phd-paper-analysis

name

phd-paper-analysis

description

Analyze exactly one academic paper and deliver a self-contained, evidence-traceable PhD-level Traditional Chinese HTML report. Use for a local or uploaded PDF, paper URL, DOI, arXiv ID, exact title, or full citation when the user wants full-paper close reading; method, architecture, equation, figure, table, experiment, or ablation interpretation; reproducibility and critical review; or an explicit assessment of transferability to audio deepfake detection. Do not use for multi-paper surveys, routine summaries, or citation-only literature searches.

PhD Paper Analysis

Objective

Turn one verified paper into one rigorous, self-contained Traditional Chinese research artifact:

one paper -> verified source set -> complete evidence inventory -> critical synthesis -> validated HTML

Optimize for traceability and technical understanding, not length. Preserve the user's requested lens. Never replace missing evidence with plausible-sounding detail.

Non-negotiable scope

• Analyze exactly one paper per run.
• Treat the main paper, its appendix, and its official supplement as one source set.
• If the input identifies multiple papers, pause and ask the user to select one. Do not silently choose the first.
• If no paper is identified, ask for one unless the user explicitly authorizes selection.
• Do not turn the task into a literature survey. Use related work only to verify context or evaluate a claim.
• Do not schedule, monitor, or create recurring tasks.

Load the report contract

Read references/report-contract.md completely before drafting. It defines the required HTML sections, evidence labels, and acceptance criteria. Apply every requirement unless the user explicitly overrides it.

Workflow

1. Resolve and verify the paper

Accept a local/uploaded PDF, URL, DOI, arXiv ID, exact title, or full citation.

1. Resolve bibliographic metadata from the paper itself and a primary or authoritative page when available.
2. Record the exact title, ordered authors, venue/source, year, identifiers, canonical URL, and local path if applicable.
3. Prefer the publisher/proceedings page, arXiv, OpenReview, PubMed, ACL Anthology, IEEE, ACM, Springer, or an official author/project page over aggregators.
4. Detect version differences between preprint, accepted manuscript, and published paper. State which version is analyzed and use one version consistently for page, figure, table, and equation references.
5. If full text is inaccessible, stop and explain what is available. Do not produce a full-paper report from metadata or an abstract alone unless the user explicitly requests a limited abstract-only analysis.

2. Acquire the complete source set

Obtain the main text plus official appendix or supplement when available. For PDF input, use PDF extraction and page rendering; visually inspect pages containing equations, figures, tables, multi-column layouts, or extraction anomalies.

Use web search only to obtain authoritative paper materials, verify released code/data, or check a narrowly scoped factual claim. Record external sources separately from paper evidence.

3. Build an internal evidence inventory

Before writing, inventory:

• section and page ranges;
• all numbered equations and unnumbered equations central to the method or loss;
• every figure and table, including appendix items used by the argument;
• datasets, splits, preprocessing, baselines, metrics, hyperparameters, compute, and code/data availability;
• main claims, supporting evidence, reported limitations, negative results, and omitted details;
• all ablations and sensitivity analyses.

Use this inventory to prevent selective reading. Do not expose raw scratch notes unless useful; summarize coverage in the report's evidence-and-limitations section.

4. Reconstruct the method

Explain the full data path from raw input to final output. For each major component, state:

• input and output, including tensor shape or signal domain when recoverable;
• operation and trainable parameters;
• role in training and inference;
• connection to the objective;
• design rationale supported by the paper;
• unresolved ambiguity or inferred interpretation.

Separate what the authors state from your reconstruction. Never describe an inferred implementation detail as reported fact.

5. Analyze mathematics faithfully

For each important equation:

1. Reproduce it in MathML without changing notation.
2. Cite its equation number and page or section when available.
3. Define every symbol, index, operator, domain, dimension, and unit that can be established.
4. Explain the computation in plain Traditional Chinese and, where justified, step by step.
5. Connect it to adjacent equations, architecture, gradients or optimization, and representation behavior.
6. State ambiguities instead of inventing missing definitions.

Derive only consequences that follow from the paper or standard mathematics. Label any added derivation as the analyst's derivation.

Use native MathML for every mathematical variable or expression in report prose. Do not emit LaTeX delimiters, LaTeX commands as text, Unicode-only equation approximations, formula screenshots, or <code> as a substitute for mathematics.

6. Analyze visual and empirical evidence

• Explain each method-defining figure and every result-bearing table used in the conclusions.
• Embed an extracted original figure as a base64 data URI when technically possible. Preserve aspect ratio and add a precise Traditional Chinese alt description and source locator.
• If extraction fails, provide a clearly labeled textual reconstruction; do not fabricate an image.
• Preserve reported values exactly. Distinguish percentages, percentage points, relative changes, means, dispersion, and confidence intervals.
• Recalculate deltas only when inputs are available, and label recalculated values.
• Evaluate whether the metric, split, baseline strength, variance, statistical testing, and experimental controls support the authors' claims.
• If no ablation, error analysis, significance test, or failure case is reported, say so explicitly and explain the resulting evidentiary limit.

7. Assess reproducibility and external relevance

Separate reproducibility into reported, verified externally, inferred, and missing. Verify code and data availability from official repositories or project pages when possible, including access date.

For audio deepfake detection, provide a transfer assessment even for an out-of-domain paper, but do not force a positive connection. Rate relevance as high, medium, low, or none; justify the rating; map reusable components to spoofing, synthetic speech, replay, or voice-conversion detection; identify domain mismatches; and propose falsifiable adaptation experiments. Label the entire section as inference unless the paper directly evaluates audio deepfake detection.

8. Write the artifact

Write a standalone UTF-8 HTML5 file in the user's requested location. Otherwise use <paper-slug>-analysis-zh-Hant.html in the current workspace. Do not paste thousands of lines of HTML into chat when a file can be delivered.

Make the document usable offline:

• set <html lang="zh-Hant"> and include <meta charset="utf-8">;
• include an informative <title>, semantic headings, table of contents, responsive CSS, print CSS, and accessible tables/figures;
• use no external script, stylesheet, font, image, analytics, or network dependency;
• use base64 data URIs for embedded figures;
• wrap wide MathML and tables in horizontally scrollable containers;
• provide clickable source links plus paper-local locators;
• put limitations next to affected claims, not only in a final disclaimer.

9. Validate before delivery

Resolve this skill's installation directory, then run:

python3 <skill-directory>/scripts/validate_report.py <report.html>

Add --require-mathml if the paper contains important mathematics and --require-embedded-images if extractable figures were available. Fix every error. Then render or open the HTML and visually inspect at least the title/TOC, one equation-heavy section, one figure or table section, and the page end at desktop and narrow widths.

Before handoff, verify:

• the analyzed identity and version are unambiguous;
• all required sections exist, including explicit “not reported” findings where applicable;
• every central claim has a paper locator, external citation, or inference label;
• no unavailable content is presented as read;
• MathML renders correctly and no placeholders remain;
• tables, figures, and long formulas do not overflow;
• the final response links to the HTML file and briefly states source limitations and validation performed.

Failure handling

• Unreadable or encrypted PDF: report the failure and request another copy or authoritative URL.
• Broken text extraction: render and inspect the affected pages; use OCR only if available and flag OCR uncertainty.
• Missing supplement: identify it as unavailable and narrow conclusions that depend on it.
• Conflicting metadata: prefer the analyzed document for content and the authoritative landing page for publication metadata; disclose the conflict.
• No relevant audio-deepfake transfer: report low or no relevance with reasons instead of inventing a connection.
• Validator unavailable: perform the same checks manually and disclose that automated validation was not run.