A printed research paper on a desk

Roadmap

Roadmap

What draft does not yet do, stated plainly, and the order we mean to close the gaps.

draft is deliberate about its limits. A tool that sells verifiability should be the first to say what it cannot yet verify. This page is the honest list, and the FAQ links here from every limitation it mentions.

Planned

Table and figure values

Docling is already available as a reader (--reader docling), but tabular values do not yet become verifiable claims. The plan is to mine table cells as claims, with the row and column headers as the quote context, and to verify numbers against the cell. Figure captions are text and gate-compatible; chart values are a later step.

LaTeX and math

Today, math around a claim can cause the surrounding text to be dropped. The first step is to preserve inline math as opaque tokens so neighbouring claims still verify; the second is to accept .tex source directly, where the math is exact text and verbatim matching is clean.

OCR for scanned PDFs

A scanned PDF has no text layer, so extraction returns nothing and the run fails. The plan is to detect this and fall back to OCR, surfacing OCR-derived claims as a distinct, flagged class rather than treating them as equal to verbatim ones, so the trust model is preserved rather than quietly diluted.

More input and output formats

.docx input already works today — built in on macOS via textutil, and on any platform with --reader docling. The remaining input gap is HTML, which reduces cleanly to text once the reader seam accepts it. For output beyond the current Markdown set, an HTML export is the likely next step; a .docx export is better left to a documented pandoc recipe.

Non-English papers

The verification gate is largely language-agnostic — UTF-8, normalisation and verbatim matching do not assume English. The English-biased parts are the house-style rules and sentence-boundary repair. The plan is to make style rules locale-configurable with --style and to document current behaviour on non-English input.

Not planned as the default

A direct API-key mode (--engine api:<provider>) is possible as an escape hatch for users with no agent CLI, but keyless operation is the whole point, so it will never be the default story.


Want something moved up the list? Open an issue on GitHub.