Paper Poster HTML

Paper Poster (HTML): measurement-gated poster generation

One HTML file styled for an exact print canvas (@page { size: W H }), rendered to PDF via Playwright print emulation. Iterate by measuring, not eyeballing — the screen preview lies; only print emulation at the correct viewport tells the truth. Core gate machinery is adapted from posterly (MIT, © 2026 Ruishuo Chen — see NOTICE.md and LICENSES/posterly-MIT.txt); ARIS adds style discipline gates, figure-provenance gates, the self-review loop, and the anti-patch-loop fix vocabulary.

Why this skill exists (the failure it prevents)

A predecessor pipeline produced a poster with 30+ colors, zero real paper figures, a screen-pixel canvas, and tiny formulas floating in oversized boxes, then spent 12+ review rounds making it worse — each round added a new badge color or bespoke SVG patch. The cure is structural, not exhortative:

1. Hard gates run before any aesthetic opinion (alignment, style, assets must PASS first — a reviewer never sees an unmeasured poster).
2. A closed fix vocabulary — visual-review fixes can only touch design tokens, whole catalogued components, content rebalance, assets, or canvas choice. New inline styles / new hex values / bespoke decorations are structurally forbidden.
3. Two-hue discipline as a machine check, not a style suggestion.
4. Real paper figures with provenance manifest, or the gate fails.

Mental model

paper (.tex / PDF) ──► content plan + claim→evidence audit (Claude, fresh self-review)
                              │
   figures extracted ─────────┤  FIGURE_MANIFEST.json (provenance, sha256)
   (real paper figures ONLY)  ▼
   template scaffold ──► fill ──► run_gates.py            ◄─── HARD, loop here
                                  preflight → style → asset → measure → polish
                              │ all hard gates PASS
                              ▼
                    Claude visual review (≤3 issues × ≤3 rounds, fix-vocabulary only)
                              │ score ≥ 9
                              ▼
                    Claude final self-review (fresh re-read, full HTML+PDF)
                              │ pass
                              ▼
                    verify-final → poster.pdf + GATE_REPORT.json

Constants

• SKILL_SCRIPTS = ${CLAUDE_SKILL_DIR}/scripts — all helpers are single-owner and ship inside this skill (Arch C). If the directory is missing the install is broken: abort and tell the user to re-install the skill (Policy A — the gates ARE the skill; never improvise replacements).
• Review backend — self-reviewed by Claude (no external model configured): every review pass (Phase 1 claim audit, Phase 6 final review) re-reads the artifacts fresh off disk, never carried forward as conversational memory across review boundaries. See auto-review-loop's "Self-Review Backend (No Second Model)" for the full rationale and honest tradeoff (no independence guarantee; self-review can't catch blind spots the way a different model could).
• CANVAS — from the venue's official spec, looked up live in Phase 0. Never assume. (Known anchor: ICLR 2026 main = 185×90 cm landscape per its official printing service; ICML/NeurIPS commonly 60×36 in landscape; workshop posters often 61×91 cm portrait. Specs change yearly — verify.)
• PALETTE — default = templates/tokens/generic.json (slate-blue #2D5F8B accent
  • gold #C9A24A highlight + neutrals) for all venues. Venue packs are opt-in via — venue-colors: true. Purple-dominant accents (hue 250–285) are banned unless the user passes — allow-purple: true.
• AUTO_PROCEED = false — wait for explicit confirmation at every 🚦 checkpoint.
• OUTPUT_DIR = poster_html/ in the working directory.

Workflow

Phase 0 — Resume, dependencies, venue spec

1. Resume: if poster_html/POSTER_STATE.json exists with status: in_progress (< 24 h), resume from the saved phase.
2. Dependencies (degradation chain, in order):
   • Playwright + bundled Chromium → if missing, python3 -m playwright install chromium → if install fails but system Chrome exists, scripts fall back to channel="chrome" → if all fail: you may produce the content plan and scaffold only, label everything "not print verified", and must NOT emit a final PDF.
   • pdfinfo missing → PyMuPDF reads PDF dimensions. At least one of pdftoppm / PyMuPDF must exist for PNG review renders.
   • MathJax: download tex-svg.js once into poster_html/assets/mathjax/ and reference it locally in the HTML. CDN is acceptable only for drafts; the measure gate hard-fails on unrendered MathJax either way.
3. Venue spec lookup (live): consult the venue's official poster-instructions page (search + fetch). Extract dimensions, orientation, font floor, logo policy, anonymity rules, file format. Record {spec, source_url, retrieved} into POSTER_STATE.json — specs change yearly; never reuse a cached spec silently.

🚦 Checkpoint: echo the venue spec table (canvas, orientation, source URL) and the chosen template. Wait.

Phase 0.5 — Design discovery (one AskUserQuestion batch)

Ask once, ≤4 questions: layout template (from templates/README.md), palette (default generic pack / venue pack / custom within constraints), logos + venue mark (paths or "none" — never fabricate; check the venue's logo policy), QR target (paper / code / project page / none — generate offline with qrencode or python-qrcode; never a remote QR-service URL). Persist answers in POSTER_STATE.json as design_decisions — re-read before any later "improvement" so deliberate choices are never reverted.

Phase 1 — Paper ingest, content plan, claim audit

1. Read the paper source (.tex ideal; PDF otherwise). Extract: title/authors/affils, the 3–5 headline numbers, core method (equations verbatim), main results (tables/figures and what they show), takeaways. Build poster_html/POSTER_CONTENT_PLAN.md — what goes in which column, word budget per card. Target density (excluding table cells, captions, author line, footer): standard poster 550–850 words; dense theory+empirical poster 750–1050 words, allowed only when ≥2 compact components are used (eqn-anatomy, flow-strip, derived-col, claim-pills, keybox--4). Warn yourself below 500 words on a 4-column landscape (it will read as sparse next to professionally dense posters) unless the template is hero/visual-first; warn above 1100 unless the user asked for dense mode. Bullets ≤ 8 words when possible — density comes from structure, not long prose. Prefer compact structure over prose: if the paper contains an explicit objective, algorithm, theorem mechanism, or baseline comparison, extract at least two of: (1) empirical objective / loss stack; (2) term-by-term equation anatomy; (3) a method-flow strip grounded in paper variables; (4) a derived-Δ column for method-vs-baseline rows; (5) a 4-up implementation/theory keybox; (6) a claim/evidence pill table for numeric-heavy posters. Do not invent an algorithm. If the paper has only an objective, label the component "objective flow" or "loss anatomy", never "algorithm".
2. Self-review content audit (fresh re-read): re-read the content plan and the paper source path(s) directly — not from memory of drafting the plan — and build a claim→evidence table: | claim on poster | paper file:line | paper says (verbatim) | match? | with match ∈ {OK, NUMERIC-MISMATCH, OVERCLAIM, MISSING-PRECONDITION, NOT-IN-PAPER, SCOPE-NARROWED}. Flag any claim you're too invested in to judge impartially. Save to poster_html/CLAIM_EVIDENCE.md.
3. Fix every non-OK row or record it as a user-acknowledged tradeoff.

🚦 Checkpoint: content plan + audit summary. Wait.

Phase 2 — Real paper figures (provenance-gated)

Source preference chain:

1. Paper source figures/ (vector SVG/PDF → convert to SVG via inkscape/pdf2svg if available, else rasterize ≥ 2× rendered px).
2. PDF-only: extract_pdf_figures.py contact-sheet + auto to list candidate regions → pick crops (🚦 human confirms crop choices) → crop at 300–450 DPI.
3. Last resort: user supplies explicit page,x0,y0,x1,y1 bboxes.

Then preprocess_figures.py --autocrop every asset. Every paper-derived image gets a FIGURE_MANIFEST.json entry (source hash, page, bbox, dpi, sha256, natural_px) and is embedded as <img data-source="paper" data-asset-id="...">.

Hard rule: ≥ 2 paper-derived visuals or the asset gate fails. Theory-only papers may waive the total-area rule (--waive-total-area) at a human checkpoint — never silently. Never draw bespoke decorative SVG "figures" as substitutes.

Figure-area bands (asset gate, fractions of body): total target 14–22 % (warn < 12 % / > 24 %, hard < 10 % / > 28 %); per ordinary figure target 4–8 % (warn

10 %, hard > 13 %); figure--duo combined 8–12 %. Hero templates pass --hero (centerpiece may take 30–40 %). The failure mode is symmetric: too small reads as decoration, too big crowds out content. Sibling figures that share axes or tell a before→after story belong in one figure--duo card, not two cards.

Phase 3 — Scaffold + tokens

cp templates/<chosen>.html poster_html/poster.html; retarget @page + .poster dims to the venue canvas (two edits, same values); apply the chosen token pack onto the :root DESIGN TOKENS block; fill content per the plan; embed manifest figures. Run preflight + style_check — both must PASS before any layout iteration. (A fresh scaffold is expected to fail measure — that gate judges a filled poster.)

Phase 4 — Layout hard loop

After every layout change:

python3 "$SKILL_SCRIPTS/run_gates.py" poster_html/poster.html \
    --tokens <pack.json> --manifest poster_html/FIGURE_MANIFEST.json \
    --report poster_html/GATE_REPORT.json

Canonical order: preflight → style → asset → measure → polish. Targets: column-bottom spread < 5 px (aim < 3), footer gap ∈ [30, 50] px, intercard gap ∈ [12, 50] px, canvas-fill ∈ [95, 101] %, poster bbox aligned to page within ±2 px. Fix guidance for each failure mode lives in the gate output and templates/COMPONENTS.md. Do not proceed while any hard gate fails. Do not let a reviewer see an unmeasured poster. Balance under-filled columns with content from the paper (Gate C), never with whitespace, space-between, or stretched cards.

Phase 5 — Claude visual review (gated aesthetics)

Render and read the result yourself:

python3 "$SKILL_SCRIPTS/render_preview.py" poster_html/poster.html
pdftoppm -r 100 poster_html/poster_preview.pdf poster_html/review_full -png -f 1 -l 1
# plus 2-4 region crops at higher res (header / one column / equations) via PIL

Calibrate first (../shared-references/taste-calibration.md): if human-curated references/good/ + references/bad/ exist under this skill dir (or the project supplies its own pair), score those 3+3 reference posters on the axes below BEFORE the target, anchoring the scale. Never select, search for, or generate anchors yourself; if no reference sets exist, proceed uncalibrated and mark CALIBRATION: none — never fabricate anchor scores. Axes (weights sum 1.0): Design 0.35 · Craft 0.30 · Functionality 0.20 · Originality 0.15. Mapping: SCORE = min(round(1 + 9 × COMPOSITE), lowest triggered cap) — caps apply AFTER the mapping, and the loop's Score ≥ 9 threshold below always reads this final capped SCORE, never the raw composite.

Score strictly 1–10. Critical caps (hard floors — a calibrated composite never overrides them): < 2 real paper figures → ≤ 3; broken canvas / clipped content / unreadable math → ≤ 4; ≥ 4 visible hue families or gradient-heavy header → ≤ 4; large blank cards or columns → ≤ 5; fabricated visual claim → ≤ 3. Checks: posterly-showcase gestalt (would this hang next to a professionally designed poster without looking like a patched dashboard?), single-accent discipline, real figures readable and central, print hierarchy (title → headline stats → figures → detail), column fill, equation prominence (no tiny math in oversized boxes), serif-body/sans-display pairing, no gradient kitsch, component consistency, 60-second narrative. Output format:

SCORE: N/10            (= min(round(1 + 9 × COMPOSITE), lowest cap); drives the loop)
COMPOSITE: 0.xx        (weighted; list the four per-axis scores)
CALIBRATION: anchored | none
GAP: <which reference poster the target falls short of / exceeds, on which axis, and why — one paragraph; omit only when CALIBRATION: none>
CAPS_TRIGGERED: ...
TOP_ISSUES: (max 3)
ALLOWED_FIX_TYPE per issue: token | component | rebalance | asset | template/canvas
PATCH_LOOP_RISK: low | medium | high

Loop: fix (fix vocabulary below) → re-run Phase 4 gates → re-score. ≤ 3 issues per round, ≤ 3 rounds. Score ≥ 9 → Phase 6. Still < 9 after 3 rounds → STOP patching; escalate to template / canvas / content re-choice (back to Phase 3) or a human decision. Never enter round 4 of cosmetic patching.

Fix vocabulary (closed set — the anti-patch-loop core)

Allowed: (a) edit a :root token value; (b) swap/remove/add a whole component instance from templates/COMPONENTS.md; (c) content rebalance (move a card across columns, trim/grow text from the paper, resize a figure within its AR band); (d) template/canvas re-choice; (e) global edits to an existing component's CSS that reference only tokens; (f) switching predefined variants (.eqn--large, .card--compact, .figure--wide, .nowrap, …); (g) asset fixes (re-crop, swap for a clearer figure from the same paper, re-preprocess).

Forbidden: new inline styles, new hex values anywhere, bespoke decorative SVG, per-element font-size overrides. A new component may not be born inside the visual loop — stop, get a human checkpoint, add it to COMPONENTS.md, re-run from Phase 3.

Phase 6 — Self-review final pass (fresh re-read)

All hard gates PASS + polish warnings zero-or-waived + visual ≥ 9 first. No external model is configured, so Claude reviews the final artifacts itself (not the content plan) in a dedicated fresh pass — re-reading them off disk, not from memory of building them: poster.html, the rendered PDF/PNG, the paper source, GATE_REPORT.json, CLAIM_EVIDENCE.md. See auto-review-loop's "Self-Review Backend (No Second Model)" for the full rationale and honest tradeoff (no independence guarantee; self-review can't catch blind spots the way a different model could). It checks: (1) fidelity & overclaims re-checked on final text (polish introduces new claims), (2) residue (\ref{, TODO, raw < in math, missing images, remote URLs), (3) visual rhetoric (headline numbers prominent, banner readable from 2 m), (4) gate-log coherence. This pass recommends; it does not edit directly. Any fix → back through Phase 4/5 gates — never straight to re-review.

Phase 7 — Final verification + report

python3 "$SKILL_SCRIPTS/poster_check.py" verify-final poster_html/poster_preview.pdf \
    --from-html poster_html/poster.html --max-size-mb 20

Page count 1, dimensions match @page, size ≤ 20 MB, no TODO/residue, no remote assets. Report: PDF path, final spread px, footer-gap range, gate summary table, unresolved waivers, self-review verdict. Update POSTER_STATE.json → done.

State persistence

poster_html/POSTER_STATE.json: {phase, venue, canvas{w,h,orientation,source_url, retrieved}, template, token_pack, design_decisions{...}, figures_selected[], visual_rounds, review_passes{audit, final}, status, timestamp} — written after every phase; enables compact-recovery resume.

Key rules

• Measure, don't eyeball. No layout claim without run_gates.py output.
• Gates before aesthetics. Every review pass only ever sees a poster whose hard gates PASS. This ordering is what kills the patch-loop death spiral.
• Never invent paper numbers or figures. Numbers come from the paper source; visuals carry manifest provenance. Fabrication = critical cap ≤ 3.
• Two hues, one system. Accent + gold + neutrals. The style gate enforces it; don't negotiate with the gate.
• Real figures are the poster. A poster without the paper's own figures is a dashboard, not a poster.
• Fix vocabulary is closed. If a fix isn't expressible as token / component / rebalance / asset / canvas, it's the wrong fix.
• Self-review verdicts. Claude drives the loop and scores visuals; acceptance of content fidelity comes from a separate fresh self-review pass (a loop can drive, never acquit its own work in the same pass). No independence guarantee — see auto-review-loop's "Self-Review Backend (No Second Model)".
• Preserve user decisions. Re-read design_decisions before "improving" anything.
• Vendor boundary. poster_check.py, render_preview.py, _posterly/ are vendored from posterly — keep diffs minimal; ARIS-side logic goes in the new scripts, not in vendored files.

Review tracing

Save every self-review pass's trace per shared-references/review-tracing.md to .aris/traces/paper-poster-html/<date>_run<NN>/ (audit + final passes, raw responses).

Output contract

poster_html/
├── poster.html              # single-file source of truth
├── poster_preview.pdf       # print-emulated, verify-final-checked
├── poster_preview.png       # thumbnail
├── POSTER_STATE.json        # resume state
├── GATE_REPORT.json         # canonical gate ledger (schema v1)
├── POSTER_CONTENT_PLAN.md   # what-goes-where + word budgets
├── CLAIM_EVIDENCE.md        # self-review claim→evidence audit (no external model configured)
├── FIGURE_MANIFEST.json     # figure provenance (sha256, page, bbox, dpi)
└── assets/{paper_figures,logos,qr,mathjax}/

When NOT to use

• Slides, not a poster → /paper-talk / /slides-polish.
• The paper's headline isn't stable yet — fix the paper first; a poster amplifies whatever story it's given.