paper

name	paper
description	Draft, review, revise, and audit research papers in LaTeX (compilable PDF) using the standard anvil lifecycle plus an optional read-only literature-search critic.
domain	paper
type	skill
user-invocable	false

anvil:paper — Research papers (LaTeX)

The paper skill produces research papers as LaTeX sources that compile to a PDF at READY state. It follows the canonical anvil lifecycle — draft → review → revise → audit → figures — with one paper-specific addition: an optional pre-draft literature-search critic (paper-litsearch) that the drafter and the reviser can consume. Litsearch is a sibling critic, not a phase that gates the state machine.

Papers carry sharper expectations than memos around evidence and citation correctness, so the audit phase (paper-audit) is a first-class part of the lifecycle (not optional), and the rubric weights rigor / evidence / citation hygiene heavily.

Artifact contract

A paper thread is a single research paper authored across one or more revisions, identified by a slug (e.g., q3-method, acme-bench-2026). Each thread occupies a portfolio directory:

<portfolio>/
  <thread>/                       Optional thread root with brief and reference material
    BRIEF.md                      Structured or freeform brief (frontmatter + prose); when absent,
                                  paper-draft bootstraps one via an interactive interview
                                  (non-interactive / --no-interview runs keep the fail-fast;
                                  see commands/paper-draft.md § BRIEF bootstrap interview)
    refs/                         Optional reference material (datasets, prior drafts, transcripts)
    refs.bib                      Optional starter bibliography supplied by the author
  <thread>.0.litsearch/           Optional pre-draft literature-search sibling (read-only)
    notes.md                      Annotated discussion of related work + positioning
    candidates.bib                Candidate BibTeX entries the drafter may merge into refs.bib
    _progress.json                Phase state (litsearch)
  <thread>.1/                     First drafted version (immutable once written)
    main.tex                      Paper body (LaTeX, uses anvil-paper.cls by default)
    refs.bib                      Bibliography (BibTeX)
    figures/                      Figure assets and source scripts
      src/                        Optional source scripts (e.g., Python plot scripts)
      *.pdf / *.png / *.svg / *.tex   Rendered figures (TikZ .tex or rasterized)
    _progress.json                Phase state for this version
    changelog.md                  (revisions only) Maps prior critic notes to changes
  <thread>.1.review/              Reviewer output for version 1 (read-only)
    verdict.md                    Top-level decision (advance / block) + total /44
    scoring.md                    Per-dimension scores against the paper rubric
    comments.md                   Line-level comments keyed to main.tex sections
  <thread>.1.audit/               Fact / citation auditor critic sibling (read-only)
    citation-audit.md             Per-\cite{} resolution + claim-support check
    numerical-audit.md            Numbers-in-text vs figures/tables consistency
    flags.md                      Critical flags (unsupported citations, numerical disagreement)
  <thread>.1.litsearch/           (optional) Re-run litsearch after reviewer flagged missing prior work
  <thread>.2/                     Revised version (after revise consumes v1 + all critic siblings)
  ...
  <thread>.{N}/                   Terminal version, marked READY in its _progress.json

Versioned dirs (<thread>.{N}/) and critic sibling dirs (<thread>.{N}.<critic>/) are immutable once their _progress.json records the phase as done. Revisions are produced as a new version dir, never by editing in place.

The litsearch sibling is intentionally allowed at .0.litsearch/ (before any drafted version) AND at .{N}.litsearch/ (after a reviewer points out missing prior work). Both follow the same "N parallel critics, one reviser" rule: when present at <thread>.{N}.litsearch/, the next memo-revise-equivalent pass (paper-revise) consumes it alongside .review/ and .audit/.

State machine

Per-thread state, derived from on-disk evidence (not flags):

EMPTY → DRAFTED → REVIEWED → REVISED → … → READY → AUDITED
        ↑
        (optional .0.litsearch/ may exist before DRAFTED; it does not gate the machine)

State	Evidence
EMPTY	No <thread>.{N}/ directories exist (a bare <thread>.0.litsearch/ is still EMPTY for state purposes)
DRAFTED	Latest <thread>.{N}/ exists with main.tex + refs.bib + _progress.json.draft == done; no sibling review at the same N
REVIEWED	<thread>.{N}.review/verdict.md exists for the latest N
REVISED	A <thread>.{N+1}/ exists after a prior <thread>.{N}.review/
READY	Latest <thread>.{N}.review/verdict.md records advance: true AND no unresolved critical flag (in either .review/ or .audit/)
AUDITED	<thread>.{N}.audit/ exists alongside a READY version AND audit/_progress.json.audit == done AND flags.md records no unresolved critical flag

Thresholds: ≥35/44 advances. <35/44 requires revision. Any critical flag (from .review/ OR .audit/) short-circuits regardless of total — block until addressed.

Iteration cap: default max_iterations: 4 (so worst-case terminal version is <thread>.5/). The cap is configurable per-thread by writing { "max_iterations": <N> } to <thread>/.anvil.json in the thread root. Exceeding the cap marks the thread BLOCKED (in the portfolio orchestrator's report) and requires human review.

<thread>/.anvil.json schema

The per-thread config supports the following optional fields:

{
  "max_iterations": 4,
  "venue": "neurips",
  "artifact_verify": {
    "commands": ["lake build", "lake exe check_theorems"],
    "cwd": "proof/",
    "timeout_s": 300
  }
}

Field	Type	Default	Effect
max_iterations	int	4	Iteration cap (see above).
venue	string	none	Target venue slug. When set, paper-review also scores the paper against the matching venue YAML and writes _review.venue.json alongside _review.json. Advisory only; does not change the /44 gate. See "Venue overlays" below.
artifact_verify	object	none	Optional external-artifact verification gate (issue #663). When declared, paper-review runs each command in commands (a list of command strings) in the resolved cwd (thread-relative or absolute; defaults to the thread root) under a per-command timeout_s budget (default 300s). A failed command (non-zero exit or timeout) emits an artifact_verify_* critical flag that blocks the review. When absent, behavior is byte-identical to today. See "External-artifact verification" below.

Venue overlays (advisory)

anvil:paper supports venue-pinned advisory rubrics in addition to the generic /44. When <thread>/.anvil.json declares a venue, the reviewer scores the paper against the matching venue YAML and writes the results as a second _review.json-shaped file (_review.venue.json) in the same .review/ sibling dir. The venue overlay is advisory only — the generic /44 rubric remains the sole driver of the advance decision, preserving the framework-wide "the skill's generic rubric is the sole advance gate" invariant. See rubric.md for the convergence-gate semantics and anvil/lib/snippets/rubric.md for the framework-wide rule.

The Python-side schema and loader live in anvil/lib/rubric.py (Rubric, load_rubric, discover_venue_rubric).

Shipped venue YAMLs

Venue	File	Total	Notes
neurips	rubrics/neurips.yaml	/16	Soundness, presentation, contribution, novelty, reproducibility.
nature	rubrics/nature.yaml	/15	Broad significance, accessibility, evidence strength, novelty.
arxiv	rubrics/arxiv.yaml	/10	Citation completeness, reproducibility, clarity, scope/category.

Each YAML cites its public source in a header comment so the overlay can be updated as venue guidelines change.

Venue discovery (three-tier search)

discover_venue_rubric (in anvil/lib/rubric.py) searches three tiers in order — first hit wins:

1. Per-thread: <thread>/.anvil/rubrics/<venue>.yaml. For a single thread that wants a non-shipped venue overlay without touching the consumer install.
2. Consumer-installed: <consumer>/.anvil/skills/paper/rubrics/<venue>.yaml (where <consumer> defaults to the portfolio directory, i.e., <thread>.parent). For consumer-wide custom venues.
3. Skill-shipped: <skill_root>/rubrics/<venue>.yaml — the framework defaults (neurips, nature, arxiv). In an installed consumer repo, this is .anvil/skills/paper/rubrics/<venue>.yaml.

Search precedence: per-thread > consumer-installed > skill-shipped. Both override tiers are consumer-controlled; the per-thread file is more specific and wins. Skill-shipped is the fallback.

When venue is set but no matching YAML is found in any tier, the reviewer emits a stdout warning and proceeds with the generic rubric only. The review is not blocked by the missing venue overlay — the generic /44 gate continues to apply unchanged.

Adding a consumer venue

A consumer who wants to ship a custom venue (e.g., ICLR) drops a Rubric-shaped YAML into one of the two override tiers above. The YAML must validate against the Rubric schema in anvil/lib/rubric.py with advisory: true. See the shipped rubrics/neurips.yaml as a worked example. Set venue: <slug> in <thread>/.anvil.json to activate.

External-artifact verification (opt-in, fail-open) — issue #663

anvil:paper's render gate + numeric-consistency checker verify the paper artifact; the reproducibility rubric dimension (dim 5) scores whether replication info is present. Neither runs the external artifact a paper's central claim rests on — a companion Lean 4 proof repository that must lake build clean, a benchmark harness that must reproduce a headline number, a dataset whose checksum must match. A review pass can score a paper highly and advance it to READY while the artifact backing its central claim is silently broken (the motivating canary incident: a paper ACCEPTed with a headline theorem that did not actually hold, caught only later by manually running lake build).

A consumer opts a thread in by declaring an artifact_verify block in <thread>/.anvil.json (schema table above). paper-review step 4f then runs each declared command as a deterministic pre-scoring gate — run the artifact, not just the PDF. This extends the existing "Build / compile failure" critical-flag class (rubric.md) to cover external companion artifacts; it is not a new rubric dimension and the rubric total stays /44.

The contract mirrors the venue-overlay discovery/fail-open shape exactly (fail-open when undeclared, one-line stdout warning when declared-but-unresolvable):

• Absent (the default for every existing thread): no subprocess call, no _artifact_verify.json file, no finding — behavior is byte-identical to a pre-#663 review. This is the load-bearing fail-open contract.
• Declared and resolvable: each command in commands runs via subprocess.run in the resolved cwd (thread-relative or absolute; defaults to the thread root) under timeout_s (default 300s), subprocess-only with no new Python dependency. ALL commands run — an earlier failure does not short-circuit the rest — so a reviewer sees every failing step in one pass. A failed command (non-zero exit OR timeout) emits a CriticalFlag with type prefix artifact_verify_<n> (<n> is the 0-based command index) that routes through the unchanged anvil/lib/critics.py::compute_verdict path and forces Verdict.BLOCK — no aggregator or schema change.
• Declared but unresolvable (missing / non-directory cwd, or a command whose executable is not launchable): the gate fails open — a one-line stdout warning + a major finding in comments.md, and the review proceeds. A broken declaration is a defect worth surfacing, but it must not be indistinguishable from "the reviewer never checked" — so it does NOT block and does NOT silently pass as "verified".

The gate's raw stdout/stderr capture is written to <thread>.{N}.review/_artifact_verify.json (mirroring the render gate's _gate.json) for CI/operator inspection — a conditional output, NOT in the review's required-files manifest. The skill-local implementation lives at anvil/skills/paper/lib/artifact_verify.py (discover_artifact_verify, verify, ArtifactVerifyResult.to_review()/to_critical_flags()), modeled on anvil/lib/render_gate.py's GateResult shape but scoped to paper per the "wait for the second consumer before generalizing" rule.

Area Chair pattern (AI-Scientist) maps to existing N-critics-one-reviser

The "Area Chair ensemble" pattern published in AI-Scientist (Sakana, 2024) — multiple reviewer agents whose outputs are merged into a single recommendation — maps directly onto the framework's existing N parallel critics, one reviser primitive in anvil/lib/critics.py::aggregate. No separate AC role is needed in this skill; the reviser pass IS the AC pass. The venue overlay above is a parallel design move: the reviser already aggregates across all <thread>.{N}.<critic>/ siblings, so the venue scorecard at _review.venue.json is consumed in the same machinery as .review/_review.json, .audit/_review.json, etc.

READY vs AUDITED distinction. Unlike memo (where AUDITED is rarely reached), paper-audit is part of the normal lifecycle. A paper is not done until it reaches AUDITED: the auditor's job is to verify every \cite{} resolves and every cited claim is actually supported. The PDF compile check (see Acceptance criteria) lives in the auditor's responsibility because it requires the bibliography to be valid.

Command dispatch

Command	Role	Reads	Writes
paper	portfolio orchestrator	all <thread>.* dirs under cwd	(none; reports state per thread + recommends next command)
paper-litsearch <thread>	literature-search critic	<thread>/BRIEF.md (+ <thread>/refs/); for re-run, the latest <thread>.{N}/main.tex and any .review/ notes about missing prior work	<thread>.0.litsearch/ (initial) or <thread>.{N}.litsearch/ (re-run)
paper-draft <thread>	drafter	<thread>/BRIEF.md (bootstrapped via interview when absent + interactive; --no-interview keeps the fail-fast), <thread>/refs.bib (if present), <thread>/refs/, AND any <thread>.0.litsearch/ sibling	<thread>.1/ with main.tex + refs.bib + figures/ (+ <thread>/BRIEF.md on the bootstrap path)
paper-review <thread>	reviewer	latest <thread>.{N}/	<thread>.{N}.review/
paper-vision <thread>	vision critic (rendered-artifact)	latest <thread>.{N}/main.tex rendered to paper.pdf + per-page PNGs	<thread>.{N}.vision/ with _review.json (kind=vision)
paper-revise <thread>	reviser	latest <thread>.{N}/ + all <thread>.{N}.*/ critic siblings	<thread>.{N+1}/ with changelog.md
paper-audit <thread>	fact / citation auditor	latest <thread>.{N}/ (after reaching READY)	<thread>.{N}.audit/
paper-figures <thread>	figurer	latest <thread>.{N}/main.tex and figures/src/	figures into <thread>.{N}/figures/

The portfolio orchestrator (paper) is the user-facing entry point for status; the lifecycle commands are dispatched from it (or invoked directly by the orchestrating agent). paper-vision is an optional rendered-artifact critic sibling (like paper-review/paper-audit): it scores figure/table/equation defects visible only in the compiled PDF and feeds the same paper-revise aggregation. It does not gate the state machine on its own. See commands/paper-vision.md.

Progress tracking

Each version dir and critic sibling dir contains _progress.json. Schema (matches memo skill, with paper-relevant phases):

{
  "version": 1,
  "thread": "<thread>",
  "phases": {
    "draft":   { "state": "done",        "started": "<ISO>", "completed": "<ISO>" },
    "figures": { "state": "in_progress", "started": "<ISO>" }
  },
  "metadata": {
    "iteration": 1,
    "max_iterations": 4
  }
}

Phase states: pending, in_progress, done, failed. Validation is by file existence (does main.tex exist? does each \includegraphics{figures/fig-1.pdf} resolve to a file?), not by flag — _progress.json is a resume hint, not a source of truth. A phase that crashed mid-write should be re-runnable from pending after deleting any partial output.

The canonical _progress.json schema, read-merge-write recipe, and crash recovery contract live in anvil/lib/snippets/progress.md (in an installed consumer repo: .anvil/anvil/lib/snippets/progress.md); every command in this skill follows that convention. The merge is shallow: a command updates one phase, preserves all others. Critic siblings (<thread>.{N}.review/, <thread>.{N}.audit/, <thread>.{N}.litsearch/) follow the human-verdict scorecard kind per anvil/lib/snippets/scorecard_kind.md.

Rubric

See rubric.md for the 9-dimension /44 scoring schema (paper-tuned weights, rigor + evidence + citation hygiene = 17/44 ≈ 38.6%), the ≥35 advance threshold, and the critical-flag short-circuit policy. rubric.md also documents the vision-owned dimensions (label_cropping, axis_legibility, palette_adherence, mathtext_artifacts) scored by the optional paper-vision critic against the rendered PDF — these are an additive overlay (like the venue overlay), not part of the /44 gate.

Skill-specific phases

paper-litsearch (optional, read-only critic). Pure-LLM literature search is prone to hallucinating citations. This role MUST refuse to invent BibTeX entries. Instead, it produces:

• notes.md — discussion of how the paper positions against the related work the author already supplied, plus a list of gaps (named missing topics or specific known papers the brief mentioned but did not supply BibTeX for).
• candidates.bib — entries that come from author-supplied refs (re-formatted for consistency) or that the role is explicitly told about in the brief. Entries from "I think there's a 2023 paper about X by someone named Smith" are forbidden — the role surfaces such gaps in notes.md for the author to fill manually (e.g., by pasting a Semantic Scholar export into <thread>/refs/).

Opt-in web search (web_search: true) — issue #424

By default this skill runs no autonomous web search — the anti-hallucination posture above. A consumer can opt a thread in by setting web_search: true (a YAML boolean, default false) in the per-thread <thread>/BRIEF.md frontmatter; in a post-#295 project layout the equivalent carrier is the web_search key on the thread's documents: entry in the project BRIEF.md (schema-validated as a strict bool by anvil/lib/project_brief.py). When the knob is absent or false, paper-litsearch and paper-review are byte-identical to their default no-web behavior.

When enabled:

• paper-litsearch may run live academic web searches, under the resolver-verified-or-dropped contract: a web-discovered candidate enters candidates.bib ONLY after anvil/lib/cite.py::resolve() returns a BibRecord (Crossref for DOIs, arXiv API for arXiv IDs). No resolver hit, no BibTeX entry. Unresolvable hits (pmid/url identifier kinds — unsupported in v0, CiteResolutionError after retries, or no extractable identifier) become leads in the ## Web leads (unverified) section of notes.md — never citations. Verified web discoveries get a provenance table row in notes.md (bib key → identifier → resolver) so author-supplied vs web-verified entries stay auditable. The drafter and reviser MUST NOT cite leads; the author promotes a lead by supplying a resolvable DOI / arXiv ID.
• paper-review may run 3–5 targeted searches while scoring D4 (related-work positioning). The reviewer stays read-only with respect to citations: discovered identifiers land in comments.md as related-work-tagged leads recommending a paper-litsearch re-run, which centralizes the resolver verification in one command.

See commands/paper-litsearch.md § "Opt-in web search" and commands/paper-review.md Inputs for the full contracts.

paper-audit (mandatory at READY). Sharper than the generic auditor:

1. Verify every \cite{} in main.tex resolves to a real entry in refs.bib.
2. Spot-check that cited papers actually support the surrounding claim (the auditor reads <thread>/refs/ for any author-supplied source PDFs / notes; for citations whose source is not on disk, the auditor flags them as "claim-support unverified — source not on disk" rather than fabricating a verification).
3. Flag claims that should have a citation but do not.
4. Flag numerical values in the text (Tables, Sec. results) that disagree with figures/tables.
5. Verify the LaTeX compiles: run pdflatex/bibtex, then rerun pdflatex to the "Label(s) may have changed. Rerun" fixpoint, capped at 5 total pdflatex passes (latexmk-style convergence loop — see commands/paper-audit.md step 4; or equivalent) and capture the log. A non-zero exit, any unresolved ?? citation in the rendered PDF, OR non-convergence at the 5-pass cap is a critical flag.

paper-figures. Writes into the current <thread>.{N}/figures/ directory (not a sibling — figures are part of the artifact). The figurer SHOULD NOT invent data. Source scripts go in figures/src/; rendered outputs go in figures/. Conventions:

• TikZ / PGFPlots (.tex files in figures/) — for diagrams and small plots; vector-native in the compiled PDF, included via \input{figures/diagram-1.tex}.
• Matplotlib (Python in figures/src/*.py) — for data plots from real datasets; saved as .pdf, included via \includegraphics{figures/fig-1.pdf}.
• External assets (.png / .svg dropped into figures/) — allowed; the figurer leaves them alone.

The auditor (paper-audit) may re-run scripts in figures/src/ to verify rendered outputs are current; this verification policy is documented in paper-audit.md.

paper-review render-gate hook (deterministic pre-flight). paper-review runs a deterministic render-gate pre-flight via anvil/lib/render_gate.py (the LaTeX-skill analog of marp_lint for the deck/slides skills). The gate checks page count (page_cap=None — paper length is venue-dependent; consumers can override per-thread via <thread>/.anvil.json: render_gate.page_cap), overfull boxes (>5.0pt threshold), compile success, and source-side placeholders (TODO / [TBD] / (figure) / .MISSING). The gate runs after paper-audit has produced main.pdf + compile-log.txt; if invoked before audit, the gate fails open with a clear stdout message and the review proceeds without enforcement. On failure, the gate emits a typed Review(kind=tool_evidence) with one CriticalFlag per failed gate dimension, which the existing anvil/lib/critics.py::compute_verdict path treats as BLOCK. See commands/paper-review.md step 4b.

Templates / assets

• Default LaTeX class: templates/anvil-paper.cls — minimal generic single-column class with \title, \author, \abstract, standard sectioning, and \bibliographystyle{plainnat} baked in. Compiles cleanly with pdflatex + bibtex from a fresh checkout, no venue-specific assumptions. Supports an anonymous option (\documentclass[anonymous]{anvil-paper}) that suppresses author/affiliation rendering for double-blind submission, and a numeric option (\documentclass[numeric]{anvil-paper}) that switches citations from the default author-year (Author, Year) style to numeric [7]-style — the two options compose (\documentclass[numeric,anonymous]{anvil-paper}).
• Bibliography: BibTeX (.bib) is the primary format, with natbib for citation commands (\citep{}, \citet{}). Most venues accept either BibTeX or biblatex; BibTeX has wider tooling compatibility.
• Venue overrides: NeurIPS, IEEE, ACM, arXiv, etc. are handled by the standard anvil override mechanism — the consumer drops neurips_2024.sty (or equivalent) into .anvil/skills/paper/templates/ in their own repo and the brief instructs the drafter to switch the documentclass line (e.g., \documentclass{neurips_2024}). This skill ships no venue style files (licensing + staleness concerns).
• Entry-point template: templates/main.tex.j2 — Jinja2-style placeholder document with frontmatter hooks ({{title}}, {{author}}, {{abstract}}) and section skeletons (\section{Introduction}, etc.). The drafter substitutes from the brief and elaborates.
• Starter bibliography: templates/refs.bib.j2 — empty .bib with a comment header explaining that entries come from either author-supplied <thread>/refs.bib or the litsearch sibling's candidates.bib.
• Smoke test brief: assets/example-brief.md — a one-page paper brief the drafter can turn into a compilable 2–4 page paper with one figure and a handful of citations. Used by the acceptance test.

Project BRIEF artifact type

paper is registered as a skill-identity artifact_type value in the shared project-BRIEF registry (anvil/lib/project_brief.py::REGISTERED_ARTIFACT_TYPES / SKILL_IDENTITY_ARTIFACT_TYPES; issue #408, following the #386 pattern for deck/slides/proposal). In a shared project BRIEF, a documents: entry with artifact_type: paper declares that this skill owns the thread. It is NOT a memo subtype: it selects no memo rubric overlay, and memo commands fail loudly when pointed at a paper-declared thread. anvil:project-migrate writes this value when its BRIEF synthesis infers a paper-class thread from a .tex body with a non-anvil-proposal \documentclass.

Legacy pub compatibility (issue #694): this skill was renamed from pub to paper. The BRIEF parser accepts the legacy string artifact_type: pub as an input alias that normalizes to the canonical paper type, so existing consumer BRIEFs authored before the rename keep parsing. New BRIEFs SHOULD write artifact_type: paper. There is no skill-level forwarding alias — re-run install-anvil.sh with --skills=paper (not --skills=pub).

Defaults and overrides

This skill ships opinionated defaults. Consumers are expected to override liberally via .anvil/skills/paper/ in their own repo:

• voice.md (optional) — Author or lab voice/style guidance the drafter reads in addition to its base prompt.
• rubric.overrides.md (optional) — Add domain-specific critical-flag examples or adjust the open-ended "any dealbreaker a sophisticated reader would catch" instruction.
• templates/<venue>.cls or templates/<venue>.sty — Venue-specific style files. The brief tells the drafter to use them.
• BRIEF.md.example — Reference brief shape; freeform prose with optional YAML frontmatter is accepted.

Migrating an existing paper

Bringing a pre-existing, hand-authored LaTeX paper into the paper grammar is a first-class, sanctioned workflow — not a workaround. There are three supported class choices, all on equal footing; pick the cheapest one that preserves your paper's rendered output:

1. \documentclass{anvil-paper} (default). Use when the paper has no load-bearing preamble and its citations are author-year. You get margins, microtype, colored hyperref, and caption styling for free.
2. \documentclass[numeric]{anvil-paper}. Use when the only blocker is citation style — the paper renders numeric [7]-style citations. The numeric option swaps natbib into numeric mode (\citep/\citet unchanged) while keeping all of anvil-paper's other styling. This is strictly cheaper than a full keep-original-class migration, so prefer it when nothing but the citation style collides. Compose with anonymous if double-blind (\documentclass[numeric,anonymous]{anvil-paper}).
3. Keep the paper's original \documentclass (e.g. \documentclass{article}) and its preamble verbatim. Use when the preamble itself collides with anvil-paper — custom theorem environments, xcolor/hyperref/caption/amsthm package-option clashes, or a citation style anvil-paper does not support even with numeric. Set documentclass: article (or the paper's original class) in the brief frontmatter and keep the paper's preamble in main.tex unchanged. This is a sanctioned path on equal footing with the venue-style-file override, not a deviation to note apologetically.

The lifecycle is class-agnostic. draft → review → revise → audit → figures, the /44 rubric scoring, and the render_gate.py pre-flight all operate on the compiled artifact — none of them inspect or require anvil-paper.cls specifically, and no rubric dimension penalizes not using it. Keeping your own class costs you the free styling anvil-paper provides, nothing else.

Decision guidance: use \documentclass[numeric]{anvil-paper} when only the citation style is the blocker; keep the original \documentclass verbatim when the preamble itself collides (theorem environments, package-option clashes, or highly customized macros). See commands/paper-draft.md § "Documentclass overrides" for how the brief frontmatter selects the class.

Anonymous / double-blind submission

When the brief sets anonymous: true, the drafter:

1. Uses \documentclass[anonymous]{anvil-paper} (or the venue override's equivalent option).
2. Replaces author names and affiliations with Author Name(s) Withheld and Affiliation Withheld.
3. Scrubs identifying language in acknowledgements and self-citations (\citet{ourpriorwork} becomes \citet{anonprior} with a note in changelog.md).

Venue overrides handle their own anonymization on top of this.

Git sync hook (opt-in, off by default)

Consumers running anvil under an external orchestrator (a sphere channel-agent, a Loom-style daemon) can opt in to a per-phase git commit hook so every lifecycle phase leaves the working tree clean: a repo-level .anvil/config.json with git.commit_per_phase: true (and optionally git.push: true) has each write-bearing paper command end its phase by staging only the dirs it wrote and committing as anvil(paper/<phase>): <thread>.{N} [<state>]. The full contract — knob shape, defaults-off rule, commit-message format, staging scope, warn-and-continue failure semantics, and ordering after the _progress.json done write and the #350 sidecar atomic rename — lives in anvil/lib/snippets/git_sync.md (.anvil/anvil/lib/snippets/git_sync.md in an installed consumer repo). All 7 write-bearing paper commands adopt it; the read-only paper portfolio orchestrator is exempt by definition. When .anvil/config.json is absent or the knob is false, behavior is byte-identical to a pre-#426 install — the hook is default off.