Skip to content

How it works

Paper Espresso separates model judgment from deterministic infrastructure. The model reads, reasons, selects, and writes. Scripts handle retrieval, archive safety, compilation, page checks, layout checks, and cleanup.

1. Resolve the exact paper

The workflow accepts a title, arXiv identifier, arXiv URL, or local PDF. Ambiguous title matches must be resolved before analysis begins.

2. Prefer original source

When arXiv TeX source is available, Paper Espresso acquires it safely and reads it as data. Original source preserves notation, labels, equations, tables, captions, bibliography, visual assets, and document structure better than extracted PDF text. Consequential figures, diagrams, and plots are inspected directly rather than inferred from captions.

Downloaded TeX is never compiled or executed.

3. Fall back carefully to PDF

When only a PDF is available, the complete document is still read. Extracted text assists navigation but is not treated as unquestioned truth. Consequential equations, tables, figures, symbols, and multi-column passages are checked against rendered pages.

If a detail cannot be extracted confidently, it is omitted or qualified rather than reconstructed by guesswork.

4. Understand before drafting

The agent does not begin after reading only the abstract, introduction, and conclusion. It first builds a coherent model of:

  • the actual problem;
  • the precise contribution;
  • the central insight and mechanism;
  • indispensable assumptions and definitions;
  • essential mathematics or conceptual structure;
  • the strongest evidence and its conditions;
  • limitations and unresolved questions.

5. Decide what survives

Material is classified as essential, supporting, or nice to have. The digest follows the paper's intellectual structure rather than its table of contents.

The prose defaults to plain wording, concrete subjects and actions, and active voice where scientifically natural. It keeps the paper's original terminology when a term names a distinct construct, preserves precision, or helps readers navigate the field, and defines that term close to first use. Simplification may reduce decoding effort but may not erase assumptions, qualifications, or technical distinctions.

The digest normally uses three to five compact visible guideposts to expose the problem and contribution, mechanism or proof spine, evidence and scope, and limits. Labels and order adapt to the paper, and related guideposts may be combined; they do not reproduce the table of contents. Mathematics remains inline when legible; standalone equations use minimal surrounding space. Light pastel backgrounds may connect a few semantic subexpressions to named phrases in adjacent prose. Color is never the only reference. Narrow figures or compact mathematical objects may be wrapped so prose fills the adjacent area. Captions and numbering appear only when indispensable.

6. Compile and inspect

The generated digest is checked against the trusted preamble before compilation. Only the preflighted document and explicitly permitted assets enter an isolated compile directory; shell escape is disabled and stale outputs are removed. Validation checks:

  • maximum page count, or exact count only when explicitly requested;
  • unresolved template placeholders;
  • unsafe source constructs;
  • undersized typography;
  • compilation errors and overfull boxes;
  • large unused bands, underfilled column bottoms, and badly imbalanced columns as diagnostics;
  • clipping, collisions, and visual readability on rendered pages.

The raster check measures the actual text body and each column independently. Its findings are warnings, not acceptance criteria. Underfill triggers one check for the next-highest-value missing understanding; if the mental model is complete, whitespace is accepted. Manual breaks, balancing packages, stretched glue, tiny type, and filler remain prohibited.

7. Apply the intellectual approval bar

Before delivery, the digest is reviewed against the complete paper. A technically competent reader should be able to explain the problem, contribution, mechanism, indispensable mathematics or assumptions, strongest evidence, and material limitations. The review also removes unnecessary jargon and indirect phrasing while checking that every retained technical term and difficult distinction remains accurate.

The final deliverables are the requested .tex and .pdf. Temporary downloads, extracted source, and rendered review pages are removed on success and cleaned up on failure.

8. Evaluate the learning product

The repository includes a three-paper benchmark spanning empirical architecture, theorem-heavy analysis, and distributed systems. It tests immediate, source-grounded comprehension with the digest available for consultation. Fidelity is a gate; unsupported load-bearing claims, altered conditions, or changed quantitative results fail the case. The benchmark deliberately excludes delayed recall and never adds test questions to the generated artifact.