Most automated writing tools produce text but not verifiable science. Spark-to-Paper changes the unit of automation from a single draft to a data-chained research workflow: it requires evidence up front, separates experiment planning from reporting, and combines model judgment with deterministic gates so that manuscript claims are accepted, revised, or abandoned according to measured outcomes.
Key Findings
- End-to-end implementation as 13 composable skills inside a coding assistant — no separate orchestration service required — lets the system span literature search, experiment planning, execution, figure generation, and LaTeX compilation. This design emphasizes reproducibility by tracing numeric outputs to source data and code.
- High empirical integrity: 99.5% citation validity and 96.4% figure editability in controlled topics. An integrity stack raised fabrication-detection from 14% (single-pass draft) to 92% (full stack); adversarial review achieved 74% precision. Those numbers demonstrate the impact of deterministic gates plus multi-stage review.
- Operational cost and scale: typical manuscript used 11.9M tokens, cost ~$8.10 in API fees, and took ~3.2 hours on average to produce — indicating lightweight resource needs compared with ad hoc human workflows.
Who it's for and tradeoffs
Great fit if you want rapid, reproducible draft generation that keeps experiments central: researchers prototyping ideas, reproducibility-focused teams, and coding-assistant integrators who can run programmatic plots and deterministic checks. Look elsewhere if you need fully autonomous novel experiments at large scale (Spark-to-Paper focuses on feasible, audit-friendly runs), or if you cannot run code inside the target coding-assistant environment (the skills assume that runtime).
Where It Fits
Compared with multi-agent pipelines that orchestrate distinct agents, Spark-to-Paper embeds composable skills inside an existing coding assistant and leans on deterministic Python gates for auditability. It complements projects that automate sections of the paper pipeline by emphasizing evidence-first experiment planning and reconstructible vector figures.
How It Works (brief)
The system: (1) converts a one-line idea or proposal into a staged plan, (2) specifies required evidence and planned experiments before execution, (3) runs feasible experiments or fills result placeholders deterministically, (4) generates editable figures via programmatic plotting and reconstruction, and (5) assembles and compiles LaTeX while enforcing citation and number-audit gates. Self-review and adversarial peer-review skills further harden outputs against fabrication.