Keywords**            | ✱        | comma-separated                              | free text, no controlled vocabulary. 🔴 **Because this is a form field, the keyword line does NOT belong in the typeset PDF**                                                                                                                  |
+| 4   | **TL;DR**               | —        | one sentence, 🔴 **hard max 250 characters** | _"a short sentence describing your paper"_. The limit is **not in the help text** — it rejects on submit. Optional, but it is the first thing a bidding reviewer reads: name the mechanism, not the topic. Count the characters before pasting |
+| 5   | **Abstract**            | ✱        | markdown + TeX, Write/Preview tabs           | paste the paper's abstract. Markdown emphasis renders; `**bold**` survives. Backticks render as code                                                                                                                                           |
+| 6   | **PDF**                 | ✱        | file upload                                  | _"Upload a PDF file that ends with .pdf"_. No supplementary field anywhere on the form — see above                                                                                                                                             |
+| 7   | **Archival**            | ✱        | radio: Archival / Non-archival               | Archival → **ACL Anthology**. _"If your paper is a cross-submission (currently under review or already published at another venue), it cannot be archival, per ACL policies."_ Changeable at camera-ready                                      |
+| 8   | **Cross Submission To** | —        | one line                                     | leave **empty** unless the paper is genuinely under review elsewhere. Filling it forfeits archival                                                                                                                                             |
+| 9   | **Serve As Reviewer**   | ✱        | profile search                               | _"reciprocal reviewing practice… nominate at least one author to serve as a reviewer using their OpenReview profile ID (e.g. ~First_Last1)"_. **A solo author nominates themself** — the field is required and there is nobody else            |
+| 10  | License                 | ✱        | fixed chip                                   | **CC BY 4.0**, not editable                                                                                                                                                                                                                    |
+| 11  | Readers                 | ✱        | fixed chips                                  | `EMNLP 2026 Workshop REALM` + `authorids`                                                                                                                                                                                                      |
+| 12  | Signatures              | ✱        | fixed chip                                   | the submitting author                                                                                                                                                                                                                          |
+
+Then **Submit** / Cancel. An Edit History block below repeats Readers + Signatures — informational.
+
+🔴 **The displayed deadline is NOT when writes stop.** An OpenReview invitation carries two dates
+and the interface shows only the first:
+
+| field                                                          | 2026 edition, UTC | at UTC+5  |
+| -------------------------------------------------------------- | ----------------- | --------- |
+| `duedate` — printed everywhere as the deadline                 | 2026-08-06 11:59  | **16:59** |
+| `expdate` — when the invitation actually stops accepting edits | 2026-08-06 12:29  | **17:29** |
+
+A **30-minute grace window**, undocumented in the CFP and invisible on the form. Read it before the
+last hour, not after — it is the difference between shipping a fix and abandoning it:
+
+```
+curl -sS "https://api2.openreview.net/invitations?id=<VENUE-PATH>/-/Submission" \
+  | python3 -c "import json,sys,datetime;[print(k, datetime.datetime.utcfromtimestamp(i[k]/1000)) \
+    for i in json.load(sys.stdin)['invitations'] for k in ('duedate','expdate') if i.get(k)]"
+```
+
+Do **not** plan around it — a venue may set `expdate == duedate`, and 30 minutes is not a schedule.
+Treat it as recovery room for a defect found at the buzzer. Re-read it every cycle.
+
+**Traps hit for real:**
+
+- 🔴 **The title lives in two places and they silently diverge.** A title change at T−14 min meant a
+  fresh PDF _and_ a retyped Title field. Change the paper first, rebuild, then fill the form — never
+  the other way round.
+- 🔴 **"Serve As Reviewer" surprises solo authors, and it is not a formality — it enrols you.** It reads
+  as though it wants a colleague. It is required, and the answer for a one-author paper is your own
+  profile ID. **The day after submitting, the program chairs assign papers and set a review deadline**
+  (2026: submissions 5 Aug, reviews due 21 Aug AoE — sixteen days). Consequences to price in BEFORE
+  submitting, not after:
+  - **ACL's AI policy applies to the reviews.** Verbatim from the chairs' email: _"the use of
+    generative AI tools to write an entire review is prohibited. While such tools may be used for
+    minor editing (e.g., grammar or clarity), the content and assessment must be the reviewer's own."_
+    Every review hour is personal. The sanction for violating it lands on submissions to the next ARR
+    cycle — the same ecosystem the paper is in.
+  - **The workshop's PC roster is not published**, so the only durable proof of the service is a
+    Web of Science verified-review record: forward the chairs' thank-you email to
+    `reviews@webofscience.com`. Set that profile up BEFORE reviewing, not after.
+  - It is nonetheless a real judging credential, distinct from authorship. One submission closes two
+    criteria — but the reviewing window is not negotiable and may collide with other commitments.
+
+  🔴 **Hit for real on `compile-rules-2026`.** This project had _decided on 2026-07-26 not to serve as
+  a REALM reviewer_, with three written reasons (no public roster, the window collides with an ASPLOS
+  artifact round, the AI policy makes it all personal hours). Eleven days later the required form field
+  enrolled him anyway, because the field was filled in from a submission checklist that did not know
+  about the decision. **A required form field can silently reverse a recorded decision.** Before
+  filling any required field, check whether the project has already decided that question.
+
+- **Keywords are a form field, not paper content** — so anything in the PDF's keyword line is
+  duplicated dead weight. Keep the list in a non-typeset comment in the source and paste it here.
+- **The abstract field accepts markdown**, so the source abstract can be pasted nearly verbatim.
+  Extracting it from the _built PDF_ instead is a mistake: a text extractor on the line-numbered
+  build interleaves the margin numbers and silently drops paragraphs.
+
+## The ARR route is not a free fallback
+
+The 31 August "ARR commitment" deadline commits a paper **that has already been reviewed by ACL
+Rolling Review**. It is not a second chance for a paper that was never submitted to ARR — that needs
+an ARR submission first, on ARR's own cycle. Do not plan around it without checking the ARR calendar.
+
+## Gotchas hit for real
+
+- **Deadline confusion.** The CFP prints "August 5th" and states the AoE rule in a _different_
+  sentence on the same page. A pass on 2026-07-31 concluded "AoE not stated, assume noon 5 Aug" and
+  was wrong by 29 hours. Read the whole page, then convert.
+- **OpenReview profile moderation is the critical path**, not the writing. An independent author with
+  no institutional email and no DBLP record has no expedite route. Start the profile weeks ahead.
+- The workshop's hybrid status is confirmed by REALM's **own** root page — do not carry it over from
+  EMNLP's general page.
+
+## Sources
+
+- `realm-workshop.github.io/call_for_papers` — CFP (fetched 2026-08-02, 2026-08-03)
+- `realm-workshop.github.io/` — root, hybrid confirmation
+- `api2.openreview.net/invitations?id=…/-/Submission` — the live form schema
+- Per-paper compliance state: `<paper-dir>/SUBMIT-CHECKLIST.md`

package/skills/submit-paper/references/venues/tex-base.jsonc

@@ -0,0 +1,50 @@
+// The TeX Live packages EVERY paper gets — and ALL a paper gets when its venue is unknown (no
+// venue.json, or a venue.json naming a venue with no profile here). A venue's own `tex.packages`
+// is added on top of this set, never instead of it.
+//
+// Read by `paperlint toolchain` (it installs this set plus every venue's) and by `paperlint build` (it checks
+// that the TeX Live it is about to use has this set plus the paper's venue's). Validated against
+// venue-profile.schema.json next door.
+//
+// 🔴 THE CONTRACT IS THE FILES, NOT THE NAMES. A package name is what tlmgr is asked for; the
+// files listed under it are what `kpsewhich` must find afterwards. An installer's exit code is not
+// the criterion — on 2026-09-01 an apt step reported success having installed zero packages.
+//
+// What the base set is FOR: a plain `article` paper with the usual tools, typeset in Type 1 fonts.
+//   cm-super — without it pdflatex silently RASTERIZES Computer Modern in some contexts (a Type 3
+//     bitmap font in the PDF; measured 2026-09-01 on an ACL paper: 1 Type 3 font before, 0 after).
+//     ACM and ACL both reject such a PDF, and the build does not complain.
+{
+  "tex": {
+    "packages": {
+      "latex": ["latex.ltx", "article.cls"],
+      "bibtex": ["plain.bst"],
+      "hyperref": ["hyperref.sty"],
+      "geometry": ["geometry.sty"],
+      "graphics": ["graphicx.sty"],
+      "tools": ["array.sty"],
+      "amsmath": ["amsmath.sty"],
+      "amsfonts": ["amssymb.sty"],
+      "url": ["url.sty"],
+      "xcolor": ["xcolor.sty"],
+      "booktabs": ["booktabs.sty"],
+      "natbib": ["natbib.sty"],
+      "microtype": ["microtype.sty"],
+      "caption": ["caption.sty"],
+      "enumitem": ["enumitem.sty"],
+      "xurl": ["xurl.sty"],
+      "pgf": ["tikz.sty"],
+      "seqsplit": ["seqsplit.sty"],
+      "cm-super": ["sfrm1000.pfb"],
+    },
+    // Programs the skills run over a paper, installed WITH TeX Live because they ARE TeX Live
+    // packages. The build does not need them, so a system TeX Live without them still builds.
+    //   texcount   — the word count a paper's length is measured by. Added to the old CTAN list on
+    //                2026-09-17 after a consumer's CI went red on a cache miss without it.
+    //   checkcites — asks whether a bibliography entry is uncited.
+    "tools": {
+      "texcount": ["texcount"],
+      "checkcites": ["checkcites"],
+    },
+  },
+}

package/skills/submit-paper/references/venues/venue-profile.schema.json

@@ -0,0 +1,74 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "https://github.com/zernie/research-paper-pipeline/venue-profile.schema.json",
+  "title": "Venue profile",
+  "description": "One venue's format profile (venues/<venue>.jsonc) and the TeX base set (venues/tex-base.jsonc). The format numbers are read by the consumer's pdf/profile rule; `tex` is read by `paperlint toolchain` and `paperlint build`.",
+  "type": "object",
+  "required": ["tex"],
+  "additionalProperties": false,
+  "properties": {
+    "tex": { "$ref": "#/definitions/tex" },
+    "template": {
+      "description": "The \\documentclass the venue's template uses.",
+      "type": "string",
+      "minLength": 1
+    },
+    "page_size": { "type": "string", "enum": ["letter", "a4"] },
+    "page_w_in": { "type": "number", "exclusiveMinimum": 0 },
+    "page_h_in": { "type": "number", "exclusiveMinimum": 0 },
+    "textblock_w_in": { "type": "number", "exclusiveMinimum": 0 },
+    "textblock_h_in": { "type": "number", "exclusiveMinimum": 0 },
+    "columns": { "type": "integer", "minimum": 1 },
+    "body_pt": { "type": "number", "exclusiveMinimum": 0 },
+    "body_pt_tol": { "type": "number", "minimum": 0 },
+    "ref_pt_min": { "type": "number", "exclusiveMinimum": 0 },
+    "ref_pt_max": { "type": "number", "exclusiveMinimum": 0 },
+    "fonts_text": { "type": "string", "minLength": 1 },
+    "fonts_title": { "type": "string", "minLength": 1 },
+    "page_limits_not_gated": { "type": "object" },
+    "kinds": {
+      "type": "object",
+      "additionalProperties": {
+        "type": "object",
+        "additionalProperties": false,
+        "properties": {
+          "body_pages_max": { "type": "integer", "minimum": 1 },
+          "ref_pages_max": { "type": "integer", "minimum": 0 }
+        }
+      }
+    }
+  },
+  "definitions": {
+    "ctanName": {
+      "description": "A package name as tlmgr knows it (CTAN / TeX Live), not a distribution package name.",
+      "type": "string",
+      "pattern": "^[a-z0-9][a-z0-9.-]*$"
+    },
+    "proofs": {
+      "type": "array",
+      "minItems": 1,
+      "uniqueItems": true,
+      "items": { "type": "string", "minLength": 1, "pattern": "^[^/\\s]+$" }
+    },
+    "tex": {
+      "description": "The TeX Live packages this template needs. Each package maps to the files that PROVE it is present: `kpsewhich` must find every one after an install, or the install failed and says which package.",
+      "type": "object",
+      "required": ["packages"],
+      "additionalProperties": false,
+      "properties": {
+        "packages": {
+          "type": "object",
+          "minProperties": 1,
+          "propertyNames": { "$ref": "#/definitions/ctanName" },
+          "additionalProperties": { "$ref": "#/definitions/proofs" }
+        },
+        "tools": {
+          "description": "Packages that ship a PROGRAM rather than a file LaTeX loads. Proved by an executable of that name in TeX Live's bin directory, not by kpsewhich.",
+          "type": "object",
+          "propertyNames": { "$ref": "#/definitions/ctanName" },
+          "additionalProperties": { "$ref": "#/definitions/proofs" }
+        }
+      }
+    }
+  }
+}

package/skills/submit-paper/submit-paper.eval.mjs

@@ -0,0 +1,19 @@
+/**
+ * submit-paper — the PAID tier: does this skill's description actually fire?
+ *
+ * COLOCATED ON PURPOSE (vigiles decides coverage by placement as of 2026-08-11).
+ * The prompts live in `.claude/lib/skill-trigger-cases.mjs` so all 21 cases
+ * are reviewed as one table where collisions between siblings are visible;
+ * copying them here would recreate the drift that rule exists to prevent.
+ *
+ * Measures recall (fires on its own territory) AND precision (stays quiet on a
+ * colliding sibling's territory), against the REAL `.claude` harness so the skill
+ * competes with every other installed description — an isolated run overstates
+ * recall and understates false positives.
+ *
+ * Costs money; not CI.
+ *   node .claude/skills/submit-paper/submit-paper.eval.mjs [trials]
+ */
+import { runSkillTriggerEval } from "../../lib/skill-eval-kit.mjs";
+
+await runSkillTriggerEval("submit-paper");

package/skills/submit-paper/submit-paper.harness.mjs

@@ -0,0 +1,23 @@
+/**
+ * submit-paper — the free, deterministic tier. No model, no network.
+ *
+ * COLOCATED ON PURPOSE. vigiles decides coverage by PLACEMENT as of 2026-08-11:
+ * a test that merely names a surface no longer counts, because that tier was
+ * crediting surfaces nothing touched. So each skill needs a file inside its own
+ * directory — this one.
+ *
+ * The assertions live in `.claude/lib/skill-checks.mjs` and are CALLED here with
+ * this skill's name. They are not copied: 22 copies of the same checks is the drift that
+ * module exists to avoid. (Until 2026-08-11 this was an env-var side channel into a
+ * 614-line file named after no surface; it is a function call now.)
+ *
+ * What this proves: this skill's frontmatter parses as strict YAML, its declared
+ * tool contract is sane, its pipeline wiring points at scripts that exist, and it
+ * announces/records under ITS OWN identity rather than a sibling's.
+ *
+ * What it does NOT prove: that the skill fires, or that its guidance produces a
+ * good result. Those need a real model — see `submit-paper.eval.mjs`.
+ */
+import { checkSkill } from "../../lib/skill-checks.mjs";
+
+await checkSkill("submit-paper");

package/skills/sweep-design-space/SKILL.md

@@ -0,0 +1,269 @@
+---
+name: sweep-design-space
+description: Find a real MECHANISM for a problem — sweep the design space across type systems, formal verification, functional programming, capabilities, specification-by-example, data integrity, authoring UX, other professions (aviation/law/audit), and historical ancestors in SE — instead of circling the same two or three obvious options. Use when the deliverable needs a solution rather than a measurement, when the last few proposals all rhymed with each other, or when the user says "your options are stale / you're thinking narrowly / think fresh". Encodes the hard filters that kill most candidates (does it re-require the thing already measured as unworkable? is it occupied? who writes the formalization? does it demand a new format? unrepresentable or merely detectable?) and the ladder pattern for when one mechanism can't cover the range. Distinct from research-ideate (validates ONE idea you already have), map-prior-work (finds competitors for a claim), and analyze-sibling-paper (deep-reads one rival) — this one GENERATES candidate mechanisms and filters them.
+context: fork
+allowed-tools: [WebSearch, WebFetch, Read, Write, Grep, Glob, Bash, Agent]
+---
+
+<!-- vigiles:sha256:1822c8c01e352f42 compiled from skills/sweep-design-space/SKILL.md.spec.ts -->
+
+# sweep-design-space — stop circling, sweep
+
+## Run me
+
+🔴 FIRST, before any other step:
+
+```
+node .claude/skills/paper-pipeline/scripts/announce.mjs sweep-design-space <dir>
+```
+
+An advisory pass cannot be observed failing — silence is both its error state and its normal
+state — so starting is an event, and events get written down.
+
+The `<dir>` argument is **optional in spirit**: this skill can run before a paper directory exists, so
+pass whichever directory the work lives in, or `.`. The ledger row is written either way, and a plain
+dot is more honest than an invented path.
+
+A model asked for "a solution" returns the three designs nearest its training prior, then variations of
+those, then variations of the variations. The user experiences this as *"you are thinking very narrowly"* —
+and they are right. The fix is not to think harder in place; it is to **sweep named source domains on
+purpose** and filter hard.
+
+Grounding lesson (2026-07-29, `the reference paper`): four consecutive proposals were reformulations of
+one idea ("check the prose"), and two of them turned out to be **already shipped in production** by
+other people. The design space was never actually swept; it was sampled from memory.
+
+## When it fires
+
+- The task needs a **mechanism**, not a measurement, and the mechanism isn't obvious.
+- Your last two or three proposals rhyme with each other.
+- The user pushes back with *"your options are stale / you're seeing only old solutions / think fresh"*.
+- A measurement phase just finished and the paper/product needs a constructive half.
+- 🔴 **You are about to recommend the first design you thought of.** That alone is the trigger.
+
+## How to run it
+
+### 1. State the defect as a STATE, not as a wish
+Not "we want rules to be enforced". Write the exact state that must not exist:
+> *a rule asserts an enforcement that nothing provides.*
+
+Everything downstream is judged by whether it makes **that state** impossible.
+
+### 2. Write the hard filters BEFORE generating
+Filters written after generation get bent to save a favourite. Standard ones — keep the ones that apply:
+
+- **🔴 Does it re-require the thing you already measured as unworkable?** The sharpest filter and the
+  easiest to violate. In the grounding case, prose-to-checker synthesis was measured leaking at
+  **84–96%** — and half the proposed designs still began with *"parse the prose"*. If your evidence
+  says X doesn't work, a design that starts with X is dead **regardless of how elegant the rest is**.
+- **Unrepresentable or merely detectable?** "We check for it afterwards" is a different (weaker) claim
+  than "it cannot occur". Say which one honestly.
+- **Who produces the formalization — a model or a human?** This, not "is a model involved", is what
+  predicts leakage. Model-synthesized formalization leaks; human-written formalization does not leak,
+  it is merely expensive. Sort candidates by this.
+- **Adoption cost.** Does it require a new format, a migration, or a new file? Count the artifacts
+  already in the wild that would have to move. A design requiring migration of a large installed base
+  has near-zero adoption regardless of merit.
+- **Scope honesty.** Which fraction of the real population does it cover, and does it *silently* cover
+  nothing for the rest? Force each candidate to say "for class X it is silent."
+
+### 3. Sweep the catalog — the point of this skill
+Take the groups below. For each, generate **3 candidates whose MECHANISMS differ** (three phrasings of
+one idea is a failed group). Delegate to a subagent with **no thread context** — the thread is exactly
+what anchors you — and hand it: the defect state, the hard filters, the measured constraints, and the
+**already-considered list**.
+
+🔴 **Hand that list as "considered, and here is exactly why each was set aside" — never as a bare
+ban.** A blanket prohibition is the wrong instrument, because some of those options were set aside
+for a *bad* reason, and a *modification* of a rejected option is often the best thing in the sweep.
+So the instruction to the agent is two-sided:
+
+- returning a considered option **unchanged** is a failure — that is the laziness the list exists to
+  prevent;
+- returning it **because you can attack the stated reason it was rejected**, or with a modification
+  that survives that reason, is a **first-class result** — say so explicitly in the brief.
+
+This shape does something a ban cannot: it surfaces bad rejections. In the grounding session, one
+option was discarded with the objection *"the generated file can still be hand-edited"* — an objection
+already solved, years earlier, by a checksum on the generated section. Under a bare ban that option
+stays dead forever, killed by a wrong reason nobody re-examines. **Record the reason next to every
+retired option precisely so it can be attacked later.**
+
+### 4. Occupancy check BEFORE you fall in love
+Every surviving candidate goes through a refutation-framed occupancy pass (*"find who already did
+this"*, not *"confirm it's free"*). In the grounding case this killed the front-runner: the
+`enforced-by:` tag design was **already shipped** — CI blocking PRs, reference resolution in Rust,
+across several independent projects. "I don't know of such work" ≠ "no such work". Search GitHub code
+search, not only papers: practice ships years before it is written up.
+
+### 5. Red-team the survivor
+Separate subagent, framed to **break** it. See the paper corpus's own refutation rules (`CLAUDE.md` beside the papers).
+
+### 6. Save to the repo
+A swept design space that lives only in a chat is re-swept from scratch next month. Write the map —
+axes, cells, occupancy verdicts, what kills each — as a file, and record **which candidates were
+retired and why**, so they are not proposed again.
+
+---
+
+## The catalog — source domains, with constructs named
+
+Do not treat this as exhaustive; it is a floor, so the agent starts from named constructs instead of
+inventing the list each time. Add groups when the problem calls for them.
+
+### A. Type systems and language design
+Smart constructors · newtype and phantom types · GADTs · **typestate** · linear and affine types
+(a permission that is *consumed*) · refinement types · totality checking · effects in the type ·
+**parse-don't-validate** · make illegal states unrepresentable · instance resolution as a gate
+(the operation *doesn't exist* when there is no backend for it).
+
+### B. Formal verification
+Coq · Lean · **TLA+ and Alloy** (model the *process*, not the code) · model checking · SMT ·
+**satisfiability as a meaningfulness test** (a constraint that removes no models is vacuous — a
+mechanical vacuity check) · refinement (does the implementation refine the spec) · proof-carrying
+artifacts · Design-by-Contract · **blame** (Findler & Felleisen — who is at fault at the boundary).
+
+### C. Functional programming
+Free monads and DSL-as-data (one AST, many interpreters) · tagless-final · **session types**
+(protocols as types — the natural home for "order of operations" rules) · algebraic effects ·
+optics · **monoids and folds** (how do rules compose, merge, override?) · lattices and semilattices
+(CRDT-style merge without conflict) · catamorphisms · **Curry–Howard** (rule as type, proof as value) ·
+property laws · denotational semantics (a rule *is* the set it forbids; empty set = vacuous).
+
+### D. Capabilities and systems
+Capability-based security · **reference monitor with complete mediation** · principle of least
+authority · object capabilities · **grants instead of prohibitions** (if the file can only *grant*,
+"don't do X" has no syntax) · sandboxing · seccomp/eBPF · information-flow control · taint.
+
+### E. Testing and specification by example
+Property-based testing · metamorphic relations · golden/approval tests · **mutation testing as an
+admission criterion** (a rule that kills no mutant is not admitted) · executable specifications ·
+example pairs as the artifact with prose as caption (machine never reads the prose) ·
+**A/B measurement of effect** (treat the artifact as opaque and measure whether it changes outcomes —
+the one way to take unparseable prose seriously without parsing it).
+
+### F. Data integrity and databases
+Constraints · **`ADD CONSTRAINT` validates existing rows** (you cannot add a constraint your current
+data violates — near-universal in databases, near-absent everywhere else) · `NOT NULL` + foreign keys ·
+schema and migrations · `NOT VALID` with a deadline · provenance and lineage · event sourcing ·
+**store the thing as data with the human-readable file as a view**.
+
+### G. Human–system interaction (authoring)
+Asymmetric cost (make the strong claim expensive and the weak one one click) · forced choice with no
+escape hatch · **changing the default** rather than adding a gate · feedback at the moment of writing
+(live "this would have matched: 0") · progressive disclosure · single entry point (the artifact can
+only be created *from* an incident, never from imagination) · what code review, commit hooks and IDE
+hints actually taught us about compliance.
+
+### H. Other professions — how they make an unenforceable norm unwritable
+- **Aviation:** Minimum Equipment List — a defect is permitted only *with* a compensating procedure
+  and a **repair deadline** by category; expiry is automatic. Also checklists, and why they work.
+- **Law:** *lex imperfecta* — an obligation with no sanction is a recognized category, not an
+  oversight. **Rules vs standards** (Kaplow 1992) — the ex-ante/ex-post distinction, half a century old.
+  Adjudicator: a standard must name who decides.
+- **Audit/finance:** SOX controls — owner + frequency + **evidence artifact**, and two separate tests
+  (*design* and *operating effectiveness*). "The control exists" stops being a state; there are three.
+- **Medicine:** protocols, contraindications, checklists, and the distinction between guideline and
+  protocol.
+- **Safety standards:** MISRA (**decidable/undecidable** is an official per-rule label), DO-178C DAL
+  A–E, ISO 26262 ASIL, Common Criteria EAL — industries that already formalized *how strongly backed*
+  something is.
+- **Construction codes, military ROE, accounting standards** — same question, different centuries.
+
+### I. 🔴 Historical ancestors in software engineering
+**Assume the problem is old.** In the grounding case, a "novel" controlled-language idea turned out to
+be a 1980s aerospace standard, and its ladder had been published in *Computational Linguistics* in 2014.
+Look here **before** generating anything:
+
+- **Gradual typing** — the canonical template for coexisting checked and unchecked in one artifact,
+  with an explicit boundary and blame, adopted across a huge installed base *because* it did not
+  require migration. If your problem is "some items are backed and some are not, and we cannot move
+  everyone", this is the closest ancestor that exists.
+- **Gradual verification**, gradual refinement types.
+- **Design by Contract** (Eiffel) and contract blame (Racket).
+- **JML · SPARK Ada · ACSL · Dafny** — and specifically **SPARK's stone → bronze → silver → gold →
+  platinum**: a shipped, mechanically-defined ladder of assurance. Steal the shape.
+- **Aspect-oriented programming** — cross-cutting requirements declared away from the code, and
+  **why it failed** (the failure is more instructive than the successes).
+- **lint (Johnson, 1978)** — the original answer to "a rule the compiler does not check".
+- **Assertions** (Floyd, Hoare) and how they became normal.
+- **Literate programming**, doctest, executable documentation.
+- **Requirements engineering** — *testable requirement* as a recognized category, traceability,
+  EARS, IEEE 830.
+
+### J. Migration and coexistence strategies
+Not a mechanism but a shape most candidates need: strangler fig · feature flags · `NOT VALID`
+constraints · gradual typing's boundary · deprecation with deadlines · advisory-then-enforcing rollout
+(how linters actually ship a new rule) · opt-in tiers.
+
+---
+
+## The ladder pattern
+
+When no single mechanism covers the range — the usual case — the answer is often a **graded scale**
+rather than one gate. A ladder is only worth anything if every rung is:
+
+1. **decided mechanically** — no reading, no judgment, no model interpreting anything;
+2. **strictly stronger** than the rung below;
+3. **precedented** — name the ancestor (SPARK levels, DAL, MISRA decidable/undecidable, gradual typing's
+   typed/untyped boundary). A ladder with invented rungs is a taxonomy, and taxonomies are usually
+   occupied.
+
+🔴 And check whether the *ladder itself* is occupied for your object. Controlled-language scales and
+assurance levels already exist; the question is whether anyone has carried them to your problem.
+
+---
+
+## Record the verdict
+
+🔴 LAST step, once the deliverable exists:
+
+```
+node .claude/skills/paper-pipeline/scripts/ledger.mjs record sweep-design-space <dir> FINDING <count> <report-path>
+node .claude/skills/paper-pipeline/scripts/ledger.mjs record sweep-design-space <dir> ABSTAINED <reason> "<one line>"
+```
+
+🔴 **There is no PASS.** A sweep that turns up survivors has not certified anything; it has found
+candidates, and candidates are findings.
+
+**FINDING** — the sweep produced something to act on; `<count>` is the number of survivors and
+`<report-path>` the sweep, including the groups that yielded nothing. Add `--blocking` when the
+whole sweep died on the filters — that is the loudest thing this skill can say.
+**ABSTAINED** — `no-witness`: the sweep ran and neither found a survivor nor established that none
+exists. `blocked`: there was no problem statement to sweep against.
+
+The skill's own rules already say negative results count. The ledger is how that survives the session:
+a recorded finding names the groups that were swept, so the next sweep starts from named constructs
+instead of from memory — which is the failure mode this skill exists to break.
+
+## Rules
+
+- **Name the forbidden options in the generation brief.** Without an explicit "do not return these"
+  list, a fresh agent returns your stale options with new names.
+- **Generate with a context-free subagent.** The thread is the anchor you are trying to escape.
+- **Three ideas per group must differ in MECHANISM**, not in phrasing.
+- **Apply the "does it re-require the unworkable thing" filter first** — it is the one most often
+  violated, including by designs that feel elegant.
+- **Occupancy before affection.** Search shipped code, not only literature.
+- **Negative results count.** "This group yields nothing here" is a real finding; record it so the
+  group is not re-swept.
+- **Prefer designs that work without anything understanding anything** — counting, required fields,
+  foreign keys, expiry dates, and grants do not leak. Interpretation does.
+- **Say what each candidate is silent about.** A design that covers a quarter of the population and
+  says nothing about the rest is fine — undeclared silence is not.
+
+## Compose with
+- `research-ideate` — validates ONE idea's research value; run after this picks a candidate.
+- `map-prior-work` / `analyze-sibling-paper` — occupancy at the level of a paper's contribution.
+- `build-benchmark` — once the mechanism exists, this measures it.
+- `paper-adversarial-review` / `pc-panel-review` — red-team the survivor.
+
+## Provenance
+Built 2026-07-29 during `the reference paper`, after a solution search stalled for hours. Four
+consecutive proposals were the same idea reworded; when finally swept properly, a context-free agent
+produced 24 candidates across 8 groups, of which the three least obvious came from **capabilities**
+(the rule's text exists only as the monitor's refusal message — no prose is stored at all),
+**databases** (`ADD CONSTRAINT` refuses when current data already violates), and
+**specification-by-example** (measure the artifact's *effect* and never read it). None of the three
+was reachable from the thread, and two of the thread's own front-runners turned out to be shipped
+already. The catalog exists so the next sweep starts from named constructs instead of from memory.