Academic Paper

---
name: academic-paper
description: "Use when turning a sparq contribution into an academic paper — identifying a genuinely-novel contribution (the re-runnable intake), classifying it and picking a venue (ISWC/ESWC Resource for the engine, PVLDB/SIGMOD/EDBT for DB-systems, arXiv/workshop for not-yet-sound ZK/MPC), drafting a single-source Typst (.typ) paper whose eval section binds to live benchmark data via --input/sys.inputs, building both a PDF and an in-site HTML page from that one source, and reviewing claims↔evidence under sparq's empirical-honesty mandate (canonical vs indicative numbers; the ZK/MPC not-yet-sound disclaimer). The paper-factory for epic sq-gum8."
---

# academic-paper — the sparq paper-factory

This is the repeatable PROCESS for producing one paper from a sparq contribution. It is
**factory machinery**, not prose advice: it enforces sparq's empirical-honesty mandate at
the *data* layer, binds every number to live benchmark data, and emits a PDF + an in-site
HTML page from one Typst source. Design record: `research/paper-factory-design.md`
(consolidates the contribution inventory + the auto-gen/PDF research). Epic **sq-gum8**.

**Two non-negotiable constraints govern every step.** They are the point of the factory; do
not soften them.

1. **No ZK/MPC security property may be claimed as proven.** The single-prover ZK verifier
   has no external audit (bead **sq-qhy4**, open); the multi-prover path is re-open
   (**sq-9hrn**); the malicious-secure MPC layer is a stub. A ZK/MPC paper today is an <!-- privacy-claims-allow: "malicious-secure" here is a negation ("is a stub"), not an achieved-property claim -->
   **honest design / limitations / negative-result** contribution, arXiv/WIP only, and
   **must cite sq-qhy4**. Never write "secure" / "verifiable" / "private" as a *proven*
   property — only as a *design goal*.
2. **Work-box (EC2 `-aws`) wall-clock numbers are NON-CANONICAL.** Only deterministic
   integer metrics are canonical today (W3C/OGC conformance floors, byte-identity, recall
   floors, gate/round/byte counts, differential-fuzz pass). A speed/memory claim needs the
   **canonical runner** (`research/ci-ec2-design.md`). Indicative numbers are **never
   co-tabulated with canonical ones** and never feed an aggregate figure-of-merit.

The factory has six stages. Stages 0/1/3/5 are the methodology this skill encodes; 2/4/6 are
scripted (see `research/paper-factory-design.md` §1–§2 for the script/route layout).

---

## Stage 0 — Intake: identify a genuinely-novel contribution (re-runnable)

Run before claiming anything is publishable. This is the re-runnable identification process;
its current ranked output is `research/paper-contributions-inventory.md` — read it first and
diff against it rather than starting from scratch.

**Scan six sources:** crates (`crates/sparq-*/src`, READMEs — what is IMPLEMENTED vs
designed-only; `NotYetImplemented`/`#[ignore]` markers); `research/*.md` (claimed technique +
prior-art + the author's own honesty hedges + negative results); beads
(`.beads/issues.jsonl` — open audit/gating beads, esp. sq-qhy4/sq-9hrn/sq-1gir);
benchmark deltas (`research/BENCHMARKS.md`, `bench/` — which numbers exist, canonical vs
work-box); the conformance scoreboard (`crates/sparq-conformance/src/scoreboard.rs` — the
only fully-canonical evidence); skills (`skills/*/SKILL.md` — distilled durable findings).

**Score four criteria — a candidate must pass all four:**

1. **Novelty vs prior art** — a new algorithm/construction/finding, or a faithful
   re-implementation of cited prior art? Name the prior systems (ACORN, NaviX, QLever,
   RDFox, CostFed, SPDZ, coZK, …). Integration *can* be a contribution — but framed as a
   *systems/integration* paper, never dressed as a new algorithm. A measured negative result
   is also a contribution.
2. **Evidence available** — canonical evidence *today*, or only work-box numbers / a design?
3. **Generality** — does the claim generalise beyond sparq's exact code, or is it a local
   tuning artefact?
4. **Honesty / soundness** — for crypto: proven, or merely implemented + internally
   reviewed? For perf: canonical or not? **This is the veto against overclaiming.**

**Assign a readiness verdict:** PUBLISHABLE-NOW (sound claim + canonical evidence today, or a
design/methodology contribution needing no benchmark) | NEEDS-CANONICAL-BENCHMARKS (honest +
real, but the headline rests on work-box numbers — publishable once re-gathered on the
canonical runner) | NOT-YET-SOUND / NEEDS-EXTERNAL-AUDIT (ZK/MPC — honest design / negative
result only; cite sq-qhy4). **Rank by readiness × impact** (a PUBLISHABLE-NOW candidate with
a real audience outranks a high-impact unaudited crypto claim).

**Re-run triggers:** a new crate/feature lands; a benchmark improves or a canonical run
completes (re-evaluate every NEEDS-CANONICAL-BENCHMARKS candidate); a conformance floor rises;
an audit bead changes state (a sq-qhy4 *pass* moves the ZK candidate out of NOT-YET-SOUND —
until then it cannot move); a new `*-measured.md`/negative-result doc appears. The re-run is
cheap: re-scan, re-score, re-verdict, re-rank, diff the inventory, open/close paper beads.

---

## Stage 1 — Classify & venue-target

Map the contribution to its family, then to a venue + track + Typst template. **Verify every
page limit / deadline / blinding policy against the current-year CFP before submitting — they
drift yearly.** Maintainer precedent (jeswr, Oxford / ODI / W3C) centres on **ISWC/ESWC**;
the engine-as-artifact maps exactly to their **Resource track**.

| Family | Best venues | Track / notes |
| --- | --- | --- |
| **A — DB/engine performance** (indexes, parallel/streaming/mmap exec, ingest, compression) | **PVLDB** (rolling monthly, single-blind, EA&B track for measurement-heavy papers); **SIGMOD** (fixed rounds, double-blind); **ICDE**; **EDBT** (most accessible first DB paper) | systems/empirical; reproducibility/artifact tracks near-universal |
| **B — Semantic Web / RDF / SPARQL** (federation, SHACL, GeoSPARQL, RSP-QL, inference, GenAI-over-RDF, RDFC-1.0) | **ISWC** (home venue; **Resource track purpose-built for a reusable engine**); **ESWC** (European sibling; note its strict no-preprint window conflicts with arXiv-first); TheWebConf; SEMANTiCS; JWS (journal) | Research / **Resource** / In-Use + posters/demos/workshops |
| **C — Security/privacy (ZK/MPC), NOT-yet-sound** | **arXiv preprint + workshop NOW** (HotPETs / ZKProof / WPES / ISWC-ESWC privacy workshops); **PoPETs** = best real target *once sound* (journal, 4 rolling deadlines/yr); USENIX/CCS/S&P aspirational | preprint/WIP today; graduate to a real venue only when sq-qhy4 (and sq-9hrn) close |

**Template:** `charged-ieee` (IEEE 2-col) | `arkheion` (arXiv-style) | `para-lipics` (LIPIcs;
accepted papers need a Typst→LaTeX conversion for the publisher) | an acmart-style template.
**Double-blind venues** (SIGMOD/VLDB/ICDE/WWW/CCS/S&P) require the anonymized build (Stage 4);
ISWC/ESWC blinding varies by track. **Rolling cadence** (PVLDB monthly, PoPETs quarterly,
SIGMOD multi-round) → target submit-when-ready, not one annual crunch.

---

## Stage 2 — Evidence capture (the honesty boundary)

Paper-bound numbers live in a **dedicated evidence file, `site/src/data/paper-evidence.json`**
— distinct from `site/src/data/benchmarks.generated.json`. The two MUST NOT be conflated:

- **`paper-evidence.json`** — the *canonical-only* paper evidence. Every record is a
  deterministic, machine-INDEPENDENT metric (a recall floor, an answer-safety invariant, a
  cost-model crossover) lifted from a named test, so each is `environment: "canonical"` and
  carries a `source` field tracing it to that test/dataset. [OPUS-4.8: as-built, PR #336.]
- **`benchmarks.generated.json`** — per-commit *timing* on the dev work-box, **all
  `environment: "indicative"`** (folded in by `site/scripts/sync-benchmarks.mjs` from the
  `benchmark-data` branch). It feeds the site's benchmark widgets, **never** a paper headline.

A paper-evidence record (a `binding` upgrades its `source` from asserted to machine-verified —
see gate 4):

```json
{ "value": 98, "unit": "passing-assertions", "environment": "canonical",
  "kind": "conformance-ratchet-floor",
  "source": "crates/sparq-shacl/tests/w3c_core.rs::BASELINE_PASS (mirrored in scoreboard.rs)",
  "binding": { "kind": "rust-anchor", "file": "crates/sparq-shacl/tests/w3c_core.rs", "anchor": "BASELINE_PASS" },
  "note": "machine-independent: a const a CI test asserts, not a timing." }
```

**The honesty gate (as-built; PR #336 landed layers 1–3, sq-gum8.13 added the binding layer 4):**

1. **Data layer — `site/scripts/build-papers.mjs::runHonestyGate()` runs FIRST**, before any
   compile: it schema-checks `paper-evidence.json` (every record needs a valid `environment`
   ∈ {canonical, indicative}, a `source`, and a `value`) so a malformed/untraceable record is
   a clear early build failure, not a Typst stack trace.
2. **Compile layer — `headline(key)` in `site/papers/_lib/bench.typ` panics the Typst
   compile** if a record's `environment != "canonical"`. `headline()` is the ONLY accessor
   allowed inside a headline result table/figure; the ungated `ev(key)` is for
   clearly-labelled *indicative* callouts only.

So any non-canonical (work-box / indicative) number can never appear as a paper's headline
evidence. Indicative and canonical numbers are never in the same table and never feed an
aggregate. Until the canonical *latency* runner exists (`research/ci-ec2-design.md`, blocked
on one IAM step), publish **only the deterministic metrics** (conformance counts, recall
floors, byte-identity / answer-safety invariants, gate/round/byte counts) — those are
canonical today and are exactly what `paper-evidence.json` holds.

3. **Coarse phrase/number gate over the paper surface — `check-no-perf-numbers.py` +
   `check-privacy-claims.sh`** (beads sq-mkza / sq-4hga / sq-mraf). Both CI gates also scan
   the paper `.typ` sources **and** the prose (`note`/free-text) fields of
   `paper-evidence.json`: a result-shaped number typed inline (not routed through the
   accessor) trips the perf gate (accessor-call spans + Typst comments are skipped, so the
   canonical data path is unflaggable); an unqualified ZK/MPC soundness/privacy phrase trips
   the privacy gate (with the same `privacy-claims-allow: <why>` inline marker + same-line
   negator/hedge exemption). They share **one** forbidden-phrase list
   (`scripts/honesty-phrases.json`), and `build-papers.mjs` re-runs both at the build
   boundary. This is a **COARSE** class only — see Stage 5 for what it does and does **not**
   catch; do not read its green as a semantic-correctness guarantee.
4. **Value↔source binding gate — `scripts/verify-paper-evidence.py`** (bead sq-gum8.13, F1).
   Layer 1 checks a `source` is *present*; this layer checks the value still *matches* it.
   Each canonical record carries a machine-checkable `binding`, one of:
   - `json-pointer` `{file, pointer}` — the value must EQUAL the RFC-6901-pointed scalar in a
     committed JSON (derivation-strength);
   - `rust-anchor` `{file, anchor}` — the anchor (const/test-fn name) must EXIST in the file AND
     the value must appear within its window (**honest limitation: literal-adjacency only** —
     it can in principle false-pass if the same literal sits nearby for another reason; strictly
     weaker than json-pointer, strictly stronger than unchecked free text);
   - `doc-anchor` `{file, quote}` — the exact quote must be present (existence-strength only).

   **Fail-closed:** a canonical record without a passing binding aborts the paper build unless
   it is on the shrink-only allowlist (`scripts/paper-evidence-binding-allowlist.json`); the
   verifier fails if that allowlist ever GROWS (F3/sq-gum8.15 drives it to empty; F2/sq-gum8.14
   exports the scoreboard consts so the conformance floors become `json-pointer`s). Wired into
   `build-papers.mjs`; `--self-test` is the CI gate (a value-DRIFT fixture and a missing-anchor
   fixture each exit non-zero). Like gate 3 it is **MECHANICAL** — value↔source match only; a
   *semantic* overclaim (a true number framed misleadingly) stays the Stage-5 human review.

---

## Stage 3 — Draft: single-source Typst, bound to live data

**Writing methodology (the order matters):**

- **Contributions list FIRST** — it drives the whole paper ("the paper substantiates the
  claims"). Each bullet is **refutable** (names *what* is achieved, specific enough to be
  disproved) and **forward-references the section that delivers the evidence**. PJ's
  contrast: NO = "We describe the WizWoz system. It is really cool."; YES = "We give the
  syntax and semantics … (§3); we prove the type system sound … (§4); … half the length of
  the Java version (§5)."
- **Abstract = 4 sentences, written last:** (1) the problem; (2) why it is interesting;
  (3) what your solution achieves; (4) what follows.
- **Introduction ≤ 1 page**, in order: what is the problem / why important / why hard (why
  naive approaches fail) / why unsolved before (what distinguishes you) / key components +
  results → ending in the bulleted contributions list. The new contribution must be clear by
  **end of page 3**. Delete "the rest of this paper is organized as follows" — the
  contributions list does that job.
- **Real examples, never `foo`/`bar`; active voice; self-contained figure captions that tell
  the reader what to notice** (reviewers skim).
- **Related work late & charitable** — "credit is not like money": be generous; compare and
  contrast, don't list; label any inferior approach explicitly and up front; write it "as if
  telling the cited authors why they should care." Defer to near the end unless a concise
  early defensive stance is needed.
- **Conclusion** with concrete numbers; **no wishlist "future work"** ("no partial credit for
  neat things you wanted to do but didn't").

**The live-data recipe (the load-bearing factory mechanism).** Author one `.typ`; the build
injects `paper-evidence.json` via `--input data=...`. Numbers are read through the shared
`site/papers/_lib/bench.typ` helpers, **never** hard-coded:

```typ
#import "_lib/bench.typ": headline, ev, provenance, authors, anon
```

`bench.typ` binds `#let evidence = json(bytes(sys.inputs.data))` once and exposes the
accessors. The eval section uses `#headline(key)` for any headline table/figure (it panics
the compile on a non-canonical record — the honesty gate), and the ungated `#ev(key)` only
inside an explicitly-labelled *indicative* callout:

```typ
The filtered top-k matches the exact-filtered ground truth at recall@10
>= #headline("filtered_ann.recall_at_10_floor").
#provenance("filtered_ann.recall_at_10_floor")
```

Run the self-check before handing off to review:

- **SIGPLAN-7 rubric** (red/yellow/green, supports judgment not a binary gate): (1) claims
  clearly stated + scoped, no implied generality; (2) suitable, fairly-configured baseline;
  (3) principled benchmark choice (established suites, justify subsets, no cherry-picking,
  don't test on the training set); (4) adequate analysis (enough trials; geometric mean for
  ratios, harmonic for rates, median under outliers; report variability/CIs); (5) relevant
  metrics (measure *all* important effects — index/compile time alongside runtime); (6) clear
  reproducible design (full platform spec, key parameters explored); (7) appropriate
  presentation (zero-based axes, right precision, distribution-reflecting summary).
- **Benchmarking-crimes self-check** (Heiser): full HW spec (CPU+µarch, cores, clock/turbo
  policy, cache levels, RAM, kernel release, storage); full SW spec (`rustc` + `Cargo.lock` +
  flags, baseline-system versions); dataset name+version+SHA-256+seed; report absolutes not
  ratios-only; never summarize 4/6/7/49% as "up to 49%"; never use a competitor's sub-optimal
  config (Heiser: that "probably constitutes scientific misconduct").

---

## Stage 4 — Build: one source → PDF + in-site HTML

`site/scripts/build-papers.mjs` (wired into `prebuild` + `dev`) drives both artifacts for
every paper in `papers.ts`, injecting the SAME `paper-evidence.json` so they cannot diverge.
The two compile invocations it runs (paths relative to `site/`, `--root site`):

```bash
# PDF (the download — static asset under public/papers/, served by the Next.js export):
typst compile papers/<source>.typ public/papers/<slug>.pdf \
  --root . --input data="$(cat src/data/paper-evidence.json)"

# In-site HTML — Typst NATIVE HTML export (NOT typst.ts):
typst compile papers/<source>.typ src/generated/papers/<slug>.html \
  --root . --format html --features html --input data="$(cat src/data/paper-evidence.json)"

# anonymized build for a double-blind venue (the .typ's authors()/anon honour this):
typst compile papers/<source>.typ /tmp/<slug>-anon.pdf \
  --root . --input data="$(cat src/data/paper-evidence.json)" --input anon=true
```

[OPUS-4.8: as-built, PR #336.] The **in-site HTML** page at `/papers/<slug>` uses **Typst's
native HTML export** — *not* typst.ts/`@myriaddreamin/typst.react`. The build extracts the
`<body>` inner HTML and the React route (`components/papers/paper-html.tsx`) injects it as a
static fragment into a scoped `.paper-prose` block (no WASM compiler shipped to the browser).
The PDF and the HTML are built from the *same* `.typ` + the *same* injected evidence, so the
numbers can't diverge. **Trade-off:** native HTML export is experimental, so **layout-only
constructs (centring, alignment, page rules / horizontal rules) drop in the HTML view** — the
expected `--features html` "experimental feature" + page-rule warnings on stderr are harmless;
those constructs are **preserved in the PDF**. Author for both: rely on semantic structure
(headings, tables, paragraphs) for meaning, treat alignment as PDF-only polish. If Typst is
not on `PATH`, the script emits an honest placeholder fragment and warns (CI installs Typst so
real artifacts always build). A `benchmark-data`/evidence change re-triggers `next build` and
both artifacts regenerate. For a venue that rejects Typst→LaTeX, the camera-ready fallback is
Pandoc/LaTeX via Tectonic against `acmart`/`IEEEtran`/`lipics`. **Do not** author the PDF with
`@react-pdf/renderer` — it duplicates the numbers (breaks single-source).

---

## Stage 5 — Review gate: claims ↔ evidence

For every claim in the intro/contributions list, identify its evidence (analysis / theorem /
measurement / case study) and confirm the forward-reference resolves. The gate (run as
subagent section-reviewers + one cross-cutting honesty/repro reviewer) **blocks** on:

- **Any indicative number in a claim** (the Stage-2 build-time honesty gate —
  `build-papers.mjs` schema-check + the `headline()` compile panic — enforces this
  mechanically; the reviewer also checks no indicative/canonical co-tabulation).
- **A C-family (ZK/MPC) draft using "secure"/"verifiable"/"private" as a proven property** —
  must be a design goal, with the sq-qhy4 soundness-gap disclaimer present.
- **Overclaiming / implied generality, weak/unfair baselines, missing ablations, no error
  bars, irreproducibility** (the SIGPLAN-7 + benchmarking-crimes findings from Stage 3).

**What is — and is NOT — mechanically gated (do not over-sell the gate).** [OPUS-4.8] As of
beads sq-mkza / sq-4hga / sq-mraf, the two CI honesty gates also scan the paper-factory's own
surface: `check-no-perf-numbers.py` scans the paper `.typ` sources (accessor-aware, result-
shaped numbers only) **and** the prose (`note`/free-text) fields of `paper-evidence.json`;
`check-privacy-claims.sh` scans those `.typ` + the evidence JSON for the absolute forbidden
ZK/MPC phrases. Both consume **one shared forbidden-phrase list**
(`scripts/honesty-phrases.json`), and `build-papers.mjs` re-runs both at the build boundary
so the factory cannot serve an un-scanned paper (`scripts/tests/test_typ_honesty_gates.sh` +
`scripts/tests/test_build_papers_honesty_boundary.sh` pin this). **But this is a COARSE,
mechanical class only** — a hard-coded result-shaped number, or one of a fixed set of
unqualified soundness/privacy phrases. A **subtle semantic overclaim** — an unsupported
"generalises to all RDF stores", an implied causal/superiority claim phrased without a
forbidden phrase or a result-unit, an unfair-baseline framing — is **NOT** caught by any
gate and **remains this Stage-5 human/subagent review's responsibility**. Treating the gate's
green as a soundness guarantee would itself be an empirical-honesty violation.

Resolve **all** findings before publish. Final check: the claims↔evidence loop is closed and
every headline number is canonical.

---

## Stage 6 — Publish & auto-update

Merge → `next build` serves `/papers/<slug>` + `/papers/<slug>.pdf` (static export, under the
`/sparq` basePath). A `paper-evidence.json` refresh re-runs the build → both artifacts
regenerate with current numbers. Each headline number carries its **provenance** inline (the
`provenance(key)` helper prints the record's `source` test + `environment`). A venue
camera-ready triggers the optional Typst→LaTeX export. To **register a new paper**, add an
entry to `site/src/data/papers.ts` (`slug` + `source`) and a `site/papers/<source>.typ` (e.g.
`filtered-ann.typ`) — the index, the per-paper route, the nav, and the PDF build
are all data-driven off `papers.ts`.

---

## Empirical-honesty rules (the rubric, condensed)

- **Canonical vs indicative** — headline tables/figures use only `environment: canonical`
  records, reported with variability. Indicative (EC2 `-aws`) numbers live only in clearly
  labelled "indicative development measurement" callouts that name the instance type + `-aws`
  kernel, state they are not the basis of any claim, and are never co-tabulated with canonical
  numbers. Never quote a speedup blending the two.
- **ZK/MPC not-yet-sound disclaimer** — every C-family paper cites **sq-qhy4**, is marked
  arXiv/WIP-only, and frames any security/privacy/integrity/attestation property as a *design
  goal*, never proven. It graduates to a real venue only when sq-qhy4 (and, for the
  multi-prover path, sq-9hrn) close.
- **Honesty norm** — be scrupulously honest; don't over-sell, hide drawbacks, or disparage
  others' work; submit only if proud to attach your name as-is; anticipate scepticism and
  state how the approach could fail.

## Pilot

The factory's first pilot is **A1 + A2** together (both PUBLISHABLE-NOW, no canonical runner
or external audit needed): **A1** = RDF-native filtered-ANN ("Filter-as-Query" — exact
transitive-BGP filter over the engine's own dict-ids + answer-safety; canonical recall
evidence today; frame as integration, not an ANN-algorithm novelty); **A2** = the honest
same-box benchmark methodology (a methods/reproducibility contribution that strengthens A1's
eval). See `research/paper-contributions-inventory.md` (Part 2) and
`research/paper-factory-design.md` (§6).

## Paper P2 instrument — the SPARQL logic-bug harness (`crates/sparq-metamorph`)

[FABLE-5] sq-gum8.6. Paper **P2** (draft bead `sq-gum8.7`) is the first dedicated
logic-bug testing paper for SPARQL engines (SQL/Datalog/Cypher are covered by
SQLancer/queryFuzz/GDsmith; SPARQL is not). Its instrument is the opt-in
`crates/sparq-metamorph` crate (`publish = false`; nothing in the shipping graph links
it): **TLP re-derived for SPARQL** — partitioning on `FILTER(c)` / `FILTER(!c)` /
`FILTER(COALESCE(IF(c, false, false), true))`, where the third branch reifies SPARQL's
evaluation-*error* outcome (type errors + unbound variables), which unlike SQL's `NULL`
is not a testable value — the spec-cited case analysis in the `tlp` module docs IS the
paper's core claim; **NoREC for SPARQL** (predicate moved to projection position);
a **differential oracle** reusing `sparq-difftest`'s engine-independent comparators;
a **seeded deterministic generator** (in-crate SplitMix64 — a ledger seed reproduces
its case exactly); SPARQL-protocol **drivers** (`protocol-drivers` cargo feature,
off by default) for Fuseki/Virtuoso/Oxigraph-class endpoints; and the **found-bug
ledger** (JSONL; an entry REQUIRES an upstream issue URL + developer-confirmation
status — the paper counts developer-confirmed bugs, never raw oracle alarms; the
wrong-result class is strictly separated from engine errors). The bug-hunting
campaign itself runs **off-CI** (work box / EC2; docker endpoints fine); CI runs the
network-free mutation self-tests (a seeded wrong-result mutant must be flagged by all
three oracles — non-vacuity). sparq is itself a campaign target: finding our own bugs
is honest and reportable.

## Notes

- This skill was **authored in-repo** (not installed). The strongest external pipeline
  (Imbad0202/academic-research-skills) is CC-BY-NC — incompatible with this permissively
  licensed repo; the MIT K-Dense scientific-writer is life-sciences-leaning and lacks
  sparq's venue map + non-canonical-benchmark handling + the live-data Typst pipeline. Those
  two sparq specifics ARE the contribution of the factory.
- This is a **site/process** surface, not a public crate API, so it is not in the
  `skills/SKILL.md` public-surface router (that router lists code-integration surfaces).
  Keep this skill in sync with `research/paper-factory-design.md` if the pipeline changes.
- Verify before relying on version-specific features: Typst's native HTML export is still
  experimental (the layout-only-drop caveat in Stage 4 — re-check each Typst release whether
  more layout survives the HTML view); the current-year CFP for any target venue.